Can large language models effectively reshape online implicit hate speech? An integrative modelling approach

Can large language models effectively reshape online implicit hate speech? An integrative modelling approach | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Can large language models effectively reshape online implicit hate speech? An integrative modelling approach Yinghui Huang, Qixia Feng, Hui Liu, Weiqing Li, Ying Ma, Zongkui Zhou This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7368894/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Implicit Hate Speech (IHS) presents major challenges for traditional content governance. Using large language models (LLMs) for reshaping has become a promising new approach. This study evaluates the ability, strategies, and risks of LLMs in reshaping IHS. We use an integrative modelling method. First, we build a predictive model to measure the external effect of LLM reshaping. Second, we design an explainable evaluation framework with four dimensions: group-specific harm, implicit emotional expression, linguistic obfuscation and extremity, and bias and implication in social interaction. Results show that LLMs (e.g., GPT-4o and DeepSeek) can strongly reshape IHS texts in topics such as threatening and inferiority, reducing toxicity by 86.2%–90.57% while keeping high semantic similarity (BERTScore F1: 82%–85%). However, reshaping is not full detoxification. It often replaces risk with new covert forms. Explicit attacks are reduced, but covert risks may appear through strategies like vague references, hiding emotions, or adding logical gaps. This study confirms the value of LLMs in IHS governance, but also reveals their “replace-rather-than-remove” pattern. The framework we propose is a useful tool to detect and manage covert risks caused by algorithms, offering both theoretical and practical guidance for creating a more civil online space. Humanities/Complex networks Social science/Complex networks Physical sciences/Mathematics and computing Biological sciences/Psychology Social science/Psychology Implicit Hate Speech Large Language Models Linguistic Style Reshaping Integrative Modelling Explainable Artificial Intelligence Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Introduction Social media has become a key space for global information exchange. The governance of online hate speech is now an urgent global issue. A report from the Anti-Defamation League (ADL) shows that 25% of adults have faced severe online harassment. Among groups such as LGBTQ + people, this number rises to 49% (League, 2024 ). In response, different regions have taken different paths. The United Nations promotes countering “hate speech” with “more speech,” focusing on education and counter-narratives (Nations, 2019 ). China has set strict rules to control “harmful information,” including hate speech, aiming to build a “clean” online space (Nations, 2019 ). China has set strict rules to control “harmful information,” including hate speech, aiming to build a “clean” online space (China, 2019 ). The European Union’s Digital Services Act (DSA) requires large platforms to assess and reduce risks, with clear rules and user rights (Council of the European Union, 2022). In this context, implicit hate speech (IHS) has become a major challenge in online governance. IHS uses hidden expression, depends on cultural context, and has complex meanings. Unlike direct attacks, it uses tools like metaphor, irony, or vague references (Ocampo, et al., 2023 ;Wen, et al., 2023 ) to avoid keyword filters while still sending a negative message to specific groups (ElSherief, et al., 2021 ). This can cause repeated psychological harm and strengthen systemic discrimination, making IHS a covert form of cyberbullying. The old “detect-and-delete” method has clear limits. Automated systems can wrongly block fair social criticism (Johnson, et al., 2019 ). Human review cannot keep up with the speed and size of online content. As a result, researchers have started using AI to reshape IHS. The goal is to balance free speech with community safety. Early methods changed word embeddings (Bolukbasi, et al., 2016 ) or hidden states in models (Zhao, et al., 2018 ). Later, the “detect-and-rewrite” method appeared (He, et al., 2021;Pryzant, et al., 2020 ). In recent years, large language models (LLMs) have shown strong skills in text style transfer (TST). They can change offensive text into neutral text by replacing aggressive words or changing sentence structure, while keeping the main meaning (Hee, et al., 2023 ;Pryzant, et al., 2020 ;Roy, et al., 2024 ). However, most research looks only at explicit hate speech. The ability and risks of handling IHS are not well studied. The need for such study comes from both the special nature of IHS and the complexity of LLMs. First, IHS depends on culture and subtle language cues (Zhang, et al., 2024 ), which are hard for models to read. Second, LLMs may keep hidden biases from training data, leading to covert toxic content. Third, there is no full framework to measure both the effect and the risks of reshaping IHS (Saakyan and Muresan, 2023 ). Solving these problems needs a mix of linguistics, computing, and ethics. Accordingly, this study seeks to address the following core research questions. RQ1:Can advanced LLMs reshape IHS with high quality? LLMs work well for explicit hate speech (Kostiuk, et al.), but for IHS, can they reduce hate while keeping meaning?Accordingly, this study seeks to address the following overarching research question: How effectively can advanced LLMs reshape implicit hate speech while preserving meaning, through which interpretable linguistic mechanisms, and with what constructive and residual risks? To operationalize this overarching question, we adopt a three-step research approach. First, we assess reshaping effectiveness by building a predictive model that quantifies changes in toxicity and semantic preservation, thereby testing whether LLMs can reduce implicit hate while maintaining core meaning. Second, we construct an interpretable linguistic-behavior framework that integrates theoretically grounded linguistic features with explainable AI to identify key language cues, enabling fine-grained measurement of reshaping strategies beyond single-score evaluations. Third, we apply this framework to pattern analysis, comparing texts before and after reshaping to uncover both constructive strategies (e.g., de-targeting, tone softening) and residual risks (e.g., obfuscation, risk displacement) in LLM-mediated transformations. To achieve these objectives, We use an integrative modelling method from computational social science (Hofman, et al., 2021 ).Using a large IHS dataset (ElSherief, et al., 2021 ), we run three stages: predictive modelling, interpretive modelling, and experimental testing. The results give both theory and practice for safe and effective AI governance. Literature Review IHS: Definition, Linguistic Features, and Offensive Strategies IHS means language that avoids open insults but still sends discrimination or hostility through soft, indirect, or coded words, such as euphemism, metaphor, or sarcasm (China, 2019 ;Ghosh, et al., 2023 ;Lin, 2022 ;Nations, 2019 ) Compared with explicit hate speech, IHS hides its attack but aims to harm the same protected groups, such as by race or ethnicity (Merriam-Webster, 2022 ). This harm is often linked to shared culture or history. One key strategy is hiding the meaning and making it more extreme. Offenders use “dog whistles” or coded words (Risius, 2024) to hide extreme ideas in harmless-looking terms, such as using a metaphor to compare people to “pests” (ElSherief, et al., 2021 ). Spotting these words is hard (Magu and Luo, 2018 ). A second is hiding emotion. Offenders may use irony, sarcasm, or “jokes” (China, 2019 ;Hee, et al., 2023 ) to cover hostility, giving them a way to deny the intent (ElSherief, et al., 2021 ). A third is bias by suggestion. Instead of direct attack, they cause group conflict using negative stereotypes (Council of the European Union, 2022;Sap, et al., 2019 ) or by telling stories about unfair treatment or being victims. These work best when the audience shares certain culture or knowledge(Ghosh, et al., 2023 ;Schmeisser-Nieto, et al., 2022). Because these strategies are indirect, IHS can pass content checks and is hard for machines to detect. Current research is still small (Nations, 2019 ). Methods have gone from static word embeddings (Magu and Luo, 2018 ) to context-aware models (Ghosh, et al., 2023 ;Kim, et al., 2022 ). But these models are often “black boxes” and can be tricked by adversarial inputs (Mathew, et al., 2021 ). This is why more work is now on making clear, explainable systems. Governing IHS: From Moderation to Reshaping The governance of IHS is undergoing a shift. Traditional keyword- or rule-based moderation has shown limited results because it cannot handle the core strategies of IHS. It struggles to detect “dog whistle” terms (Marten Risius, 2024 ) hidden through linguistic masking and extremisation, and it fails to understand biased suggestions in social interactions that depend heavily on shared cultural and social backgrounds (Ghosh, et al., 2023 ;Schmeisser-Nieto, et al., 2022). These structural gaps create systematic blind spots in content review(Hartmann, et al., 2025 ). As a result, governance is moving towards constructive interventions centred on “speech reshaping.” The goal is no longer simple deletion, but the use of TST techniques (Hu, et al., 2022 ;Johnson, et al., 2019 ) to directly address the linguistic strategies that harm specific groups. This can include reshaping to remove hostility from implicit emotional expressions—such as softening sarcasm—or reconstructing sentences to break biased suggestions in social interactions, for example by reducing stereotypes, while keeping the non-hateful core meaning (Pryzant, et al., 2020 ). Given the complexity of geopolitical and cultural contexts, banning IHS through a single set of rules is not only impractical but may even backfire (Johnson, et al., 2019 ). Proactive reshaping through TST has therefore become a viable new approach (Hu, et al., 2022 ;Jin, et al., 2022). This strategy aims to turn biased content into neutral expressions while preserving the original intent. Within this approach, in addition to strategies such as “counter speech” (Cepollaro, et al., 2023 ) that seek to change the overall discourse environment, the use of LLMs for “detoxification” (He, et al., 2021;Pryzant, et al., 2020 ) has become a key technology. However, current research almost entirely focuses on explicit hate speech (Kostiuk, et al.;Roy, et al., 2024 ), while work on reshaping IHS is still at an early stage. Common issues include loss of semantic coherence and insufficient modelling of context (Johnson, et al., 2019 ). This makes the systematic evaluation and improvement of LLM performance in IHS reshaping an urgent research gap to address. LLMs and Their Evaluation Methods The new generation of LLMs, represented by the GPT series (Achiam, et al., 2023 ) and DeepSeek (Guo, et al., 2025 ), is characterised by emergent abilities. These abilities arise from the massive expansion of model parameters and training data, enabling advanced functions—such as complex reasoning and deep contextual understanding—that were absent in smaller-scale models (Achiam, et al., 2023 ). Such capabilities provide the technical foundation for addressing culturally sensitive tasks like IHS.Evaluating these complex models requires multi-dimensional approaches. These include human evaluation, which is the gold standard but costly, and computational evaluation, which is more scalable but relies on automated metrics. Traditional automated metrics, however, often fail to capture the deeper quality of text, especially whether an LLM truly understands and neutralises the complex strategies of IHS. For example, a model might produce fluent and harmless-looking text, but the metrics cannot tell whether it has identified the real intent hidden behind linguistic masking, or whether it has simply imitated a harmless surface style. To address these gaps, integrated evaluation methods have emerged—such as using one LLM to assess the output of another. In the context of IHS, an ideal evaluation system must go beyond “black-box” classification (Mathew, et al., 2021 ) and be able to interpret the socio-cultural meanings encoded in language (Ghosh, et al., 2023 ;Lin, 2022 ). This means that LLMs are not only the subject of evaluation but can also serve as tools for building interpretable evaluation frameworks.A human-like computational evaluation system should have the capacity to recognise the strategies underlying IHS. It should not only determine whether a text ultimately harms a specific group but also explain how that harm is conveyed—whether through implicit emotional expression or biased suggestions. In this way, evaluation moves beyond “black-box” judgements (Mathew, et al., 2021 ) and achieves genuine interpretability. Formulation of Research Questions With the rise of online interaction, IHS has become a major barrier to constructive dialogue in digital spaces. Existing studies show that IHS spreads hostility through complex linguistic strategies, such as coded language, sarcasm, negative stereotypes, and dehumanising metaphors (Lemmens, et al., 2021 ;Wiegand, et al., 2023 ), making it more difficult to detect and manage.In this context, using generative artificial intelligence to “reshape” or “detoxify” such language (Pryzant, et al., 2020 ) has emerged as a cutting-edge approach to promoting a more civil online environment. However, most detoxification research has focused on explicit hate speech (Kostiuk, et al.;Roy, et al., 2024 ). Assessing the capabilities of LLM in the more complex task of reshaping IHS remains a clear research gap(Johnson, et al., 2019 ).Because the offensiveness of IHS is embedded in its linguistic structure, an effective evaluation system must examine whether an LLM can recognise and neutralise these underlying harmful strategies. To address this, the present study proposes a dual evaluation framework that combines external indicators (reshaping outcomes) with internal indicators (treatment of linguistic strategies). The internal framework focuses on four core operational dimensions:group-specific harm, implicit emotional expression, linguistic obfuscation and extremity, and bias and implication in social interaction. This design enables a systematic assessment of the capabilities, strategies, and behaviours of LLM in reshaping IHS. Accordingly, this study first examines whether LLM can effectively achieve a high success rate in reshaping IHS. To address this, an AI-based predictive model will be developed, following the text detoxification paradigm(Yuan, et al., 2025 ), to conduct external evaluation across two dimensions: hate mitigation and semantic preservation.Second, the study investigates how to construct a comprehensive evaluation framework that integrates the core linguistic strategies of IHS for internal assessment of reshaping performance. IHS will be deconstructed into four strategic categories—group-specific harm, implicit emotional expression, linguistic obfuscation and extremity, and bias and implication in social interaction. Using explainable AI methods, the study will identify key linguistic indicators closely associated with these strategies, thereby building a computational internal evaluation framework.Finally, this framework will be applied to examine whether, in the process of reshaping IHS, LLM can effectively neutralise the strategies found in human-generated IHS. Comparative analysis of texts before and after reshaping will be conducted to explore the extent to which these strategies and behaviours are mitigated through language transformation. Methods and Experiments This study adopts the integrative modelling approach from computational social science (Hofman, et al., 2021 ) to address the above research questions. This approach combines predictive and interpretive modelling, aiming to achieve both accurate forecasting of phenomena and in-depth explanation of their underlying mechanisms. In this research, we use machine learning and deep learning algorithms to build predictive models, integrate interpretability techniques to explore underlying mechanisms, and apply a systematic experimental design to validate the findings. Specifically, based on the high-quality Latent Hatred benchmark dataset (Section 3.1), we follow a three-stage research process to systematically assess the overall performance of LLMs in reshaping IHS. The process is illustrated in Fig. 1 . Predictive modelling stage : Build and train a high-performance deep learning model to enable automated and large-scale evaluation of LLM reshaping outcomes.Interpretive modelling stage : Apply interpretability methods to identify and quantify key linguistic cues influencing IHS classification, thereby constructing a computational framework for evaluating LLM linguistic behaviours.Experimental validation stage: Conduct dual verification and in-depth analysis of LLM reshaping performance using multiple indicators, including toxicity reduction, semantic retention, and the linguistic cues identified above. Dataset This study employed the Latent Hatred benchmark dataset developed by ElSherief et al. (ElSherief, et al., 2021 ). The dataset was collected from the Twitter and contains a total of 27,157 tweets, each annotated for hate speech category (implicit hate, explicit hate, and non-hate). Among these, 12,671 are IHS texts (including 7,100 unique tweets). In addition, the dataset includes annotations for the target group and specific type of hate speech, covering six categories: incitement, inferiority, irony, stereotypical, threatening, and white grievance. For the reshaping experiments, all IHS samples were selected, and two representative LLMs were used to perform text reshaping: GPT-4o, released by OpenAI in May 2024, and Deepseek-V2, released by DeepSeek in May 2024. All reshaped texts corresponding to the IHS tweets were obtained via the official APIs. Detailed statistical information on the dataset is provided in Table 1 . Table 1 Statistical Information of the Dataset Statistics Tweet Length Length of IHS Tweets Deepseek Reshaped Text Length GPT-4o Reshaped Text Length Count 27,157 12,035 7,100 7,100 Mean 90.545 31.299 160.878 126.684 Standard deviation 41.493 11.446 69.562 42.005 Minimum 4 6 6 14 25th percentile 63 24 114 97 Median 90 29 147 122 75th percentile 113 36 192 151 Maximum 801 186 572 418 Predictive Modelling: Automated Evaluation of LLM Reshaping Capability To quantitatively assess the effectiveness of LLMs in reshaping IHS, this stage aims to develop an automated classification model. The core task of this model is to determine the final speech category of the text reshaped by LLMs (e.g., “non-hate” or “still implicit hate”). In model construction, a comparative experimental design was adopted. First, at the feature engineering stage, input features were prepared for two types of algorithms: (1) psycho-linguistic features extracted using the LIWC tool for training traditional machine learning models; and (2) word embedding vectors generated by the RoBERTa model as input for deep learning models. Second, in algorithm selection, a wide range of models was examined, including traditional machine learning algorithms such as Logistic Regression (LR), Support Vector Machine (SVM), Random Forest (RF), and XGBoost, as well as multiple deep learning architectures such as RoBERTa-BiLSTM-Attention. The performance of all candidate models was comprehensively evaluated using multiple metrics, including accuracy, precision, recall, F1-score, and ROC AUC, with cross-validation employed to ensure generalisation capability. In addition to classification performance, another key indicator—semantic retention—was also assessed. This study applied the BERTScore method based on RoBERTa-large to compute the F1 semantic similarity between the original and reshaped texts, thereby quantifying the extent to which LLMs preserved the core semantics during the reshaping process. Interpretive Modelling: Construction of the IHS Linguistic Behaviour Evaluation Framework The objective of this stage is to move beyond a surface-level assessment of “whether it works” to explore “why it works” and “where potential risks lie.” This entails constructing an interpretable framework for evaluating linguistic behaviours in IHS detection. The framework aims to identify key linguistic cues that play a decisive role in determining IHS and, based on these cues, to evaluate the specific linguistic operations employed by GPT-4o and Deepseek during the reshaping process. The construction process was as follows. First, with IHS category (yes/no) as the prediction target and over one hundred linguistic features extracted via LIWC as input variables, a series of machine learning classifiers were developed. Second, to select the most influential linguistic cues from high-dimensional features, a systematic feature selection procedure was designed. We initially applied the intrinsic feature importance ranking of the XGBoost model to conduct a coarse-grained filtering, removing low-contribution redundant features. Subsequently, on the resulting candidate subset, we employed a stepwise forward selection strategy, starting from an empty set and iteratively adding the feature that yielded the greatest improvement in model performance, as measured by F1-score and ROC AUC, until no significant performance gain was observed. This process ensured that the final feature subset was both parsimonious and effective. Finally, for the optimal classification model built on this refined feature subset—determined through grid search and cross-validation—the SHAP (Shapley Additive Explanations) method was applied for in-depth interpretation. SHAP quantifies both the magnitude and the direction (positive/negative) of each linguistic cue’s contribution to every individual prediction, enabling precise identification of the core linguistic cues most influential in IHS determination. These identified cues, together with their association patterns with IHS, collectively constitute the proposed IHS linguistic behaviour evaluation framework. Framework Validation: Dual Evaluation of LLM Linguistic Behaviour To validate the effectiveness and practical utility of the linguistic behaviour evaluation framework developed in the previous stage, we conducted a dual evaluation. The core objective was to examine whether the linguistic changes captured by the framework aligned with externally perceived changes in textual risk. Specifically, for both the original and reshaped texts generated by GPT-4o and Deepseek, we conducted two complementary assessments. First, an internal linguistic feature analysis was carried out using the proposed evaluation framework to measure the average percentage change in key linguistic cues (e.g., sentiment polarity, frequency of specific lexical items) before and after reshaping, thereby revealing what the LLMs did at the linguistic level. Second, an external risk metric assessment was performed using established third-party tools (e.g., Perspective API) to calculate the average percentage change in toxicity scores, reflecting the actual impact of reshaping on external perceptions of risk. By comparing the results from internal and external analyses, we were able to verify whether the linguistic behaviours identified by the framework were indeed the primary drivers of reduced hateful attributes in the text. This not only confirmed the validity of the framework but also provided robust evidence for a deeper understanding of the mechanisms underlying LLM reshaping. Results Predictive Modelling Analysis: Evaluating LLM Reshaping Capability Performance Comparison of Automated Evaluation Models The comparative analysis produced several key findings(Table 2 ) .First, among the deep learning models using RoBERTa word embedding features, the RoBERTa-BiLSTM-Attention model achieved the best overall performance, with both accuracy and ROC AUC reaching 0.864. Notably, although the XLNet model had a slightly lower accuracy (0.853), it achieved the highest recall (0.901), indicating stronger ability in identifying positive cases.Secondly, for traditional machine learning models based on LIWC psycho-linguistic features, ensemble learning algorithms performed best. In particular, XGBoost outperformed other models in ROC AUC (0.838), recall (0.713), and F1 score (0.741), showing the strongest discriminative ability, while Random Forest achieved the highest accuracy (0.756).Finally, when combining LIWC and RoBERTa features, the effects varied by model type. For traditional machine learning models, this combination significantly improved performance limits. For example, the XGBoost model’s ROC AUC increased to 0.895, while its accuracy and F1 score improved by 7.7% and 7.3%, respectively, compared with using LIWC features alone. However, this positive effect did not appear in deep learning models, suggesting that for models already capable of strong semantic representation, adding LIWC features brought no further gains. Table 2 Performance of Different Classification Algorithms Features Algorithm Accuracy Precision Recall F1 ROC AUC LIWC features Logistic Regression 0.738 0.747 0.711 0.729 0.807 Support Vector Machine 0.751 0.824 0.632 0.715 0.819 Random Forest 0.756 0.821 0.649 0.725 0.834 XGBoost 0.753 0.770 0.713 0.741 0.838 RoBERTa word embedding features RoBERTa 0.858 0.840 0.885 0.862 0.858 RoBERTa-LSTM-Attention 0.862 0.845 0.886 0.865 0.862 XLNet 0.853 0.822 0.901 0.850 0.853 RoBERTa-BiLSTM 0.860 0.843 0.885 0.864 0.860 RoBERTa-CNN 0.861 0.846 0.882 0.864 0.861 RoBERTa-BiLSTM-Attention 0.864 0.850 0.883 0.866 0.864 Combined LIWC and RoBERTa word embedding features Logistic Regression 0.761 0.787 0.708 0.746 0.833 Support Vector Machine 0.752 0.829 0.630 0.716 0.820 Random Forest 0.785 0.905 0.632 0.744 0.865 XGBoost 0.811 0.859 0.740 0.795 0.895 RoBERTa 0.795 0.798 0.795 0.795 0.795 RoBERTa-LSTM-Attention 0.805 0.813 0.804 0.804 0.804 RoBERTa-BiLSTM 0.798 0.815 0.796 0.794 0.796 RoBERTa-BiLSTM-Attention 0.802 0.805 0.802 0.801 0.802 Evaluation of LLM Reshaping Performance Based on the Optimal Model Based on the above comparison, this study selected the RoBERTa-BiLSTM-Attention model—identified as having the best overall performance—as the final automated evaluation tool to determine whether a text reshaped by an LLM still belongs to IHS. To conduct a rigorous statistical test, we applied the McNemar test to examine whether the distribution of IHS categories for the same set of tweets changed significantly before and after LLM reshaping. The results are shown in Table 3 . Table 3 Paired-Sample Chi-Square Test of the Number of Aggressive Items Between OL and OL AUTO-Cot, and Semantic Preservation Results LLM Topic Type Prompt Type No-Hate Count Hate Count Success Rate of No-Hate Conversion (%) p-value Semantic Preservation Deepseek Incitement OL 0 1240 0.000 0.832 OL AUTO-Cot 1235 5 99.60 Inferiority OL 0 859 0.000 0.834 OL AUTO-Cot 849 10 98.84 Irony OL 0 794 0.000 0.853 OL AUTO-Cot 750 44 94.41 Other OL 0 79 0.000 0.842 OL AUTO-Cot 76 3 96.20 Stereotypical OL 0 1104 0.000 0.829 OL AUTO-Cot 1095 9 99.19 Threatening OL 0 666 0.000 0.822 OL AUTO-Cot 661 5 99.25 White grievance OL 0 1506 0.000 0.829 OL AUTO-Cot 1502 4 99.73 GPT-4o Incitement OL 0 1240 0.000 0.829 OL AUTO-Cot 1238 2 99.84 Inferiority OL 0 859 0.000 0.831 OL AUTO-Cot 853 6 99.30 Irony OL 0 794 0.000 0.842 OL AUTO-Cot 775 19 97.60 Other OL 0 79 0.000 0.832 OL AUTO-Cot 79 0 100 Stereotypical OL 0 1104 0.000 0.823 OL AUTO-Cot 1098 6 99.46 Threatening OL 0 666 0.000 0.823 OL AUTO-Cot 664 2 99.70 White grievance OL 0 1506 0.000 0.823 OL AUTO-Cot 1503 3 99.80 The statistical results clearly indicate that, whether reshaped by Deepseek or GPT-4o, the distribution of all IHS topic types (e.g., incitement, inferiority, irony) showed highly significant changes before and after reshaping (all p < 0.001). This finding confirms that LLMs can effectively and significantly reduce implicit hate attributes in text.Alongside evaluating reshaping effectiveness, we also examined semantic preservation. The results show that both LLMs achieved a semantic preservation rate of 82%–85% (measured by the BERTScore F1 value) when performing reshaping tasks. This means that LLMs can remove implicit bias while maintaining the core meaning of the original text. Taken together, these findings provide clear empirical support for our first research question (RQ1). Interpretability Modelling Analysis: Development of the IHS Linguistic Behaviour Evaluation Framework To gain deeper insight into the linguistic behaviours of LLMs when reshaping IHS, this study developed an interpretable classification model designed to identify the key linguistic cues that determine IHS. We first evaluated the performance of several candidate classification models (see Table 4 ). The results showed that the XGBClassifier achieved the highest accuracy (0.714) and precision (0.722), making it the optimal model. Accordingly, this model was selected as the basis for the subsequent in-depth interpretability analysis. Table 4 Comparison of Different Classification Models in Assessing IHS Reshaping Performance Algorithm Accuracy Precision Recall F1 ROC AUC Logistic Regression 0.699 0.703 0.940 0.803 0.594 Ridge Regression 0.699 0.701 0.944 0.804 0.592 Support Vector Classifier 0.695 0.687 0.987 0.809 0.568 Multi-layer Perceptron 0.686 0.735 0.816 0.773 0.630 Random Forest Classifier 0.710 0.706 0.959 0.812 0.601 XGBoost Classifier 0.714 0.722 0.918 0.808 0.626 We applied the SHAP (Shapley Additive Explanations) method to the optimal XGBClassifier model to conduct a global interpretability analysis, aiming to evaluate how different linguistic features influence IHS classification predictions. Figure 2 presents the top 25 most influential linguistic cues identified by the model, along with their overall contribution. To further examine the direction and magnitude of each cue’s impact, we plotted a SHAP summary diagram (see Fig. 3 ). This diagram shows that these cues influence IHS classification mainly in two opposite ways: Positively correlated cues (increasing the probability of being classified as IHS):The results indicate that when the text contains more words referring to specific groups—such as ethnicity, visual, body, religion, the pronoun “they,” and informal pronouns (ipron)—as well as more adjectives (adj), exclamation marks (Exclam), emotional words (Affect), absolute terms (allnone), and power-related words (power), the probability of the text being classified as IHS increases. This suggests that these linguistic cues play a key role in driving the expression of implicit hate. Negatively correlated cues (decreasing the probability of being classified as IHS):In contrast, when the text contains more features related to tone (Tone), prepositions (prep), past tense (focuspast), social behaviour (socbehav), communication (comm), and work-related terms (work), the probability of the text being classified as IHS decreases. This indicates that the presence of these cues often signals a shift towards more neutral or constructive content. To systematise the above quantitative results, this study developed an IHS speech behaviour evaluation framework (Table 5 ) based on SHAP analysis and relevant literature. Instead of merely listing linguistic cues, the framework organises them into four broader theoretical dimensions, as shown in Table 6 . The first dimension, group-specific harm, relates to references and descriptions of specific groups, such as “ethnicity” and “visual features”. The second, implicit emotional expression, covers cues reflecting the intensity and polarity of emotion, with “adjectives” and “exclamation marks” as positive indicators and “tone” as a negative indicator. The third, linguistic obfuscation and extremity, involves strategies that make statements more vague or absolute, such as “commas” and “all-or-none terms”. The fourth, bias and implication in social interaction, captures potential bias in describing social interactions, such as “motivation words” and “power words”. This framework addresses RQ2 by not only identifying key linguistic behaviours influencing IHS but also providing a quantifiable, multi-dimensional system for analysing and assessing the specific language operations applied by LLMs during the reshaping process. Table 5 Analytical framework for IHS linguistic behaviours in online Concept Definition Application example Applicable factors Group-specific harm Sentences imply discriminatory expressions toward specific groups by referring to protected attributes such as race, religion, appearance, and physical traits, often through vague references. Understand the groups, references, and related historical or social background in the text; avoid negative or discriminatory statements; use positive or neutral expressions for protected traits like race, religion, or appearance. Protected traits, historical and social background Implicit emotional expression Implicitly convey aggressive emotions through the use of emotional words, exclamation marks, and other punctuation. Use emotional words and punctuation to judge emotional tone, identify inflammatory remarks, and replace them with positive or neutral words to communicate inclusively. Emotional words, punctuation Linguistic obfuscation and extremity Convey extreme positions or hide true intentions implicitly through complex sentence structures, euphemisms, punctuation use, and absolute terms. Identify unreasonable elements in text semantics or structure, as well as exaggerated, euphemistic, vague, or metaphorical expressions; ensure language is clear, direct, and inclusive. Sentence structure, punctuation use, absolute terms Bias and implication in social interaction Implicitly convey bias or negative attitudes toward specific groups through words related to social behaviours and communication. Use positive or neutral language in communication, social behaviour, and work discussions, considering social context and background knowledge, to avoid conflict, inequality, and negative stereotypes. Social behaviours, communication words Table 6 Comprehensive evaluation framework for IHS Concept Rank Category in LIWC Internal indicator Importance Relationship between indicator and IHS Group-specific harm 2 Culture Ethnicity 0.307 Positive 3 Perception Visual 0.14 Positive 4 Physical Physical 0.129 Positive 5 Pronouns They 0.115 Positive 6 Culture Relig 0.106 Positive 17 Linguistic Dimensions Ipron 0.058 Positive 24 Perception Perception 0.045 Positive Implicit emotional expression 16 Linguistic Dimensions Adj 0.064 Positive 12 Punctuation Exclam 0.08 Positive 13 Psychological Processes Affect 0.075 Positive 19 Punctuation AllPunc 0.052 Positive 23 Summary Variables Tone 0.046 Negative Linguistic obfuscation and extremity 1 Punctuation Comma 0.366 Positive 14 Psychological Processes Discrep 0.068 Positive 20 Linguistic Dimensions Prep 0.052 Negative 21 Psychological Processes Allnone 0.05 Positive Bias and implication in social interaction 10 Time Orientation Focuspast 0.085 Negative 11 Psychological Processes Drives 0.085 Positive 9 Social Processes Socbehav 0.093 Negative 15 Social Processes Comm 0.065 Negative 22 Lifestyle Work 0.048 Negative 25 Psychological Processes Power 0.035 Positive Uncategorised words 7 Summary Variables Wc 0.106 Positive 8 Summary Variables Dic 0.1 Negative 18 Cognition Space 0.057 Positive Validation of the comprehensive evaluation framework based on intrinsic indicators To closely examine the language behaviour of LLMs during the reshaping process, this study applied the previously established evaluation framework to compare the mean values of each linguistic cue (intrinsic indicator) between the original IHS tweets and the texts reshaped by LLMs (Deepseek and GPT-4o). This analysis aimed to reveal the specific language strategies used to reduce hateful content, the commonalities and differences between models, and potential limitations. Detailed results are shown in Table 7 . The findings show that the models adopted a “de-targeting” strategy, reducing the frequency of group-specific words and the plural pronoun “they” to weaken attack targets, while increasing non-personal pronouns linked to bias, which dissolved explicit targeting but left bias residue. In the emotional dimension, they reduced exclamation marks and increased neutral words to lower emotional intensity, yet their internal logic for achieving this differed, leading to conflicting strategies. For language style, the models reduced absolutist words and increased prepositions, which eased extreme expression, but their “increased complexity” approach—raising the use of commas and discrepancy words—unintentionally heightened concealment, resulting in eased extremes but stronger masking. In the social interaction dimension, the models increased constructive words and reduced power-related words to lower social aggressiveness. However, their key “dilution” strategy—expanding total text length and vocabulary to weaken hateful content—may indicate a risk displacement effect, as text length is also linked to hate speech detection, meaning reduced aggression but shifted risk. These results answer RQ3, showing that LLMs can reshape IHS-related strategies, but the reshaping is not a full “removal” and instead resembles a complex “replacement”. The models blunt the sharpest edges of hate speech, but at the cost of introducing more ambiguous and covert language risks. Table 7 Mean Differences in Linguistic Cues Between IHS and AUTO Concept Rank Category in LIWC Internal indicator Importance Relationship between indicator and IHS Mean Percentage Difference (AUTO-CoT, Deepseek) Mean Percentage Difference (AUTO-CoT, GPT-4o) Group-specific harm 2 Culture Ethnicity 2 Positive -79.939 -82.555 3 Perception Visual 3 Positive -76.912 -79.172 4 Physical Physical 4 Positive -74.291 -74.685 5 Pronouns They 5 Positive -39.022 -46.710 6 Culture Relig 6 Positive -74.081 -77.541 17 Linguistic Dimensions Ipron 17 Positive 62.156 37.993 24 Perception Perception 24 Positive -28.066 -17.265 Implicit emotional expression 16 Linguistic Dimensions Adj 16 Positive -6.249 7.998 12 Punctuation Exclam 12 Positive -92.396 -89.045 13 Psychological Processes Affect 13 Positive 4.243 -4.022 19 Punctuation AllPunc 19 Positive -36.087 -38.235 23 Summary Variables Tone 23 Negative 133.165 147.157 1 Punctuation Comma 1 Positive 182.547 129.378 14 Psychological Processes Discrep 14 Positive 3.542 10.839 20 Linguistic Dimensions Prep 20 Negative 32.463 31.673 21 Psychological Processes Allnone 21 Positive -13.646 -24.644 Linguistic obfuscation and extremity 10 Time Orientation Focuspast 10 Negative -11.693 -3.291 11 Psychological Processes Drives 11 Positive 28.233 25.806 9 Social Processes Socbehav 9 Negative 36.467 35.350 15 Social Processes Comm 15 Negative 35.618 62.061 22 Lifestyle Work 22 Negative 70.342 71.093 25 Psychological Processes Power 25 Positive -24.599 -24.369 Bias and implication in social interaction 7 Summary Variables Wc 7 Positive 49.631 15.672 8 Summary Variables Dic 8 Negative 1.870 2.289 18 Cognition Space 18 Positive -21.713 2.754 Note:(Average percentage differences in linguistic cue statistics between AUTO-CoT and the original text are shown. Positive cues increase IHS aggressiveness, while Negative cues reduce aggressiveness. A positive value means AUTO-CoT used the feature more frequently; a negative value means the original text used it more frequently.) Result Validation: External Validity Test Based on Toxicity Scores We calculated the Toxicity Score for the original IHS tweets and their LLM-reformed versions, and conducted paired-sample t-tests and Tukey HSD tests to analyse score differences. The overall results (see Table 8 and Fig. 4 ) show that LLM reformulation significantly reduced textual toxicity. Compared with the original tweets, Deepseek reduced toxicity by an average of 86.21%, while GPT-4o achieved a reduction of 90.57%. The differences in toxicity scores for both models reached a highly significant statistical level (p < 0.001). In terms of distribution, the median toxicity scores and overall score ranges of the reformed texts showed a clear downward trend. Table 8 Average Percentage Difference in Toxicity Scores and Significance Tests: OL AUTO-CoT vs. OL LLM Average Percentage Difference* t-value p-value Deepseek -86.210 166.082 0.000 GPT-4o -90.568 174.814 0.000 To further examine model performance across different contexts, this study repeated the toxicity score comparison for each IHS topic category (Table 9 and Fig. 5 ). The analysis shows that in all topic categories, LLMs exhibited strong toxicity reduction, with average decreases ranging from 81.64% to 94.7%. Notably, both models performed exceptionally well in high-risk categories, but with some performance differences. GPT-4o achieved the greatest reduction in Threatening (− 94.7%),Derogation (− 92.7%), and Stereotypical (− 90.1%) categories. DeepSeek also showed strong results in Threatening (− 93.7%) and Derogation (− 90.4%). Overall, under the AUTO-CoT prompting framework, GPT-4o generally outperformed DeepSeek, particularly in handling high-risk, highly aggressive IHS. Table 9 Mean Percentage Differences in Toxicity Reduction Effects of LLM–Prompt Combinations Across Different Topic Types Topic Type Mean Percentage Difference (AUTO-CoT, Deepseek) (AUTO-CoT, GPT-4o) Incitement -86.657 -90.705 Inferiority -90.392 -92.681 Irony -85.983 -88.421 Other -81.636 -89.639 Stereotypical -85.06 -90.118 Threatening -93.6651 -94.654 White grievance -83.719 -89.948 Discussion This study examines the ability, strategies, and potential limitations of LLMs in reshaping online IHS. The results confirm that LLMs can reshape IHS. However, this process is not a simple purification, but rather a complex displacement of risk. Effectiveness and Complexity of LLMs as Tools for Governing IHS This study first addresses RQ1 by confirming that LLMs, represented by GPT-4o and DeepSeek, can effectively reduce the hateful attributes of IHS while largely preserving the core semantics of the original text. This finding provides empirical support for adopting “speech reshaping” as a constructive intervention, surpassing the systemic limitations of traditional moderation when addressing core IHS strategies such as linguistic obfuscation and bias cues in social interaction (Hartmann, et al., 2025 ). The core contribution, however, lies in revealing the underlying mechanisms of this reshaping process. By constructing and applying the IHS linguistic behavior evaluation framework (RQ2), this study examined the specific linguistic operations performed by LLMs, thereby answering RQ3. The analysis shows that LLM reshaping is not a complete elimination but a complex displacement. The models successfully blunt the most direct attacks in hate speech, but at the cost of introducing new, more covert linguistic risks. This suggests that current LLM handling of IHS is closer to pattern-based risk avoidance than to value alignment grounded in deep understanding. The models learn to identify and replace language features strongly associated with hate labels, yet fail to fully grasp the socio-cultural connotations and bias roots behind these features (Ghosh, et al., 2023 ;Lin, 2022 ) . Four Types of Language Changes in the Reshaping Process This study found four kinds of replacement behaviors that show the inner limits of LLM reshaping strategies. First, target removal but bias remains. The models cut direct mentions of certain groups, an effective “de-targeting” move. But they added impersonal pronouns linked to bias, turning clear targeting into vague bias hints. This may lower scores in automatic checks, but may not remove the bias that readers can still understand in context (Ghosh, et al., 2023 ;Schmeisser-Nieto, et al., 2022) . Second, lower intensity but mixed methods. The models reduced exclamation marks and used more neutral words to soften tone. Yet they used different and sometimes opposite ways to do this. This shows that emotion control is still at a surface level, without a steady and clear method, mostly just cutting visible emotion signs. Third, less extreme but more hidden. The models used fewer absolute words, making text less extreme. But their “safe output” method of adding complexity—more commas and mismatch words—made the message more hidden. This is like “dog-whistle” or coded speech (Marten Risius, 2024 ), where complex form hides true meaning. Fourth, weaker attacks but risk moved. The models used more positive words and fewer power words to cut social attack. But their main “dilution” method—adding length and extra words—may create new risks. Longer text can also link to hate speech labels. This “padding” may be a hidden form of risk change. Theoretical and Practical Implications Theoretically, this study builds and applies an IHS Speech Behavior Assessment Framework to reveal the inner tensions in LLMs’“detoxification” process. It introduces the concept of risk displacement, challenging the simplified view of this process as a one-way purification. By quantifying and operationalizing the complex attack strategies of IHS (China, 2019 ;Ghosh, et al., 2023 ;Lemmens, et al., 2021 ), the framework offers a replicable method for opening the “black box” of LLM decision-making (Mathew, et al., 2021 ) and for analyzing their specific behaviors when handling socially sensitive language tasks, thereby advancing research in explainable AI governance. Practically, the study provides empirical support for shifting online content governance from traditional “censorship” to a restorative paradigm of “speech reshaping” (Hartmann, et al., 2025 ;Johnson, et al., 2019 ). More importantly, the identified language replacement paths—such as “target removal but bias remains” and “less extreme but more hidden”—serve as a direct audit checklist for platforms and developers. These findings call for governance to go beyond simple technical deployment toward fine-grained management that incorporates continuous risk monitoring and secondary review, offering clear guidance for the responsible use of this technology. Limitations and Future Directions This study has several limitations. First, it only covers two specific LLMs, so the findings may not apply to other models. Second, the dataset is based mainly on English tweets. Future work should test the framework on more languages and different types of online texts to check its validity across cultures and platforms. Future research can take several paths. One is to study how these “replaced” language risks affect real human perception. Does reshaped, more vague text lower readers’ guard and spread bias in subtle ways? Another is to develop new training or fine-tuning methods to solve the problem of risk displacement at its root, guiding models to remove hate without adding new risks. A third is to combine the framework with cognitive experiments to better understand the social and psychological mechanisms behind IHS language strategies. Ethical Risks and Potential Biases Using generative AI to reshape IHS raises serious ethical concerns. The main risk is that models may inherit and amplify cultural and social biases from their training data (Lin, 2022 ;Yadav and Singh, 2024 ). This can lead to misunderstanding non-mainstream contexts or reproducing stereotypes about marginalized groups, reinforcing inequality. The lack of transparency in LLM decision-making—the “algorithmic black box” (Roychowdhury and Gupta, 2023 ) —also makes it hard for users to know why their speech was changed and what values guided the change. Without transparency and fair appeal mechanisms, such hidden language control can threaten freedom of expression and the right to information. Any purely technical solution must be part of a broader socio-technical governance system that includes ethical review, bias audits, and user rights protection, to prevent solving old problems while creating new and hidden risks. Conclusion This study used an integrative modelling approach to examine the ability and mechanisms of LLMs in reshaping IHS. It first confirmed that LLMs can effectively reduce the hateful features of IHS while keeping its core meaning. However, with the IHS Speech Behavior Assessment Framework developed in this study, the core finding shows that the reshaping process is not a simple removal but a complex risk displacement. While LLMs can weaken direct attacks, they may also introduce new and more hidden risks through strategies like vague targeting and greater language masking. The contributions are both theoretical and practical. Theoretically, the concept of risk displacement deepens understanding of the “detoxification” process, and the framework provides an explainable tool for future research in this area. Practically, the study supports “speech reshaping” as a governance model but also defines its risk boundaries. The identified replacement paths can serve as a direct checklist for technical audits. Future research should go beyond performance evaluation and focus on resolving these replaced language risks, as well as exploring better human–AI collaboration in governance, to align technological development with social values. Declarations Author Contribution Yinghui Huang, Qixia Feng and Hui Liu were responsible for conceptualization, methodology, and writing—original draft preparation. Hui Liu, Weiqing Li, Zongkui Zhou, and Ying Ma contributed to writing—review and editing. Yinghui Huang, Weiqing Li and Ying Ma acquired the funding. All authors have read and agreed to the published version of the manuscript. Data Availability This study uses the publicly available Latent Hatred benchmark dataset (ElSherief et al., 2021), available at: https://www.dropbox.com/scl/fi/76gx54lbv9dnbqr7nmy2c/implicit-hate-corpus.zip?rlkey=qxp26bhci0v8o0tldzxot1ref&e=1&dl=0. The LLM-reshaped implicit hate speech texts generated in this study using GPT-4o and DeepSeek-V2 are available from the corresponding author upon reasonable request. Acknowledgments This research was supported by the National Natural Science Foundation of China (Grant Nos. 72204095 and 72304090), the Humanities and Social Science Young Scientist Program sponsored by the Ministry of Education of the People’s Republic of China (Grant No. 22YJC880022), the Fundamental Research Funds for the Central Universities (Grant No. 2024VA059), and the Ministry of Education of Humanities and Social Science project (Grant No. 23YJAZH098). Artificial Intelligence usage During the preparation of this manuscript, we used an artificial intelligence (AI) tool (ChatGPT, OpenAI) to improve the readability and style of the text. The AI assistance was limited to language polishing (grammar, spelling, punctuation, and tone). All content was generated by the authors, and the final version was thoroughly reviewed and approved by the authors to ensure accuracy and integrity. References Achiam J, Adler S, Agarwal S, Ahmad L, Akkaya I, Aleman FL et al (2023) Gpt-4 technical report. arXiv preprint arXiv:2303.08774 Bolukbasi T, Chang K-W, Zou JY, Saligrama V, Kalai AT (2016) Man is to computer programmer as woman is to homemaker? debiasing word embeddings. Advances in neural information processing systems 29 Cepollaro B, Lepoutre M, Simpson RM (2023) Counterspeech. Philos Compass 18:e12890 China CA (2019) o. Provisions on the Governance of the Online Information Content Ecosystem [网络信息内容生态治理规定] . Institution, Beijing Council of the European Union, E. P (2022) Regulation (EU) 2022/2065 on a Single Market For Digital Services and amending Directive 2000/31/EC (Digital Services Act). Institution ElSherief M, Ziems C, Muchlinski D, Anupindi V, Seybolt J, De Choudhury M et al (2021) Latent hatred: A benchmark for understanding implicit hate speech. arXiv preprint arXiv:2109.05322 Ghosh S, Suri M, Chiniya P, Tyagi U, Kumar S, Manocha D (2023) CoSyn: Detecting implicit hate speech in online conversations using a context synergized hyperbolic network. arXiv preprint arXiv:2303.03387 Guo D, Yang D, Zhang H, Song J, Zhang R, Xu R et al (2025) Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948 Hartmann D, Oueslati A, Staufer D, Pohlmann L, Munzert S, Heuer H (2025) Lost in moderation: How commercial content moderation apis over-and under-moderate group-targeted hate speech and linguistic variations. In: (eds) Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems , vol pp. 1–26 He Z, Majumder BP (2021) Detect and perturb: Neutral rewriting of biased and sensitive text via gradient-based decoding. arXiv preprint arXiv:2109.11708 Hee MS, Chong W-H, Lee RK-W (2023) Decoding the underlying meaning of multimodal hateful memes. arXiv preprint arXiv:2305.17678 Hofman JM, Watts DJ, Athey S, Garip F, Griffiths TL, Kleinberg J et al (2021) Integrating explanation and prediction in computational social science. Nature 595:181–188 Hu Z, Lee RK-W, Aggarwal CC, Zhang A (2022) Text style transfer: A review and experimental evaluation. ACM SIGKDD Explorations Newsl 24:14–45 Johnson NF, Leahy R, Restrepo NJ, Velásquez N, Zheng M, Manrique P et al (2019) Hidden resilience and adaptive dynamics of the global online hate ecology. Nature 573:261–265 Kim Y, Park S, Han Y-S (2022) Generalizable implicit hate speech detection using contrastive learning. In: (eds) Proceedings of the 29th international conference on computational linguistics , vol pp. 6667–6679 Kostiuk Y, Tonja AL, Sidorov G, Kolesnikova O Reframing social media discourse: Converting hate speech to non-hate speech. J Intell Fuzzy Syst : JIFS–219348 League A-D (2024) Online Hate and Harassment: The American Experience 2024 . Report No. 26–42 Lemmens J, Markov I, Daelemans W (2021) Improving hate speech type and target detection with hateful metaphor features. In: (eds) Proceedings of the Fourth Workshop on NLP for Internet Freedom: Censorship, Disinformation, and Propaganda , vol pp. 7–16 Lin J (2022) Leveraging world knowledge in implicit hate speech detection. arXiv preprint arXiv:2212.14100 Magu R, Luo J (2018) Determining code words in euphemistic hate speech using word embedding networks. In: (eds) Proceedings of the 2nd workshop on abusive language online (ALW2) , vol pp. 93–100 Marten Risius MN (2024) Substitution: Extremists’ New Form of Implicit Hate Speech to Avoid Detection. Saeed Akhlaghpour and Hetiao (Slim) Xie. Global Network on Extremism & Technology Mathew B, Saha P, Yimam SM, Biemann C, Goyal P, Mukherjee A (2021) Hatexplain: A benchmark dataset for explainable hate speech detection. In: (eds) Proceedings of the AAAI conference on artificial intelligence , vol 35 . pp. 14867–14875 Merriam-Webster (2022) Nations U (2019) United Nations Strategy and Plan of Action on Hate Speech. Report No. United Nations Ocampo NB, Cabrio E, Villata S (2023) Unmasking the hidden meaning: Bridging implicit and explicit hate speech embedding representations. In: (eds) Findings of the Association for Computational Linguistics: EMNLP 2023 , vol pp. 6626–6637 Pryzant R, Martinez RD, Dass N, Kurohashi S, Jurafsky D, Yang D (2020) Automatically neutralizing subjective bias in text. In: (eds) Proceedings of the aaai conference on artificial intelligence , vol 34 . pp. 480–489 Roy A, Khanna D, Mahapatra D, Das A (2024) Do the Right Thing, Just Debias! Multi-Category Bias Mitigation Using LLMs. arXiv preprint arXiv:2409.16371 Roychowdhury S, Gupta V (2023) Data-efficient methods for improving hate speech detection. In: (eds) Findings of the Association for Computational Linguistics: EACL 2023 , vol pp. 125–132 Saakyan A, Muresan S (2023) Iclef: In-context learning with expert feedback for explainable style transfer. arXiv preprint arXiv:2309.08583 Sap M, Card D, Gabriel S, Choi Y, Smith NA (2019) The risk of racial bias in hate speech detection. In: (eds) Proceedings of the 57th annual meeting of the association for computational linguistics , vol pp. 1668–1678 Schmeisser-Nieto W, Nofre M (2022) Criteria for the annotation of implicit stereotypes. In: (eds) Proceedings of the Thirteenth Language Resources and Evaluation Conference , vol pp. 753–762 Wen J, Ke P, Sun H, Zhang Z, Li C, Bai J et al (2023) Unveiling the implicit toxicity in large language models. arXiv preprint arXiv:2311.17391 Wiegand M, Kampfmeier J, Eder E, Ruppenhofer J (2023) Euphemistic abuse–a new dataset and classification experiments for implicitly abusive language. In: (eds) Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , vol pp. 16280–16297 Yadav A, Singh V (2024) HateFusion: Harnessing Attention-Based Techniques for Enhanced Filtering and Detection of Implicit Hate Speech. IEEE Trans Comput Social Syst Yuan S, Nie E, Kouba L, Kangen AY, Schmid H, Schütze H et al (2025) LLM in the Loop: Creating the PARADEHATE Dataset for Hate Speech Detoxification. arXiv preprint arXiv:2506.01484 Zhang M, He J, Ji T, Lu C-T (2024) Don't Go To Extremes: Revealing the Excessive Sensitivity and Calibration Limitations of LLMs in Implicit Hate Speech Detection. arXiv preprint arXiv:2402.11406 Zhao J, Wang T, Yatskar M, Ordonez V, Chang K-W (2018) Gender bias in coreference resolution: Evaluation and debiasing methods. arXiv preprint arXiv:1804.06876 Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7368894","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":573873372,"identity":"bfe3a6c8-f8a6-4c6d-8163-8f1be086704e","order_by":0,"name":"Yinghui Huang","email":"","orcid":"","institution":"Wuhan University of Technology","correspondingAuthor":false,"prefix":"","firstName":"Yinghui","middleName":"","lastName":"Huang","suffix":""},{"id":573873373,"identity":"87bab72d-38ad-4f50-8b5c-cc21537631e7","order_by":1,"name":"Qixia Feng","email":"","orcid":"","institution":"Nanjing University","correspondingAuthor":false,"prefix":"","firstName":"Qixia","middleName":"","lastName":"Feng","suffix":""},{"id":573873374,"identity":"b78b4778-cb5e-4d6e-b26a-8a13e07913ec","order_by":2,"name":"Hui Liu","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA5ElEQVRIiWNgGAWjYBACA3YwZcPAwMzcABeVwKuFGUylAbUwkqblMBATq8Wcmfnh58K289H87YzNHxhq7PIMDjAfvM3DYJeHS4tlM5ux9My227kzDjO2STAcSy42OMCWbM3DkFyM02GHediYeYFaGoBagG47kLjhAI+ZNA/DgcQG/FrO5c4/DHQYRAv/N2K0HMjdcJixQQJqCxsBLUC/8JxLzt0I8kvCseTEmUARyzkGybi1HG9++JmnzC533vnDhz98qLFL7AOK3HhTYYdTCypIABHgmDIgSv0oGAWjYBSMAhwAAMBnUvQGa/3WAAAAAElFTkSuQmCC","orcid":"","institution":"Hubei Normal University","correspondingAuthor":true,"prefix":"","firstName":"Hui","middleName":"","lastName":"Liu","suffix":""},{"id":573873375,"identity":"1765e5a4-b139-48b0-a1c1-9e115879bd1e","order_by":3,"name":"Weiqing Li","email":"","orcid":"","institution":"Hubei University of Technology","correspondingAuthor":false,"prefix":"","firstName":"Weiqing","middleName":"","lastName":"Li","suffix":""},{"id":573873376,"identity":"4d26ece6-7c83-48d4-9f19-5c7ea38612dd","order_by":4,"name":"Ying Ma","email":"","orcid":"","institution":"Center for Data Science and Intelligent Decision-Making, Wuhan University of Technology","correspondingAuthor":false,"prefix":"","firstName":"Ying","middleName":"","lastName":"Ma","suffix":""},{"id":573873377,"identity":"b28c2a22-78cd-4e18-900c-c308fd509466","order_by":5,"name":"Zongkui Zhou","email":"","orcid":"","institution":"Key Laboratory of Adolescent Cyberpsychology and Behavior (Ministry of Education)","correspondingAuthor":false,"prefix":"","firstName":"Zongkui","middleName":"","lastName":"Zhou","suffix":""}],"badges":[],"createdAt":"2025-08-14 02:08:17","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7368894/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7368894/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":100372107,"identity":"d3a3855e-49ab-4d32-bb61-b1f915cd0d51","added_by":"auto","created_at":"2026-01-16 08:11:40","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1140572,"visible":true,"origin":"","legend":"","description":"","filename":"CanlargelanguagemodelseffectivelyreshapeonlineimplicithatespeechAnintegrativemodellingapproach5.docx","url":"https://assets-eu.researchsquare.com/files/rs-7368894/v1/da94f6a52b26cfaed2a3c7bf.docx"},{"id":100371139,"identity":"e8f24cf9-9d67-4638-ac50-9199425e01ca","added_by":"auto","created_at":"2026-01-16 08:09:31","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":7624,"visible":true,"origin":"","legend":"","description":"","filename":"6139d6c3b0be4e5aa3b8a2ddfb132ea3.json","url":"https://assets-eu.researchsquare.com/files/rs-7368894/v1/5c0d5bc0d7374d7d46d09388.json"},{"id":100252180,"identity":"e5bcbb6f-f627-4cdb-a6cf-894f7e6891b4","added_by":"auto","created_at":"2026-01-14 15:18:48","extension":"xml","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":163223,"visible":true,"origin":"","legend":"","description":"","filename":"6139d6c3b0be4e5aa3b8a2ddfb132ea31enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-7368894/v1/e9c10b9d0ccc63549bbd7351.xml"},{"id":100371901,"identity":"33158202-c280-45cb-8f50-b13b56cbf3fd","added_by":"auto","created_at":"2026-01-16 08:11:11","extension":"png","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":101357,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7368894/v1/5dc106e3501836810df4ad16.png"},{"id":100372106,"identity":"6a01fbdf-432b-4cdf-ba1a-ade7420d1681","added_by":"auto","created_at":"2026-01-16 08:11:39","extension":"png","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":71308,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7368894/v1/ff3fcd84cc74870604fc10d5.png"},{"id":100371116,"identity":"8c28e272-1ad6-4398-a20b-24d7de6ecdf7","added_by":"auto","created_at":"2026-01-16 08:09:27","extension":"png","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":477834,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7368894/v1/4e1932db6014d55ce53c036c.png"},{"id":100252175,"identity":"e644c5e9-3d98-42c8-b599-55c69dd1bfb6","added_by":"auto","created_at":"2026-01-14 15:18:48","extension":"png","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":128663,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7368894/v1/5c7d132cd956d4ea5024eac3.png"},{"id":100252169,"identity":"a9f8be8b-c720-469d-9a63-4dd4c2741910","added_by":"auto","created_at":"2026-01-14 15:18:48","extension":"jpeg","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":12258,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage5.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7368894/v1/6b65e840c6d463f321e824e5.jpeg"},{"id":100252187,"identity":"72cc135d-57d0-4b89-af1d-b06c0cb21dbf","added_by":"auto","created_at":"2026-01-14 15:18:48","extension":"png","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":235874,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-7368894/v1/4cc409452ff4df826d475dbb.png"},{"id":100371021,"identity":"da0d8e6c-4698-4e00-9352-9fa526a197b3","added_by":"auto","created_at":"2026-01-16 08:09:12","extension":"jpeg","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":12258,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage5.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7368894/v1/fadb41de72730d7d91676b3d.jpeg"},{"id":100371294,"identity":"0f55cd66-087a-4aff-941d-b9e7530657ac","added_by":"auto","created_at":"2026-01-16 08:09:46","extension":"png","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":52483,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7368894/v1/ef7230e91fe67e8a006aede8.png"},{"id":100252183,"identity":"6b3ce6e6-9b49-4fa0-b73f-c174b7d64bb9","added_by":"auto","created_at":"2026-01-14 15:18:48","extension":"png","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":28055,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7368894/v1/f9e6a557b8c437d126a746fb.png"},{"id":100252184,"identity":"02758bc2-7ae9-4637-bd1e-47066574a658","added_by":"auto","created_at":"2026-01-14 15:18:48","extension":"png","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":130077,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7368894/v1/b3f8c98a3223041da52d713d.png"},{"id":100371877,"identity":"32eb805a-096c-45be-b90f-8ef087804266","added_by":"auto","created_at":"2026-01-16 08:11:08","extension":"png","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":32316,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7368894/v1/7bf3e1a912a7fc43a495628a.png"},{"id":100371732,"identity":"5f81c252-915f-432c-a555-450004cbc52e","added_by":"auto","created_at":"2026-01-16 08:10:46","extension":"png","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":3016,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-7368894/v1/030c88831af63b5344a9ebb5.png"},{"id":100252182,"identity":"6af4680d-81e5-483e-afc2-8ca9aa84abc9","added_by":"auto","created_at":"2026-01-14 15:18:48","extension":"png","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":61954,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-7368894/v1/8b2728cb9f70ff359cf6e348.png"},{"id":100252178,"identity":"49cef6ad-a85b-4c7f-a6c7-a2728150c59e","added_by":"auto","created_at":"2026-01-14 15:18:48","extension":"png","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":3016,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-7368894/v1/a0d01e2ed6f74ff0cfaee2d2.png"},{"id":100371644,"identity":"d3d4f3b3-2292-4589-9e83-61b7a9dff4a5","added_by":"auto","created_at":"2026-01-16 08:10:38","extension":"xml","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":161225,"visible":true,"origin":"","legend":"","description":"","filename":"6139d6c3b0be4e5aa3b8a2ddfb132ea31structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-7368894/v1/72c0aeecf2622e3226ed719f.xml"},{"id":100252186,"identity":"9b7598f2-a75d-4039-9ac5-308377b43e83","added_by":"auto","created_at":"2026-01-14 15:18:48","extension":"html","order_by":18,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":170433,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-7368894/v1/ed675a98031895cc582087ff.html"},{"id":100252165,"identity":"140b09fa-9c1d-46ee-8cf6-b2b65fb7053f","added_by":"auto","created_at":"2026-01-14 15:18:48","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":114592,"visible":true,"origin":"","legend":"\u003cp\u003eResearch method and process.\u003c/p\u003e","description":"","filename":"1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7368894/v1/b600add8e5bc9c8be8ea2251.jpg"},{"id":100371736,"identity":"5b016413-6667-4ad7-8285-4e07cdc79557","added_by":"auto","created_at":"2026-01-16 08:10:46","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":64643,"visible":true,"origin":"","legend":"\u003cp\u003ePerformance of the predictive model across different combinations of top-N features.\u003c/p\u003e","description":"","filename":"2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7368894/v1/d0a53e19a161be1f987c3410.jpg"},{"id":100252166,"identity":"e478afa1-3aa9-40f6-b6b5-573868da1b1e","added_by":"auto","created_at":"2026-01-14 15:18:48","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":63186,"visible":true,"origin":"","legend":"\u003cp\u003eOverall interpretability SHAP plot for the IHS transformation.\u003c/p\u003e\n\u003cp\u003eNote. The plot is composed of thousands of individual points from the training dataset. Feature values are colour-coded, with higher values shown in red and lower values in blue, as indicated by the colour bar on the right. Thus, if the points on one side of the central line gradually shift towards red or blue, this suggests that an increase or decrease in the feature value respectively pushes the prediction of IHS in that direction. In this study, a predicted value of 0 indicates IHS, and 1 indicates non-IHS.\u003c/p\u003e","description":"","filename":"3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7368894/v1/405f9b69744f28c6e9dd5d2a.jpg"},{"id":100252168,"identity":"49b65447-0b1c-4888-ad3a-77f45ad19e94","added_by":"auto","created_at":"2026-01-14 15:18:48","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":50010,"visible":true,"origin":"","legend":"\u003cp\u003eDistribution of toxicity scores by category and topic type.\u003c/p\u003e","description":"","filename":"4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7368894/v1/021dea789f111f1f246d74c1.jpg"},{"id":100371306,"identity":"450ae875-40a5-4f0b-8a8d-a25f48ffb0ca","added_by":"auto","created_at":"2026-01-16 08:09:47","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":82248,"visible":true,"origin":"","legend":"\u003cp\u003eDistribution of toxicity scores.\u003c/p\u003e","description":"","filename":"5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7368894/v1/c2fa406c9ae7b070268efc59.jpg"},{"id":103693976,"identity":"5315feaa-952b-46eb-a290-6f46d5b9b144","added_by":"auto","created_at":"2026-03-01 15:10:31","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2232730,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7368894/v1/f7e6242e-e823-4f16-86d1-e7ad5764f179.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Can large language models effectively reshape online implicit hate speech? An integrative modelling approach","fulltext":[{"header":"Introduction","content":"\u003cp\u003eSocial media has become a key space for global information exchange. The governance of online hate speech is now an urgent global issue. A report from the Anti-Defamation League (ADL) shows that 25% of adults have faced severe online harassment. Among groups such as LGBTQ\u0026thinsp;+\u0026thinsp;people, this number rises to 49% (League, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). In response, different regions have taken different paths. The United Nations promotes countering \u0026ldquo;hate speech\u0026rdquo; with \u0026ldquo;more speech,\u0026rdquo; focusing on education and counter-narratives (Nations, \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). China has set strict rules to control \u0026ldquo;harmful information,\u0026rdquo; including hate speech, aiming to build a \u0026ldquo;clean\u0026rdquo; online space (Nations, \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). China has set strict rules to control \u0026ldquo;harmful information,\u0026rdquo; including hate speech, aiming to build a \u0026ldquo;clean\u0026rdquo; online space (China, \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). The European Union\u0026rsquo;s Digital Services Act (DSA) requires large platforms to assess and reduce risks, with clear rules and user rights (Council of the European Union, 2022).\u003c/p\u003e \u003cp\u003eIn this context, implicit hate speech (IHS) has become a major challenge in online governance. IHS uses hidden expression, depends on cultural context, and has complex meanings. Unlike direct attacks, it uses tools like metaphor, irony, or vague references (Ocampo, et al., \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e2023\u003c/span\u003e;Wen, et al., \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) to avoid keyword filters while still sending a negative message to specific groups (ElSherief, et al., \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). This can cause repeated psychological harm and strengthen systemic discrimination, making IHS a covert form of cyberbullying.\u003c/p\u003e \u003cp\u003eThe old \u0026ldquo;detect-and-delete\u0026rdquo; method has clear limits. Automated systems can wrongly block fair social criticism (Johnson, et al., \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). Human review cannot keep up with the speed and size of online content. As a result, researchers have started using AI to reshape IHS. The goal is to balance free speech with community safety. Early methods changed word embeddings (Bolukbasi, et al., \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2016\u003c/span\u003e) or hidden states in models (Zhao, et al., \u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). Later, the \u0026ldquo;detect-and-rewrite\u0026rdquo; method appeared (He, et al., 2021;Pryzant, et al., \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). In recent years, large language models (LLMs) have shown strong skills in text style transfer (TST). They can change offensive text into neutral text by replacing aggressive words or changing sentence structure, while keeping the main meaning (Hee, et al., \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2023\u003c/span\u003e;Pryzant, et al., \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2020\u003c/span\u003e;Roy, et al., \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). However, most research looks only at explicit hate speech. The ability and risks of handling IHS are not well studied.\u003c/p\u003e \u003cp\u003eThe need for such study comes from both the special nature of IHS and the complexity of LLMs. First, IHS depends on culture and subtle language cues (Zhang, et al., \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e2024\u003c/span\u003e), which are hard for models to read. Second, LLMs may keep hidden biases from training data, leading to covert toxic content. Third, there is no full framework to measure both the effect and the risks of reshaping IHS (Saakyan and Muresan, \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). Solving these problems needs a mix of linguistics, computing, and ethics.\u003c/p\u003e \u003cp\u003eAccordingly, this study seeks to address the following core research questions. RQ1:Can advanced LLMs reshape IHS with high quality? LLMs work well for explicit hate speech (Kostiuk, et al.), but for IHS, can they reduce hate while keeping meaning?Accordingly, this study seeks to address the following overarching research question: How effectively can advanced LLMs reshape implicit hate speech while preserving meaning, through which interpretable linguistic mechanisms, and with what constructive and residual risks? To operationalize this overarching question, we adopt a three-step research approach. First, we assess reshaping effectiveness by building a predictive model that quantifies changes in toxicity and semantic preservation, thereby testing whether LLMs can reduce implicit hate while maintaining core meaning. Second, we construct an interpretable linguistic-behavior framework that integrates theoretically grounded linguistic features with explainable AI to identify key language cues, enabling fine-grained measurement of reshaping strategies beyond single-score evaluations. Third, we apply this framework to pattern analysis, comparing texts before and after reshaping to uncover both constructive strategies (e.g., de-targeting, tone softening) and residual risks (e.g., obfuscation, risk displacement) in LLM-mediated transformations.\u003c/p\u003e \u003cp\u003eTo achieve these objectives, We use an integrative modelling method from computational social science (Hofman, et al., \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2021\u003c/span\u003e).Using a large IHS dataset (ElSherief, et al., \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2021\u003c/span\u003e), we run three stages: predictive modelling, interpretive modelling, and experimental testing. The results give both theory and practice for safe and effective AI governance.\u003c/p\u003e"},{"header":"Literature Review","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eIHS: Definition, Linguistic Features, and Offensive Strategies\u003c/h2\u003e \u003cp\u003eIHS means language that avoids open insults but still sends discrimination or hostility through soft, indirect, or coded words, such as euphemism, metaphor, or sarcasm (China, \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2019\u003c/span\u003e;Ghosh, et al., \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2023\u003c/span\u003e;Lin, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2022\u003c/span\u003e;Nations, \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e2019\u003c/span\u003e) Compared with explicit hate speech, IHS hides its attack but aims to harm the same protected groups, such as by race or ethnicity (Merriam-Webster, \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). This harm is often linked to shared culture or history.\u003c/p\u003e \u003cp\u003eOne key strategy is hiding the meaning and making it more extreme. Offenders use \u0026ldquo;dog whistles\u0026rdquo; or coded words (Risius, 2024) to hide extreme ideas in harmless-looking terms, such as using a metaphor to compare people to \u0026ldquo;pests\u0026rdquo; (ElSherief, et al., \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). Spotting these words is hard (Magu and Luo, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). A second is hiding emotion. Offenders may use irony, sarcasm, or \u0026ldquo;jokes\u0026rdquo; (China, \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2019\u003c/span\u003e;Hee, et al., \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) to cover hostility, giving them a way to deny the intent (ElSherief, et al., \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). A third is bias by suggestion. Instead of direct attack, they cause group conflict using negative stereotypes (Council of the European Union, 2022;Sap, et al., \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e2019\u003c/span\u003e) or by telling stories about unfair treatment or being victims. These work best when the audience shares certain culture or knowledge(Ghosh, et al., \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2023\u003c/span\u003e;Schmeisser-Nieto, et al., 2022).\u003c/p\u003e \u003cp\u003eBecause these strategies are indirect, IHS can pass content checks and is hard for machines to detect. Current research is still small (Nations, \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). Methods have gone from static word embeddings (Magu and Luo, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2018\u003c/span\u003e) to context-aware models (Ghosh, et al., \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2023\u003c/span\u003e;Kim, et al., \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). But these models are often \u0026ldquo;black boxes\u0026rdquo; and can be tricked by adversarial inputs (Mathew, et al., \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). This is why more work is now on making clear, explainable systems.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eGoverning IHS: From Moderation to Reshaping\u003c/h3\u003e\n\u003cp\u003eThe governance of IHS is undergoing a shift. Traditional keyword- or rule-based moderation has shown limited results because it cannot handle the core strategies of IHS. It struggles to detect \u0026ldquo;dog whistle\u0026rdquo; terms (Marten Risius, \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) hidden through linguistic masking and extremisation, and it fails to understand biased suggestions in social interactions that depend heavily on shared cultural and social backgrounds (Ghosh, et al., \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2023\u003c/span\u003e;Schmeisser-Nieto, et al., 2022). These structural gaps create systematic blind spots in content review(Hartmann, et al., \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2025\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eAs a result, governance is moving towards constructive interventions centred on \u0026ldquo;speech reshaping.\u0026rdquo; The goal is no longer simple deletion, but the use of TST techniques (Hu, et al., \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2022\u003c/span\u003e;Johnson, et al., \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2019\u003c/span\u003e) to directly address the linguistic strategies that harm specific groups. This can include reshaping to remove hostility from implicit emotional expressions\u0026mdash;such as softening sarcasm\u0026mdash;or reconstructing sentences to break biased suggestions in social interactions, for example by reducing stereotypes, while keeping the non-hateful core meaning (Pryzant, et al., \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2020\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eGiven the complexity of geopolitical and cultural contexts, banning IHS through a single set of rules is not only impractical but may even backfire (Johnson, et al., \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). Proactive reshaping through TST has therefore become a viable new approach (Hu, et al., \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2022\u003c/span\u003e;Jin, et al., 2022). This strategy aims to turn biased content into neutral expressions while preserving the original intent. Within this approach, in addition to strategies such as \u0026ldquo;counter speech\u0026rdquo; (Cepollaro, et al., \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) that seek to change the overall discourse environment, the use of LLMs for \u0026ldquo;detoxification\u0026rdquo; (He, et al., 2021;Pryzant, et al., \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2020\u003c/span\u003e) has become a key technology. However, current research almost entirely focuses on explicit hate speech (Kostiuk, et al.;Roy, et al., \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2024\u003c/span\u003e), while work on reshaping IHS is still at an early stage. Common issues include loss of semantic coherence and insufficient modelling of context (Johnson, et al., \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). This makes the systematic evaluation and improvement of LLM performance in IHS reshaping an urgent research gap to address.\u003c/p\u003e\n\u003ch3\u003eLLMs and Their Evaluation Methods\u003c/h3\u003e\n\u003cp\u003eThe new generation of LLMs, represented by the GPT series (Achiam, et al., \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) and DeepSeek (Guo, et al., \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2025\u003c/span\u003e), is characterised by emergent abilities. These abilities arise from the massive expansion of model parameters and training data, enabling advanced functions\u0026mdash;such as complex reasoning and deep contextual understanding\u0026mdash;that were absent in smaller-scale models (Achiam, et al., \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). Such capabilities provide the technical foundation for addressing culturally sensitive tasks like IHS.Evaluating these complex models requires multi-dimensional approaches. These include human evaluation, which is the gold standard but costly, and computational evaluation, which is more scalable but relies on automated metrics. Traditional automated metrics, however, often fail to capture the deeper quality of text, especially whether an LLM truly understands and neutralises the complex strategies of IHS. For example, a model might produce fluent and harmless-looking text, but the metrics cannot tell whether it has identified the real intent hidden behind linguistic masking, or whether it has simply imitated a harmless surface style.\u003c/p\u003e \u003cp\u003eTo address these gaps, integrated evaluation methods have emerged\u0026mdash;such as using one LLM to assess the output of another. In the context of IHS, an ideal evaluation system must go beyond \u0026ldquo;black-box\u0026rdquo; classification (Mathew, et al., \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2021\u003c/span\u003e) and be able to interpret the socio-cultural meanings encoded in language (Ghosh, et al., \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2023\u003c/span\u003e;Lin, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). This means that LLMs are not only the subject of evaluation but can also serve as tools for building interpretable evaluation frameworks.A human-like computational evaluation system should have the capacity to recognise the strategies underlying IHS. It should not only determine whether a text ultimately harms a specific group but also explain how that harm is conveyed\u0026mdash;whether through implicit emotional expression or biased suggestions. In this way, evaluation moves beyond \u0026ldquo;black-box\u0026rdquo; judgements (Mathew, et al., \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2021\u003c/span\u003e) and achieves genuine interpretability.\u003c/p\u003e\n\u003ch3\u003eFormulation of Research Questions\u003c/h3\u003e\n\u003cp\u003eWith the rise of online interaction, IHS has become a major barrier to constructive dialogue in digital spaces. Existing studies show that IHS spreads hostility through complex linguistic strategies, such as coded language, sarcasm, negative stereotypes, and dehumanising metaphors (Lemmens, et al., \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2021\u003c/span\u003e;Wiegand, et al., \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2023\u003c/span\u003e), making it more difficult to detect and manage.In this context, using generative artificial intelligence to \u0026ldquo;reshape\u0026rdquo; or \u0026ldquo;detoxify\u0026rdquo; such language (Pryzant, et al., \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2020\u003c/span\u003e) has emerged as a cutting-edge approach to promoting a more civil online environment. However, most detoxification research has focused on explicit hate speech (Kostiuk, et al.;Roy, et al., \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Assessing the capabilities of LLM in the more complex task of reshaping IHS remains a clear research gap(Johnson, et al., \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2019\u003c/span\u003e).Because the offensiveness of IHS is embedded in its linguistic structure, an effective evaluation system must examine whether an LLM can recognise and neutralise these underlying harmful strategies. To address this, the present study proposes a dual evaluation framework that combines external indicators (reshaping outcomes) with internal indicators (treatment of linguistic strategies). The internal framework focuses on four core operational dimensions:group-specific harm, implicit emotional expression, linguistic obfuscation and extremity, and bias and implication in social interaction. This design enables a systematic assessment of the capabilities, strategies, and behaviours of LLM in reshaping IHS.\u003c/p\u003e \u003cp\u003eAccordingly, this study first examines whether LLM can effectively achieve a high success rate in reshaping IHS. To address this, an AI-based predictive model will be developed, following the text detoxification paradigm(Yuan, et al., \u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e2025\u003c/span\u003e), to conduct external evaluation across two dimensions: hate mitigation and semantic preservation.Second, the study investigates how to construct a comprehensive evaluation framework that integrates the core linguistic strategies of IHS for internal assessment of reshaping performance. IHS will be deconstructed into four strategic categories\u0026mdash;group-specific harm, implicit emotional expression, linguistic obfuscation and extremity, and bias and implication in social interaction. Using explainable AI methods, the study will identify key linguistic indicators closely associated with these strategies, thereby building a computational internal evaluation framework.Finally, this framework will be applied to examine whether, in the process of reshaping IHS, LLM can effectively neutralise the strategies found in human-generated IHS. Comparative analysis of texts before and after reshaping will be conducted to explore the extent to which these strategies and behaviours are mitigated through language transformation.\u003c/p\u003e\n\u003ch3\u003eMethods and Experiments\u003c/h3\u003e\n\u003cp\u003eThis study adopts the integrative modelling approach from computational social science (Hofman, et al., \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2021\u003c/span\u003e) to address the above research questions. This approach combines predictive and interpretive modelling, aiming to achieve both accurate forecasting of phenomena and in-depth explanation of their underlying mechanisms. In this research, we use machine learning and deep learning algorithms to build predictive models, integrate interpretability techniques to explore underlying mechanisms, and apply a systematic experimental design to validate the findings.\u003c/p\u003e \u003cp\u003eSpecifically, based on the high-quality Latent Hatred benchmark dataset (Section 3.1), we follow a three-stage research process to systematically assess the overall performance of LLMs in reshaping IHS. The process is illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. Predictive modelling stage : Build and train a high-performance deep learning model to enable automated and large-scale evaluation of LLM reshaping outcomes.Interpretive modelling stage : Apply interpretability methods to identify and quantify key linguistic cues influencing IHS classification, thereby constructing a computational framework for evaluating LLM linguistic behaviours.Experimental validation stage: Conduct dual verification and in-depth analysis of LLM reshaping performance using multiple indicators, including toxicity reduction, semantic retention, and the linguistic cues identified above.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eDataset\u003c/h2\u003e \u003cp\u003eThis study employed the Latent Hatred benchmark dataset developed by ElSherief et al. (ElSherief, et al., \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). The dataset was collected from the Twitter and contains a total of 27,157 tweets, each annotated for hate speech category (implicit hate, explicit hate, and non-hate). Among these, 12,671 are IHS texts (including 7,100 unique tweets). In addition, the dataset includes annotations for the target group and specific type of hate speech, covering six categories: incitement, inferiority, irony, stereotypical, threatening, and white grievance.\u003c/p\u003e \u003cp\u003eFor the reshaping experiments, all IHS samples were selected, and two representative LLMs were used to perform text reshaping: GPT-4o, released by OpenAI in May 2024, and Deepseek-V2, released by DeepSeek in May 2024. All reshaped texts corresponding to the IHS tweets were obtained via the official APIs. Detailed statistical information on the dataset is provided in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eStatistical Information of the Dataset\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStatistics\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTweet Length\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eLength of IHS Tweets\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eDeepseek Reshaped Text Length\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eGPT-4o Reshaped Text Length\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCount\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e27,157\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e12,035\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e7,100\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e7,100\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMean\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e90.545\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e31.299\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e160.878\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e126.684\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStandard deviation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e41.493\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e11.446\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e69.562\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e42.005\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMinimum\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e14\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e25th percentile\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e63\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e24\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e114\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e97\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMedian\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e29\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e147\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e122\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e75th percentile\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e113\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e36\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e192\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e151\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMaximum\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e801\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e186\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e572\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e418\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003ePredictive Modelling: Automated Evaluation of LLM Reshaping Capability\u003c/h3\u003e\n\u003cp\u003eTo quantitatively assess the effectiveness of LLMs in reshaping IHS, this stage aims to develop an automated classification model. The core task of this model is to determine the final speech category of the text reshaped by LLMs (e.g., \u0026ldquo;non-hate\u0026rdquo; or \u0026ldquo;still implicit hate\u0026rdquo;).\u003c/p\u003e \u003cp\u003eIn model construction, a comparative experimental design was adopted. First, at the feature engineering stage, input features were prepared for two types of algorithms: (1) psycho-linguistic features extracted using the LIWC tool for training traditional machine learning models; and (2) word embedding vectors generated by the RoBERTa model as input for deep learning models. Second, in algorithm selection, a wide range of models was examined, including traditional machine learning algorithms such as Logistic Regression (LR), Support Vector Machine (SVM), Random Forest (RF), and XGBoost, as well as multiple deep learning architectures such as RoBERTa-BiLSTM-Attention.\u003c/p\u003e \u003cp\u003eThe performance of all candidate models was comprehensively evaluated using multiple metrics, including accuracy, precision, recall, F1-score, and ROC AUC, with cross-validation employed to ensure generalisation capability. In addition to classification performance, another key indicator\u0026mdash;semantic retention\u0026mdash;was also assessed. This study applied the BERTScore method based on RoBERTa-large to compute the F1 semantic similarity between the original and reshaped texts, thereby quantifying the extent to which LLMs preserved the core semantics during the reshaping process.\u003c/p\u003e\n\u003ch3\u003eInterpretive Modelling: Construction of the IHS Linguistic Behaviour Evaluation Framework\u003c/h3\u003e\n\u003cp\u003eThe objective of this stage is to move beyond a surface-level assessment of \u0026ldquo;whether it works\u0026rdquo; to explore \u0026ldquo;why it works\u0026rdquo; and \u0026ldquo;where potential risks lie.\u0026rdquo; This entails constructing an interpretable framework for evaluating linguistic behaviours in IHS detection. The framework aims to identify key linguistic cues that play a decisive role in determining IHS and, based on these cues, to evaluate the specific linguistic operations employed by GPT-4o and Deepseek during the reshaping process.\u003c/p\u003e \u003cp\u003eThe construction process was as follows. First, with IHS category (yes/no) as the prediction target and over one hundred linguistic features extracted via LIWC as input variables, a series of machine learning classifiers were developed. Second, to select the most influential linguistic cues from high-dimensional features, a systematic feature selection procedure was designed. We initially applied the intrinsic feature importance ranking of the XGBoost model to conduct a coarse-grained filtering, removing low-contribution redundant features. Subsequently, on the resulting candidate subset, we employed a stepwise forward selection strategy, starting from an empty set and iteratively adding the feature that yielded the greatest improvement in model performance, as measured by F1-score and ROC AUC, until no significant performance gain was observed. This process ensured that the final feature subset was both parsimonious and effective.\u003c/p\u003e \u003cp\u003eFinally, for the optimal classification model built on this refined feature subset\u0026mdash;determined through grid search and cross-validation\u0026mdash;the SHAP (Shapley Additive Explanations) method was applied for in-depth interpretation. SHAP quantifies both the magnitude and the direction (positive/negative) of each linguistic cue\u0026rsquo;s contribution to every individual prediction, enabling precise identification of the core linguistic cues most influential in IHS determination. These identified cues, together with their association patterns with IHS, collectively constitute the proposed IHS linguistic behaviour evaluation framework.\u003c/p\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eFramework Validation: Dual Evaluation of LLM Linguistic Behaviour\u003c/h2\u003e \u003cp\u003eTo validate the effectiveness and practical utility of the linguistic behaviour evaluation framework developed in the previous stage, we conducted a dual evaluation. The core objective was to examine whether the linguistic changes captured by the framework aligned with externally perceived changes in textual risk.\u003c/p\u003e \u003cp\u003eSpecifically, for both the original and reshaped texts generated by GPT-4o and Deepseek, we conducted two complementary assessments. First, an internal linguistic feature analysis was carried out using the proposed evaluation framework to measure the average percentage change in key linguistic cues (e.g., sentiment polarity, frequency of specific lexical items) before and after reshaping, thereby revealing what the LLMs did at the linguistic level. Second, an external risk metric assessment was performed using established third-party tools (e.g., Perspective API) to calculate the average percentage change in toxicity scores, reflecting the actual impact of reshaping on external perceptions of risk.\u003c/p\u003e \u003cp\u003eBy comparing the results from internal and external analyses, we were able to verify whether the linguistic behaviours identified by the framework were indeed the primary drivers of reduced hateful attributes in the text. This not only confirmed the validity of the framework but also provided robust evidence for a deeper understanding of the mechanisms underlying LLM reshaping.\u003c/p\u003e \u003c/div\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003ePredictive Modelling Analysis: Evaluating LLM Reshaping Capability\u003c/h2\u003e \u003cdiv id=\"Sec14\" class=\"Section3\"\u003e \u003ch2\u003ePerformance Comparison of Automated Evaluation Models\u003c/h2\u003e \u003cp\u003eThe comparative analysis produced several key findings(Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e) .First, among the deep learning models using RoBERTa word embedding features, the RoBERTa-BiLSTM-Attention model achieved the best overall performance, with both accuracy and ROC AUC reaching 0.864. Notably, although the XLNet model had a slightly lower accuracy (0.853), it achieved the highest recall (0.901), indicating stronger ability in identifying positive cases.Secondly, for traditional machine learning models based on LIWC psycho-linguistic features, ensemble learning algorithms performed best. In particular, XGBoost outperformed other models in ROC AUC (0.838), recall (0.713), and F1 score (0.741), showing the strongest discriminative ability, while Random Forest achieved the highest accuracy (0.756).Finally, when combining LIWC and RoBERTa features, the effects varied by model type. For traditional machine learning models, this combination significantly improved performance limits. For example, the XGBoost model\u0026rsquo;s ROC AUC increased to 0.895, while its accuracy and F1 score improved by 7.7% and 7.3%, respectively, compared with using LIWC features alone. However, this positive effect did not appear in deep learning models, suggesting that for models already capable of strong semantic representation, adding LIWC features brought no further gains.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003ePerformance of Different Classification Algorithms\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFeatures\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAlgorithm\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePrecision\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eRecall\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eF1\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eROC AUC\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003e\u003cb\u003eLIWC features\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLogistic Regression\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.738\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.747\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.711\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.729\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.807\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSupport Vector Machine\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.751\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.824\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.632\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.715\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.819\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRandom Forest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.756\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.821\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.649\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.725\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.834\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eXGBoost\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.753\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.770\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.713\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.741\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.838\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"5\" rowspan=\"6\"\u003e \u003cp\u003e\u003cb\u003eRoBERTa word embedding features\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRoBERTa\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.858\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.840\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.885\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.862\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.858\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRoBERTa-LSTM-Attention\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.862\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.845\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.886\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.865\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.862\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eXLNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.853\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.822\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.901\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.850\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.853\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRoBERTa-BiLSTM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.860\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.843\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.885\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.864\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.860\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRoBERTa-CNN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.861\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.846\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.882\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.864\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.861\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRoBERTa-BiLSTM-Attention\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.864\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.850\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.883\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.866\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.864\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"7\" rowspan=\"8\"\u003e \u003cp\u003e\u003cb\u003eCombined LIWC and RoBERTa word embedding features\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLogistic Regression\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.761\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.787\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.708\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.746\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.833\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSupport Vector Machine\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.752\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.829\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.630\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.716\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.820\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRandom Forest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.785\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.905\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.632\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.744\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.865\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eXGBoost\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.811\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.859\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.740\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.795\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.895\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRoBERTa\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.795\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.798\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.795\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.795\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.795\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRoBERTa-LSTM-Attention\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.805\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.813\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.804\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.804\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.804\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRoBERTa-BiLSTM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.798\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.815\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.796\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.794\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.796\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRoBERTa-BiLSTM-Attention\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.802\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.805\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.802\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.801\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.802\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003eEvaluation of LLM Reshaping Performance Based on the Optimal Model\u003c/h2\u003e \u003cp\u003eBased on the above comparison, this study selected the RoBERTa-BiLSTM-Attention model\u0026mdash;identified as having the best overall performance\u0026mdash;as the final automated evaluation tool to determine whether a text reshaped by an LLM still belongs to IHS. To conduct a rigorous statistical test, we applied the McNemar test to examine whether the distribution of IHS categories for the same set of tweets changed significantly before and after LLM reshaping. The results are shown in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003ePaired-Sample Chi-Square Test of the Number of Aggressive Items Between OL and OL AUTO-Cot, and Semantic Preservation Results\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"8\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLLM\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTopic Type\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePrompt Type\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eNo-Hate Count\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eHate Count\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eSuccess Rate of No-Hate Conversion (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003ep-value\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003eSemantic Preservation\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"13\" rowspan=\"14\"\u003e \u003cp\u003eDeepseek\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eIncitement\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eOL\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1240\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e0.000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e0.832\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eOL AUTO-Cot\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e1235\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e99.60\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eInferiority\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eOL\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e859\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e0.000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e0.834\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eOL AUTO-Cot\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e849\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e10\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e98.84\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eIrony\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eOL\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e794\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e0.000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e0.853\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eOL AUTO-Cot\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e750\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e44\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e94.41\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eOther\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eOL\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e79\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e0.000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e0.842\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eOL AUTO-Cot\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e76\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e96.20\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eStereotypical\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eOL\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1104\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e0.000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e0.829\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eOL AUTO-Cot\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e1095\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e99.19\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eThreatening\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eOL\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e666\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e0.000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e0.822\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eOL AUTO-Cot\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e661\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e99.25\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eWhite grievance\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eOL\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1506\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e0.000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e0.829\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eOL AUTO-Cot\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e1502\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e99.73\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"13\" rowspan=\"14\"\u003e \u003cp\u003eGPT-4o\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eIncitement\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eOL\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1240\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e0.000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e0.829\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eOL AUTO-Cot\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e1238\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e99.84\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eInferiority\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eOL\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e859\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e0.000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e0.831\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eOL AUTO-Cot\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e853\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e99.30\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eIrony\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eOL\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e794\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e0.000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e0.842\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eOL AUTO-Cot\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e775\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e19\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e97.60\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eOther\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eOL\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e79\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e0.000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e0.832\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eOL AUTO-Cot\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e79\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e100\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eStereotypical\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eOL\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1104\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e0.000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e0.823\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eOL AUTO-Cot\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e1098\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e99.46\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eThreatening\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eOL\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e666\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e0.000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e0.823\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eOL AUTO-Cot\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e664\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e99.70\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eWhite grievance\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eOL\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1506\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e0.000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e0.823\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eOL AUTO-Cot\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e1503\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e99.80\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eThe statistical results clearly indicate that, whether reshaped by Deepseek or GPT-4o, the distribution of all IHS topic types (e.g., incitement, inferiority, irony) showed highly significant changes before and after reshaping (all p\u0026thinsp;\u0026lt;\u0026thinsp;0.001). This finding confirms that LLMs can effectively and significantly reduce implicit hate attributes in text.Alongside evaluating reshaping effectiveness, we also examined semantic preservation. The results show that both LLMs achieved a semantic preservation rate of 82%\u0026ndash;85% (measured by the BERTScore F1 value) when performing reshaping tasks. This means that LLMs can remove implicit bias while maintaining the core meaning of the original text. Taken together, these findings provide clear empirical support for our first research question (RQ1).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003eInterpretability Modelling Analysis: Development of the IHS Linguistic Behaviour Evaluation Framework\u003c/h2\u003e \u003cp\u003eTo gain deeper insight into the linguistic behaviours of LLMs when reshaping IHS, this study developed an interpretable classification model designed to identify the key linguistic cues that determine IHS. We first evaluated the performance of several candidate classification models (see Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). The results showed that the XGBClassifier achieved the highest accuracy (0.714) and precision (0.722), making it the optimal model. Accordingly, this model was selected as the basis for the subsequent in-depth interpretability analysis.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eComparison of Different Classification Models in Assessing IHS Reshaping Performance\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAlgorithm\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePrecision\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRecall\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eF1\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eROC AUC\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLogistic Regression\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.699\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.703\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.940\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.803\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.594\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRidge Regression\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.699\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.701\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.944\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.804\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.592\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSupport Vector Classifier\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.695\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.687\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.987\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.809\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.568\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMulti-layer Perceptron\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.686\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.735\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.816\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.773\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.630\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRandom Forest Classifier\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.710\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.706\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.959\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.812\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.601\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eXGBoost Classifier\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.714\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.722\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.918\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.808\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.626\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"6\"\u003eWe applied the SHAP (Shapley Additive Explanations) method to the optimal XGBClassifier model to conduct a global interpretability analysis, aiming to evaluate how different linguistic features influence IHS classification predictions. Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e presents the top 25 most influential linguistic cues identified by the model, along with their overall contribution. To further examine the direction and magnitude of each cue\u0026rsquo;s impact, we plotted a SHAP summary diagram (see Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e). This diagram shows that these cues influence IHS classification mainly in two opposite ways:\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003ePositively correlated cues (increasing the probability of being classified as IHS):The results indicate that when the text contains more words referring to specific groups\u0026mdash;such as ethnicity, visual, body, religion, the pronoun \u0026ldquo;they,\u0026rdquo; and informal pronouns (ipron)\u0026mdash;as well as more adjectives (adj), exclamation marks (Exclam), emotional words (Affect), absolute terms (allnone), and power-related words (power), the probability of the text being classified as IHS increases. This suggests that these linguistic cues play a key role in driving the expression of implicit hate.\u003c/p\u003e \u003cp\u003eNegatively correlated cues (decreasing the probability of being classified as IHS):In contrast, when the text contains more features related to tone (Tone), prepositions (prep), past tense (focuspast), social behaviour (socbehav), communication (comm), and work-related terms (work), the probability of the text being classified as IHS decreases. This indicates that the presence of these cues often signals a shift towards more neutral or constructive content.\u003c/p\u003e \u003cp\u003eTo systematise the above quantitative results, this study developed an IHS speech behaviour evaluation framework (Table\u0026nbsp;\u003cspan refid=\"Tab5\" class=\"InternalRef\"\u003e5\u003c/span\u003e) based on SHAP analysis and relevant literature. Instead of merely listing linguistic cues, the framework organises them into four broader theoretical dimensions, as shown in Table\u0026nbsp;\u003cspan refid=\"Tab6\" class=\"InternalRef\"\u003e6\u003c/span\u003e. The first dimension, group-specific harm, relates to references and descriptions of specific groups, such as \u0026ldquo;ethnicity\u0026rdquo; and \u0026ldquo;visual features\u0026rdquo;. The second, implicit emotional expression, covers cues reflecting the intensity and polarity of emotion, with \u0026ldquo;adjectives\u0026rdquo; and \u0026ldquo;exclamation marks\u0026rdquo; as positive indicators and \u0026ldquo;tone\u0026rdquo; as a negative indicator. The third, linguistic obfuscation and extremity, involves strategies that make statements more vague or absolute, such as \u0026ldquo;commas\u0026rdquo; and \u0026ldquo;all-or-none terms\u0026rdquo;. The fourth, bias and implication in social interaction, captures potential bias in describing social interactions, such as \u0026ldquo;motivation words\u0026rdquo; and \u0026ldquo;power words\u0026rdquo;. This framework addresses RQ2 by not only identifying key linguistic behaviours influencing IHS but also providing a quantifiable, multi-dimensional system for analysing and assessing the specific language operations applied by LLMs during the reshaping process.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab5\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 5\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eAnalytical framework for IHS linguistic behaviours in online\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eConcept\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDefinition\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eApplication example\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eApplicable factors\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGroup-specific harm\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSentences imply discriminatory expressions toward specific groups by referring to protected attributes such as race, religion, appearance, and physical traits, often through vague references.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eUnderstand the groups, references, and related historical or social background in the text; avoid negative or discriminatory statements; use positive or neutral expressions for protected traits like race, religion, or appearance.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eProtected traits, historical and social background\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eImplicit emotional expression\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eImplicitly convey aggressive emotions through the use of emotional words, exclamation marks, and other punctuation.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eUse emotional words and punctuation to judge emotional tone, identify inflammatory remarks, and replace them with positive or neutral words to communicate inclusively.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eEmotional words, punctuation\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLinguistic obfuscation and extremity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eConvey extreme positions or hide true intentions implicitly through complex sentence structures, euphemisms, punctuation use, and absolute terms.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eIdentify unreasonable elements in text semantics or structure, as well as exaggerated, euphemistic, vague, or metaphorical expressions; ensure language is clear, direct, and inclusive.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSentence structure, punctuation use, absolute terms\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBias and implication in social interaction\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eImplicitly convey bias or negative attitudes toward specific groups through words related to social behaviours and communication.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eUse positive or neutral language in communication, social behaviour, and work discussions, considering social context and background knowledge, to avoid conflict, inequality, and negative stereotypes.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSocial behaviours, communication words\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab6\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 6\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eComprehensive evaluation framework for IHS\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eConcept\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRank\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCategory in LIWC\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eInternal indicator\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eImportance\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eRelationship between indicator and IHS\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"6\" rowspan=\"7\"\u003e \u003cp\u003eGroup-specific harm\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCulture\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eEthnicity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.307\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePerception\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eVisual\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.14\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePhysical\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePhysical\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.129\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePronouns\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eThey\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.115\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCulture\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRelig\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.106\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e17\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eLinguistic Dimensions\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eIpron\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.058\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e24\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePerception\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePerception\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.045\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"4\" rowspan=\"5\"\u003e \u003cp\u003eImplicit emotional\u003c/p\u003e \u003cp\u003eexpression\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e16\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eLinguistic Dimensions\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAdj\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.064\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e12\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePunctuation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eExclam\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.08\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e13\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePsychological Processes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAffect\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.075\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e19\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePunctuation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAllPunc\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.052\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e23\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSummary Variables\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eTone\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.046\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eNegative\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003eLinguistic obfuscation and extremity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePunctuation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eComma\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.366\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e14\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePsychological Processes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eDiscrep\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.068\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e20\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eLinguistic Dimensions\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePrep\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.052\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eNegative\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e21\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePsychological Processes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAllnone\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.05\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"5\" rowspan=\"6\"\u003e \u003cp\u003eBias and implication in social interaction\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e10\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTime Orientation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eFocuspast\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.085\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eNegative\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e11\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePsychological Processes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eDrives\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.085\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSocial Processes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSocbehav\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.093\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eNegative\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e15\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSocial Processes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eComm\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.065\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eNegative\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e22\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eLifestyle\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eWork\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.048\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eNegative\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e25\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePsychological Processes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePower\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.035\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eUncategorised words\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSummary Variables\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eWc\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.106\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSummary Variables\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eDic\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eNegative\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e18\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCognition\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSpace\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.057\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003eValidation of the comprehensive evaluation framework based on intrinsic indicators\u003c/h2\u003e \u003cp\u003eTo closely examine the language behaviour of LLMs during the reshaping process, this study applied the previously established evaluation framework to compare the mean values of each linguistic cue (intrinsic indicator) between the original IHS tweets and the texts reshaped by LLMs (Deepseek and GPT-4o). This analysis aimed to reveal the specific language strategies used to reduce hateful content, the commonalities and differences between models, and potential limitations. Detailed results are shown in Table\u0026nbsp;\u003cspan refid=\"Tab7\" class=\"InternalRef\"\u003e7\u003c/span\u003e. The findings show that the models adopted a \u0026ldquo;de-targeting\u0026rdquo; strategy, reducing the frequency of group-specific words and the plural pronoun \u0026ldquo;they\u0026rdquo; to weaken attack targets, while increasing non-personal pronouns linked to bias, which dissolved explicit targeting but left bias residue. In the emotional dimension, they reduced exclamation marks and increased neutral words to lower emotional intensity, yet their internal logic for achieving this differed, leading to conflicting strategies. For language style, the models reduced absolutist words and increased prepositions, which eased extreme expression, but their \u0026ldquo;increased complexity\u0026rdquo; approach\u0026mdash;raising the use of commas and discrepancy words\u0026mdash;unintentionally heightened concealment, resulting in eased extremes but stronger masking. In the social interaction dimension, the models increased constructive words and reduced power-related words to lower social aggressiveness. However, their key \u0026ldquo;dilution\u0026rdquo; strategy\u0026mdash;expanding total text length and vocabulary to weaken hateful content\u0026mdash;may indicate a risk displacement effect, as text length is also linked to hate speech detection, meaning reduced aggression but shifted risk. These results answer RQ3, showing that LLMs can reshape IHS-related strategies, but the reshaping is not a full \u0026ldquo;removal\u0026rdquo; and instead resembles a complex \u0026ldquo;replacement\u0026rdquo;. The models blunt the sharpest edges of hate speech, but at the cost of introducing more ambiguous and covert language risks.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab7\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 7\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eMean Differences in Linguistic Cues Between IHS and AUTO\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"8\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eConcept\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRank\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCategory in LIWC\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eInternal indicator\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eImportance\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eRelationship between indicator and IHS\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eMean Percentage Difference (AUTO-CoT, Deepseek)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003eMean Percentage Difference (AUTO-CoT, GPT-4o)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"6\" rowspan=\"7\"\u003e \u003cp\u003eGroup-specific harm\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCulture\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eEthnicity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e-79.939\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e-82.555\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePerception\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eVisual\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e-76.912\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e-79.172\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePhysical\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePhysical\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e-74.291\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e-74.685\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePronouns\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eThey\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e-39.022\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e-46.710\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCulture\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRelig\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e-74.081\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e-77.541\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e17\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eLinguistic Dimensions\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eIpron\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e17\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e62.156\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e37.993\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e24\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePerception\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePerception\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e24\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e-28.066\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e-17.265\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"8\" rowspan=\"9\"\u003e \u003cp\u003eImplicit emotional\u003c/p\u003e \u003cp\u003eexpression\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e16\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eLinguistic Dimensions\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAdj\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e16\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e-6.249\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e7.998\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e12\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePunctuation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eExclam\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e12\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e-92.396\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e-89.045\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e13\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePsychological Processes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAffect\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e13\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e4.243\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e-4.022\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e19\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePunctuation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAllPunc\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e19\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e-36.087\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e-38.235\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e23\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSummary Variables\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eTone\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e23\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eNegative\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e133.165\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e147.157\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePunctuation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eComma\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e182.547\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e129.378\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e14\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePsychological Processes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eDiscrep\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e14\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e3.542\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e10.839\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e20\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eLinguistic Dimensions\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePrep\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e20\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eNegative\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e32.463\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e31.673\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e21\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePsychological Processes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAllnone\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e21\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e-13.646\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e-24.644\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"5\" rowspan=\"6\"\u003e \u003cp\u003eLinguistic obfuscation and extremity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e10\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTime Orientation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eFocuspast\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e10\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eNegative\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e-11.693\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e-3.291\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e11\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePsychological Processes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eDrives\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e11\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e28.233\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e25.806\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSocial Processes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSocbehav\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eNegative\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e36.467\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e35.350\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e15\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSocial Processes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eComm\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e15\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eNegative\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e35.618\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e62.061\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e22\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eLifestyle\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eWork\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e22\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eNegative\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e70.342\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e71.093\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e25\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePsychological Processes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePower\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e25\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e-24.599\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e-24.369\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eBias and implication in social interaction\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSummary Variables\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eWc\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e49.631\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e15.672\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSummary Variables\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eDic\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eNegative\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e1.870\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e2.289\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e18\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCognition\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSpace\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e18\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e-21.713\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e2.754\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"8\"\u003eNote:(Average percentage differences in linguistic cue statistics between AUTO-CoT and the original text are shown. Positive cues increase IHS aggressiveness, while Negative cues reduce aggressiveness. A positive value means AUTO-CoT used the feature more frequently; a negative value means the original text used it more frequently.)\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003eResult Validation: External Validity Test Based on Toxicity Scores\u003c/h2\u003e \u003cp\u003eWe calculated the Toxicity Score for the original IHS tweets and their LLM-reformed versions, and conducted paired-sample t-tests and Tukey HSD tests to analyse score differences. The overall results (see Table\u0026nbsp;\u003cspan refid=\"Tab8\" class=\"InternalRef\"\u003e8\u003c/span\u003e and Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e) show that LLM reformulation significantly reduced textual toxicity. Compared with the original tweets, Deepseek reduced toxicity by an average of 86.21%, while GPT-4o achieved a reduction of 90.57%. The differences in toxicity scores for both models reached a highly significant statistical level (p\u0026thinsp;\u0026lt;\u0026thinsp;0.001). In terms of distribution, the median toxicity scores and overall score ranges of the reformed texts showed a clear downward trend.\u003c/p\u003e\u003cp\u003eTable 8\u0026emsp;Average Percentage Difference in Toxicity Scores and Significance Tests: OL AUTO-CoT vs. OL\u003c/p\u003e\u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Taba\" border=\"1\"\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLLM\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAverage Percentage Difference*\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003et-value\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003ep-value\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDeepseek\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e-86.210\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e166.082\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.000\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGPT-4o\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e-90.568\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e174.814\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.000\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eTo further examine model performance across different contexts, this study repeated the toxicity score comparison for each IHS topic category (Table\u0026nbsp;\u003cspan refid=\"Tab9\" class=\"InternalRef\"\u003e9\u003c/span\u003e and Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e). The analysis shows that in all topic categories, LLMs exhibited strong toxicity reduction, with average decreases ranging from 81.64% to 94.7%. Notably, both models performed exceptionally well in high-risk categories, but with some performance differences. GPT-4o achieved the greatest reduction in Threatening (\u0026minus;\u0026thinsp;94.7%),Derogation (\u0026minus;\u0026thinsp;92.7%), and Stereotypical (\u0026minus;\u0026thinsp;90.1%) categories. DeepSeek also showed strong results in Threatening (\u0026minus;\u0026thinsp;93.7%) and Derogation (\u0026minus;\u0026thinsp;90.4%). Overall, under the AUTO-CoT prompting framework, GPT-4o generally outperformed DeepSeek, particularly in handling high-risk, highly aggressive IHS.\u003c/p\u003e\u003cp\u003eTable 9 Mean Percentage Differences in Toxicity Reduction Effects of LLM\u0026ndash;Prompt Combinations Across Different Topic Types\u003c/p\u003e\u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Tabb\" border=\"1\"\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eTopic Type\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e \u003cp\u003eMean Percentage Difference\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003e(AUTO-CoT, Deepseek)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003e(AUTO-CoT, GPT-4o)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eIncitement\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e-86.657\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e-90.705\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eInferiority\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e-90.392\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e-92.681\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eIrony\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e-85.983\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e-88.421\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOther\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e-81.636\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e-89.639\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStereotypical\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e-85.06\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e-90.118\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eThreatening\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e-93.6651\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e-94.654\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eWhite grievance\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e-83.719\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e-89.948\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eThis study examines the ability, strategies, and potential limitations of LLMs in reshaping online IHS. The results confirm that LLMs can reshape IHS. However, this process is not a simple purification, but rather a complex displacement of risk.\u003c/p\u003e \u003cdiv id=\"Sec20\" class=\"Section2\"\u003e \u003ch2\u003eEffectiveness and Complexity of LLMs as Tools for Governing IHS\u003c/h2\u003e \u003cp\u003eThis study first addresses RQ1 by confirming that LLMs, represented by GPT-4o and DeepSeek, can effectively reduce the hateful attributes of IHS while largely preserving the core semantics of the original text. This finding provides empirical support for adopting \u0026ldquo;speech reshaping\u0026rdquo; as a constructive intervention, surpassing the systemic limitations of traditional moderation when addressing core IHS strategies such as linguistic obfuscation and bias cues in social interaction (Hartmann, et al., \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2025\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe core contribution, however, lies in revealing the underlying mechanisms of this reshaping process. By constructing and applying the IHS linguistic behavior evaluation framework (RQ2), this study examined the specific linguistic operations performed by LLMs, thereby answering RQ3. The analysis shows that LLM reshaping is not a complete elimination but a complex displacement. The models successfully blunt the most direct attacks in hate speech, but at the cost of introducing new, more covert linguistic risks. This suggests that current LLM handling of IHS is closer to pattern-based risk avoidance than to value alignment grounded in deep understanding. The models learn to identify and replace language features strongly associated with hate labels, yet fail to fully grasp the socio-cultural connotations and bias roots behind these features (Ghosh, et al., \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2023\u003c/span\u003e;Lin, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) .\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec21\" class=\"Section2\"\u003e \u003ch2\u003eFour Types of Language Changes in the Reshaping Process\u003c/h2\u003e \u003cp\u003eThis study found four kinds of replacement behaviors that show the inner limits of LLM reshaping strategies. First, target removal but bias remains. The models cut direct mentions of certain groups, an effective \u0026ldquo;de-targeting\u0026rdquo; move. But they added impersonal pronouns linked to bias, turning clear targeting into vague bias hints. This may lower scores in automatic checks, but may not remove the bias that readers can still understand in context (Ghosh, et al., \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2023\u003c/span\u003e;Schmeisser-Nieto, et al., 2022) .\u003c/p\u003e \u003cp\u003eSecond, lower intensity but mixed methods. The models reduced exclamation marks and used more neutral words to soften tone. Yet they used different and sometimes opposite ways to do this. This shows that emotion control is still at a surface level, without a steady and clear method, mostly just cutting visible emotion signs.\u003c/p\u003e \u003cp\u003eThird, less extreme but more hidden. The models used fewer absolute words, making text less extreme. But their \u0026ldquo;safe output\u0026rdquo; method of adding complexity\u0026mdash;more commas and mismatch words\u0026mdash;made the message more hidden. This is like \u0026ldquo;dog-whistle\u0026rdquo; or coded speech (Marten Risius, \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2024\u003c/span\u003e), where complex form hides true meaning.\u003c/p\u003e \u003cp\u003eFourth, weaker attacks but risk moved. The models used more positive words and fewer power words to cut social attack. But their main \u0026ldquo;dilution\u0026rdquo; method\u0026mdash;adding length and extra words\u0026mdash;may create new risks. Longer text can also link to hate speech labels. This \u0026ldquo;padding\u0026rdquo; may be a hidden form of risk change.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec22\" class=\"Section2\"\u003e \u003ch2\u003eTheoretical and Practical Implications\u003c/h2\u003e \u003cp\u003eTheoretically, this study builds and applies an IHS Speech Behavior Assessment Framework to reveal the inner tensions in LLMs\u0026rsquo;\u0026ldquo;detoxification\u0026rdquo; process. It introduces the concept of risk displacement, challenging the simplified view of this process as a one-way purification. By quantifying and operationalizing the complex attack strategies of IHS (China, \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2019\u003c/span\u003e;Ghosh, et al., \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2023\u003c/span\u003e;Lemmens, et al., \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2021\u003c/span\u003e), the framework offers a replicable method for opening the \u0026ldquo;black box\u0026rdquo; of LLM decision-making (Mathew, et al., \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2021\u003c/span\u003e) and for analyzing their specific behaviors when handling socially sensitive language tasks, thereby advancing research in explainable AI governance.\u003c/p\u003e \u003cp\u003ePractically, the study provides empirical support for shifting online content governance from traditional \u0026ldquo;censorship\u0026rdquo; to a restorative paradigm of \u0026ldquo;speech reshaping\u0026rdquo; (Hartmann, et al., \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2025\u003c/span\u003e;Johnson, et al., \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). More importantly, the identified language replacement paths\u0026mdash;such as \u0026ldquo;target removal but bias remains\u0026rdquo; and \u0026ldquo;less extreme but more hidden\u0026rdquo;\u0026mdash;serve as a direct audit checklist for platforms and developers. These findings call for governance to go beyond simple technical deployment toward fine-grained management that incorporates continuous risk monitoring and secondary review, offering clear guidance for the responsible use of this technology.\u003c/p\u003e \u003cdiv id=\"Sec23\" class=\"Section3\"\u003e \u003ch2\u003eLimitations and Future Directions\u003c/h2\u003e \u003cp\u003eThis study has several limitations. First, it only covers two specific LLMs, so the findings may not apply to other models. Second, the dataset is based mainly on English tweets. Future work should test the framework on more languages and different types of online texts to check its validity across cultures and platforms.\u003c/p\u003e \u003cp\u003eFuture research can take several paths. One is to study how these \u0026ldquo;replaced\u0026rdquo; language risks affect real human perception. Does reshaped, more vague text lower readers\u0026rsquo; guard and spread bias in subtle ways? Another is to develop new training or fine-tuning methods to solve the problem of risk displacement at its root, guiding models to remove hate without adding new risks. A third is to combine the framework with cognitive experiments to better understand the social and psychological mechanisms behind IHS language strategies.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec24\" class=\"Section2\"\u003e \u003ch2\u003eEthical Risks and Potential Biases\u003c/h2\u003e \u003cp\u003eUsing generative AI to reshape IHS raises serious ethical concerns. The main risk is that models may inherit and amplify cultural and social biases from their training data (Lin, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2022\u003c/span\u003e;Yadav and Singh, \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). This can lead to misunderstanding non-mainstream contexts or reproducing stereotypes about marginalized groups, reinforcing inequality. The lack of transparency in LLM decision-making\u0026mdash;the \u0026ldquo;algorithmic black box\u0026rdquo; (Roychowdhury and Gupta, \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) \u0026mdash;also makes it hard for users to know why their speech was changed and what values guided the change. Without transparency and fair appeal mechanisms, such hidden language control can threaten freedom of expression and the right to information. Any purely technical solution must be part of a broader socio-technical governance system that includes ethical review, bias audits, and user rights protection, to prevent solving old problems while creating new and hidden risks.\u003c/p\u003e \u003c/div\u003e"},{"header":"Conclusion","content":"\u003cp\u003eThis study used an integrative modelling approach to examine the ability and mechanisms of LLMs in reshaping IHS. It first confirmed that LLMs can effectively reduce the hateful features of IHS while keeping its core meaning. However, with the IHS Speech Behavior Assessment Framework developed in this study, the core finding shows that the reshaping process is not a simple removal but a complex risk displacement. While LLMs can weaken direct attacks, they may also introduce new and more hidden risks through strategies like vague targeting and greater language masking.\u003c/p\u003e \u003cp\u003eThe contributions are both theoretical and practical. Theoretically, the concept of risk displacement deepens understanding of the \u0026ldquo;detoxification\u0026rdquo; process, and the framework provides an explainable tool for future research in this area. Practically, the study supports \u0026ldquo;speech reshaping\u0026rdquo; as a governance model but also defines its risk boundaries. The identified replacement paths can serve as a direct checklist for technical audits. Future research should go beyond performance evaluation and focus on resolving these replaced language risks, as well as exploring better human\u0026ndash;AI collaboration in governance, to align technological development with social values.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eYinghui Huang, Qixia Feng and Hui Liu were responsible for conceptualization, methodology, and writing\u0026mdash;original draft preparation. Hui Liu, Weiqing Li, Zongkui Zhou, and Ying Ma contributed to writing\u0026mdash;review and editing. Yinghui Huang, Weiqing Li and Ying Ma acquired the funding. All authors have read and agreed to the published version of the manuscript.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eThis study uses the publicly available Latent Hatred benchmark dataset (ElSherief et al., 2021), available at: https://www.dropbox.com/scl/fi/76gx54lbv9dnbqr7nmy2c/implicit-hate-corpus.zip?rlkey=qxp26bhci0v8o0tldzxot1ref\u0026amp;e=1\u0026amp;dl=0. The LLM-reshaped implicit hate speech texts generated in this study using GPT-4o and DeepSeek-V2 are available from the corresponding author upon reasonable request.\u003c/p\u003e\u003ch1\u003eAcknowledgments\u003c/h1\u003e\n\u003cp\u003eThis research was supported by the National Natural Science Foundation of China (Grant Nos. 72204095 and 72304090), the Humanities and Social Science Young Scientist Program sponsored by the Ministry of Education of the People’s Republic of China (Grant No. 22YJC880022), the Fundamental Research Funds for the Central Universities (Grant No. 2024VA059), and the Ministry of Education of Humanities and Social Science project (Grant No. 23YJAZH098).\u003c/p\u003e\n\u003ch1\u003eArtificial Intelligence usage\u003c/h1\u003e\n\u003cp\u003eDuring the preparation of this manuscript, we used an artificial intelligence (AI) tool (ChatGPT, OpenAI) to improve the readability and style of the text. The AI assistance was limited to language polishing (grammar, spelling, punctuation, and tone). All content was generated by the authors, and the final version was thoroughly reviewed and approved by the authors to ensure accuracy and integrity.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eAchiam J, Adler S, Agarwal S, Ahmad L, Akkaya I, Aleman FL et al (2023) Gpt-4 technical report. \u003cem\u003earXiv preprint arXiv:2303.08774\u003c/em\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBolukbasi T, Chang K-W, Zou JY, Saligrama V, Kalai AT (2016) Man is to computer programmer as woman is to homemaker? debiasing word embeddings. \u003cem\u003eAdvances in neural information processing systems 29\u003c/em\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCepollaro B, Lepoutre M, Simpson RM (2023) Counterspeech. Philos Compass 18:e12890\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChina CA (2019) o. \u003cem\u003eProvisions on the Governance of the Online Information Content Ecosystem [网络信息内容生态治理规定]\u003c/em\u003e. Institution, Beijing\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCouncil of the European Union, E. P (2022) Regulation (EU) 2022/2065 on a Single Market For Digital Services and amending Directive 2000/31/EC (Digital Services Act). Institution\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eElSherief M, Ziems C, Muchlinski D, Anupindi V, Seybolt J, De Choudhury M et al (2021) Latent hatred: A benchmark for understanding implicit hate speech. \u003cem\u003earXiv preprint arXiv:2109.05322\u003c/em\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGhosh S, Suri M, Chiniya P, Tyagi U, Kumar S, Manocha D (2023) CoSyn: Detecting implicit hate speech in online conversations using a context synergized hyperbolic network. \u003cem\u003earXiv preprint arXiv:2303.03387\u003c/em\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGuo D, Yang D, Zhang H, Song J, Zhang R, Xu R et al (2025) Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. \u003cem\u003earXiv preprint arXiv:2501.12948\u003c/em\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHartmann D, Oueslati A, Staufer D, Pohlmann L, Munzert S, Heuer H (2025) Lost in moderation: How commercial content moderation apis over-and under-moderate group-targeted hate speech and linguistic variations. In: (eds) \u003cem\u003eProceedings of the 2025 CHI Conference on Human Factors in Computing Systems\u003c/em\u003e, \u003cem\u003evol\u003c/em\u003e pp. 1\u0026ndash;26\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHe Z, Majumder BP (2021) Detect and perturb: Neutral rewriting of biased and sensitive text via gradient-based decoding. \u003cem\u003earXiv preprint arXiv:2109.11708\u003c/em\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHee MS, Chong W-H, Lee RK-W (2023) Decoding the underlying meaning of multimodal hateful memes. \u003cem\u003earXiv preprint arXiv:2305.17678\u003c/em\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHofman JM, Watts DJ, Athey S, Garip F, Griffiths TL, Kleinberg J et al (2021) Integrating explanation and prediction in computational social science. Nature 595:181\u0026ndash;188\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHu Z, Lee RK-W, Aggarwal CC, Zhang A (2022) Text style transfer: A review and experimental evaluation. ACM SIGKDD Explorations Newsl 24:14\u0026ndash;45\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJohnson NF, Leahy R, Restrepo NJ, Vel\u0026aacute;squez N, Zheng M, Manrique P et al (2019) Hidden resilience and adaptive dynamics of the global online hate ecology. Nature 573:261\u0026ndash;265\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKim Y, Park S, Han Y-S (2022) Generalizable implicit hate speech detection using contrastive learning. In: (eds) \u003cem\u003eProceedings of the 29th international conference on computational linguistics\u003c/em\u003e, \u003cem\u003evol\u003c/em\u003e pp. 6667\u0026ndash;6679\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKostiuk Y, Tonja AL, Sidorov G, Kolesnikova O Reframing social media discourse: Converting hate speech to non-hate speech. J Intell Fuzzy Syst : JIFS\u0026ndash;219348\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLeague A-D (2024) \u003cem\u003eOnline Hate and Harassment: The American Experience 2024\u003c/em\u003e. Report No. 26\u0026ndash;42\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLemmens J, Markov I, Daelemans W (2021) Improving hate speech type and target detection with hateful metaphor features. In: (eds) \u003cem\u003eProceedings of the Fourth Workshop on NLP for Internet Freedom: Censorship, Disinformation, and Propaganda\u003c/em\u003e, \u003cem\u003evol\u003c/em\u003e pp. 7\u0026ndash;16\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLin J (2022) Leveraging world knowledge in implicit hate speech detection. \u003cem\u003earXiv preprint arXiv:2212.14100\u003c/em\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMagu R, Luo J (2018) Determining code words in euphemistic hate speech using word embedding networks. In: (eds) \u003cem\u003eProceedings of the 2nd workshop on abusive language online (ALW2)\u003c/em\u003e, \u003cem\u003evol\u003c/em\u003e pp. 93\u0026ndash;100\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMarten Risius MN (2024) Substitution: Extremists\u0026rsquo; New Form of Implicit Hate Speech to Avoid Detection. Saeed Akhlaghpour and Hetiao (Slim) Xie. Global Network on Extremism \u0026amp; Technology\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMathew B, Saha P, Yimam SM, Biemann C, Goyal P, Mukherjee A (2021) Hatexplain: A benchmark dataset for explainable hate speech detection. In: (eds) \u003cem\u003eProceedings of the AAAI conference on artificial intelligence\u003c/em\u003e, \u003cem\u003evol 35\u003c/em\u003e. pp. 14867\u0026ndash;14875\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMerriam-Webster (2022)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNations U (2019) United Nations Strategy and Plan of Action on Hate Speech. Report No. United Nations\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOcampo NB, Cabrio E, Villata S (2023) Unmasking the hidden meaning: Bridging implicit and explicit hate speech embedding representations. In: (eds) \u003cem\u003eFindings of the Association for Computational Linguistics: EMNLP 2023\u003c/em\u003e, \u003cem\u003evol\u003c/em\u003e pp. 6626\u0026ndash;6637\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePryzant R, Martinez RD, Dass N, Kurohashi S, Jurafsky D, Yang D (2020) Automatically neutralizing subjective bias in text. In: (eds) \u003cem\u003eProceedings of the aaai conference on artificial intelligence\u003c/em\u003e, \u003cem\u003evol 34\u003c/em\u003e. pp. 480\u0026ndash;489\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRoy A, Khanna D, Mahapatra D, Das A (2024) Do the Right Thing, Just Debias! Multi-Category Bias Mitigation Using LLMs. \u003cem\u003earXiv preprint arXiv:2409.16371\u003c/em\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRoychowdhury S, Gupta V (2023) Data-efficient methods for improving hate speech detection. In: (eds) \u003cem\u003eFindings of the Association for Computational Linguistics: EACL 2023\u003c/em\u003e, \u003cem\u003evol\u003c/em\u003e pp. 125\u0026ndash;132\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSaakyan A, Muresan S (2023) Iclef: In-context learning with expert feedback for explainable style transfer. \u003cem\u003earXiv preprint arXiv:2309.08583\u003c/em\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSap M, Card D, Gabriel S, Choi Y, Smith NA (2019) The risk of racial bias in hate speech detection. In: (eds) \u003cem\u003eProceedings of the 57th annual meeting of the association for computational linguistics\u003c/em\u003e, \u003cem\u003evol\u003c/em\u003e pp. 1668\u0026ndash;1678\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSchmeisser-Nieto W, Nofre M (2022) Criteria for the annotation of implicit stereotypes. In: (eds) \u003cem\u003eProceedings of the Thirteenth Language Resources and Evaluation Conference\u003c/em\u003e, \u003cem\u003evol\u003c/em\u003e pp. 753\u0026ndash;762\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWen J, Ke P, Sun H, Zhang Z, Li C, Bai J et al (2023) Unveiling the implicit toxicity in large language models. \u003cem\u003earXiv preprint arXiv:2311.17391\u003c/em\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWiegand M, Kampfmeier J, Eder E, Ruppenhofer J (2023) Euphemistic abuse\u0026ndash;a new dataset and classification experiments for implicitly abusive language. In: (eds) \u003cem\u003eProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing\u003c/em\u003e, \u003cem\u003evol\u003c/em\u003e pp. 16280\u0026ndash;16297\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYadav A, Singh V (2024) HateFusion: Harnessing Attention-Based Techniques for Enhanced Filtering and Detection of Implicit Hate Speech. IEEE Trans Comput Social Syst\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYuan S, Nie E, Kouba L, Kangen AY, Schmid H, Sch\u0026uuml;tze H et al (2025) LLM in the Loop: Creating the PARADEHATE Dataset for Hate Speech Detoxification. \u003cem\u003earXiv preprint arXiv:2506.01484\u003c/em\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang M, He J, Ji T, Lu C-T (2024) Don't Go To Extremes: Revealing the Excessive Sensitivity and Calibration Limitations of LLMs in Implicit Hate Speech Detection. \u003cem\u003earXiv preprint arXiv:2402.11406\u003c/em\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhao J, Wang T, Yatskar M, Ordonez V, Chang K-W (2018) Gender bias in coreference resolution: Evaluation and debiasing methods. \u003cem\u003earXiv preprint arXiv:1804.06876\u003c/em\u003e\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Implicit Hate Speech, Large Language Models, Linguistic Style Reshaping, Integrative Modelling, Explainable Artificial Intelligence","lastPublishedDoi":"10.21203/rs.3.rs-7368894/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7368894/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eImplicit Hate Speech (IHS) presents major challenges for traditional content governance. Using large language models (LLMs) for reshaping has become a promising new approach. This study evaluates the ability, strategies, and risks of LLMs in reshaping IHS. We use an integrative modelling method. First, we build a predictive model to measure the external effect of LLM reshaping. Second, we design an explainable evaluation framework with four dimensions: group-specific harm, implicit emotional expression, linguistic obfuscation and extremity, and bias and implication in social interaction. Results show that LLMs (e.g., GPT-4o and DeepSeek) can strongly reshape IHS texts in topics such as threatening and inferiority, reducing toxicity by 86.2%\u0026ndash;90.57% while keeping high semantic similarity (BERTScore F1: 82%\u0026ndash;85%). However, reshaping is not full detoxification. It often replaces risk with new covert forms. Explicit attacks are reduced, but covert risks may appear through strategies like vague references, hiding emotions, or adding logical gaps. This study confirms the value of LLMs in IHS governance, but also reveals their \u0026ldquo;replace-rather-than-remove\u0026rdquo; pattern. The framework we propose is a useful tool to detect and manage covert risks caused by algorithms, offering both theoretical and practical guidance for creating a more civil online space.\u003c/p\u003e","manuscriptTitle":"Can large language models effectively reshape online implicit hate speech? An integrative modelling approach","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-01-14 15:18:43","doi":"10.21203/rs.3.rs-7368894/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"a48f0f9a-c9f6-42e8-9396-2ff61cbf4508","owner":[],"postedDate":"January 14th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":61060654,"name":"Humanities/Complex networks"},{"id":61060655,"name":"Social science/Complex networks"},{"id":61060656,"name":"Physical sciences/Mathematics and computing"},{"id":61060657,"name":"Biological sciences/Psychology"},{"id":61060658,"name":"Social science/Psychology"}],"tags":[],"updatedAt":"2026-03-01T15:09:22+00:00","versionOfRecord":[],"versionCreatedAt":"2026-01-14 15:18:43","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7368894","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7368894","identity":"rs-7368894","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}