BCGH-Net: A Hierarchical Neural Framework for Fake News Detection Using BERT and Attention Fusion

BCGH-Net: A Hierarchical Neural Framework for Fake News Detection Using BERT and Attention Fusion | 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 Research Article BCGH-Net: A Hierarchical Neural Framework for Fake News Detection Using BERT and Attention Fusion Hussein Al-kaabi, Fuqdan Al-ibrahimi, Ali kadhim jasim, Zainab S. Idan This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6893733/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 The proliferation of misinformation and fake news on digital platforms has emerged as a critical challenge, undermining public trust and exacerbating societal polarization. Traditional verification methods are inadequate for rapid information dissemination on social media, necessitating advanced automated solutions. This paper proposes a novel fake news detection model, BCGH-Net, which integrates BERT embeddings with a multi-level feature extraction framework. The model employs Convolutional Neural Networks (CNN) for word-level feature extraction and Bidirectional Gated Recurrent Units (Bi-GRU) for sequential sentence-level dependencies, fused via a Hierarchical Attention Network (HAN) to prioritize salient textual elements. Evaluated on four benchmark datasets (FakeNewsNet, ISOT, CoAID, and ReCOVery), BCGH-Net achieves superior performance across four benchmark datasets, demonstrating its robustness and generalizability, with accuracies of 95.8%, 99.3%, 98.7%, and 92.4%, respectively. The results demonstrate its robustness across diverse domains, including political and health-related misinformation. Key contributions include a two-stage feature extraction approach, hierarchical attention fusion, and superior performance metrics (precision, recall, F1-score) compared to existing models. Despite its computational complexity, BCGH-Net offers a scalable solution for real-world deployment, addressing the urgent need for reliable fake news detection systems. Fake News Detection BERT Embedding Hierarchical Attention Network (HAN) Deep Learning Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 1. Introduction The spread of misinformation and fake news has been a significant concern for societies throughout history. While the advent of print media in the 15th century marked a considerable revolution in information dissemination, it also posed challenges regarding the accuracy and credibility of the information shared. The mass media's rise in the 20th century deepened these concerns, as the large-scale dissemination of factual and misleading information became more accessible [ 1 ]. However, the emergence of digital platforms, mainly social media, in the early 21st century dramatically accelerated the spread of fake news. The ease of content creation and distribution on platforms like Facebook, Twitter, and Instagram has made it increasingly difficult to differentiate between legitimate and misleading information [ 2 ]. The detection of fake news has thus become a pressing issue. If left unchecked, misinformation can erode trust in institutions, exacerbate societal polarization, and shape political outcomes [ 3 ]. Traditional verification methods are no longer sufficient in the age of social media, where news spreads rapidly and can be subtly manipulated. Consequently, there is an urgent need for automated systems that can efficiently detect fake news and provide the public with accurate and reliable information [ 4 ]. Artificial Intelligence (AI), specifically Natural Language Processing (NLP), has emerged as a powerful tool in the fight against fake news. With advances in machine learning algorithms and deep learning architectures, AI can now analyze vast amounts of textual data, identify patterns, and discern subtle cues that distinguish real news from fake news [ 5 ]. Techniques such as sentiment analysis [ 6 ] and spam detection [ 7 ] have enabled AI models to evaluate news content unimaginably. This paper proposes a novel approach for fake news detection that leverages BERT (Bidirectional Encoder Representations from Transformers), a state-of-the-art contextual embedding model [ 8 ]. Our method incorporates a multi-level architecture that encompasses domain-specific pre-processing, feature extraction utilizing Convolutional Neural Networks (CNN) and Bidirectional Gated Recurrent Units (Bi-GRU), and fusion via a Hierarchical Attention Network (HAN). We aim to enhance the performance of fake news detection models by incorporating social media-specific features, such as hashtags, emojis, and user mentions, and by fine-tuning BERT to improve its understanding of fake news in different contexts. Our work contributes to developing more accurate and efficient systems for detecting fake news, which can be deployed on real-world social media platforms. The key contributions of this paper are as follows: Multi-Level Feature Extraction: We develop a two-stage feature extraction approach that combines CNN for local word-level features and Bi-GRU for sequential sentence-level dependencies. Hierarchical Attention Fusion: We employ a Hierarchical Attention Network (HAN) to fuse multi-granular features effectively, allowing the model to focus on the most informative words and sentences. Superior Performance on Benchmark Datasets: The proposed model significantly improves fake news detection accuracy, precision, recall, and F1-score compared to state-of-the-art methods on the FakeNewsNet, ISOT, CoAID, and ReCOVery datasets. The remainder of this paper is organized as follows: Section 2 reviews the related work in fake news detection, particularly the role of AI and NLP. Section 3 describes the methodology employed in this study, including the data pre-processing, feature extraction, and model architecture. Section 4 presents our experiments' experimental setup, dataset, and results. Section 5 discusses the implications of our findings, and Section 6 concludes the paper with suggestions for future work in the field. 2. Related work The growing prevalence of fake news has led to extensive research into automated detection methods, with machine learning (ML) and deep learning (DL) emerging as dominant approaches. ML techniques rely on handcrafted features and traditional classifiers, while DL models leverage neural networks for feature extraction and classification [ 9 ]. This section reviews recent advancements in ML- and DL-based fake news detection methodologies. 2.1 Machine Learning-Based Approaches Recent studies have explored machine learning (ML) techniques for fake news detection using various feature extraction and classification methods. Traditional ML models, such as SVM, LR, and KNN, have been employed with TF-IDF, chi-square, and ANOVA for feature selection [ 10 – 12 ]. However, ML-based approaches have limitations, including reliance on handcrafted features that may miss semantic nuances, sensitivity to training data quality, and difficulty capturing deep contextual relationships. These challenges have driven the shift toward deep learning, which offers automatic feature extraction and improved accuracy [ 13 ]. 2.2 Deep Learning-Based Approaches Deep learning (DL) techniques have significantly enhanced text classification tasks by integrating CNN, RNN, and transformer-based models. Saleh et al. [ 14 ] proposed an Optimized Convolutional Neural Network (OPCNN-FAKE) model to enhance fake news detection through deep learning. The study evaluated its performance against traditional machine learning models (Decision Tree, Logistic Regression, K-Nearest Neighbor, Random Forest, Support Vector Machine, and Naïve Bayes) as well as deep learning models (Recurrent Neural Network and Long Short-Term Memory) across four benchmark datasets: Kaggle Fake News Dataset [ 15 ], FakeNewsNet [ 16 ], FA-KES [ 17 ], and ISOT [ 18 ], demonstrating its effectiveness in improving detection accuracy. Kaliyar et al. [ 19 ] proposed FNDNet, a deep Convolutional Neural Network (CNN) model explicitly designed for fake news detection. Unlike traditional approaches that rely on hand-crafted features, FNDNet can automatically learn discriminatory features for fake news classification through its multiple hidden layers by leveraging CNN’s hierarchical feature extraction capabilities. Nithya et al. [ 20 ] proposed a two-phase feature extraction approach, combining BERT-OPCNN for deep feature extraction and fastText with ant colony optimization (FIAC) for word embedding. A customized Bi-LSTM classifies the resulting vectors, achieving 55.69% accuracy on the LIAR [ 21 ] dataset and 98.24% on the ISOT dataset, demonstrating its effectiveness in fake news detection. Almarashy et al. [ 22 ] suggest a new model to detect fake news by extracting and combining global features (using TF-IDF), spatial features (using CNN), and temporal features (using BiLSTM). These features are then classified using a Fast Learning Network (FLN). The model was tested on two publicly available fake news datasets: ISOT and FA-KES. Kaliyar et al. [ 23 ] proposed FakeBERT, a deep learning model to tackle ambiguity in natural language processing. By integrating BERT with advanced deep learning methods, FakeBERT achieves an impressive accuracy of 98.90% for the fake news dataset, surpassing the performance of existing models. Luvembee et al. [ 24 ] proposed Complementary Attention Fusion with an Optimized Deep Neural Network (CAF-ODNN), which captures subtle cross-modal relationships using image captioning and bidirectional attention. CAF-ODNN improves feature extraction and fusion, outperforming existing methods on real-world datasets containing tweets and fact-checked news from PolitiFact and GossipCop websites. Truică et al. [ 25 ] proposed a Deep Neural Network Ensemble Architecture for social and textual context-aware fake news detection (DANES). DANES integrates a Text Branch for textual content and a Social Branch for social context, creating a novel Network Embedding. Preliminary ablation studies on three real-world datasets, such as Twitter15 and Twitter16, demonstrate its effectiveness. Thaokar et al. [ 26 ] proposed an approach comprising basic natural language processing techniques with different deep learning models for categorizing content as real or fake. LSTM, Bi-LSTM, and Bi-LSTM with an attention mechanism are employed. The study compared results using one-hot word embedding and a pre-trained GloVe technique. On the benchmark LIAR dataset, the LSTM achieved an accuracy of 67.2%, while Bi-LSTM with GloVe embedding reached 67%. An accuracy of 98.22% was achieved using Bi-LSTM, and 97.98% using LSTM on the Real-Fake dataset. Shu et al. [ 27 ] introduced FakeNewsNet, a comprehensive data repository integrating news content, social context, and dynamic information, achieving approximately 65% accuracy with baseline models. In a comparative analysis, Qian et al. [ 28 ] demonstrated that Graph Neural Networks and Transformer-based architectures can achieve up to 86.16% accuracy on the same dataset. Zhang et al. [ 29 ] proposed MCNNet, a multichannel neural network model that surpassed 90% accuracy by effectively integrating multiple modalities. Similarly, Ghenai et al. [ 30 ] improved fake news detection by converting text into image representations, achieving 99.3% accuracy on the ISOT dataset. While highly effective, this method requires significant computational resources, limiting its practicality in low-resource or real-time settings. Addressing COVID-19 misinformation, Wang et al. [ 31 ] developed a machine-learned solution utilizing the CoAID dataset, employing a Bi-LSTM model trained on 1,257 entries and tested on 419 entries. This approach achieved a weighted average F1-score of 0.93, with equal precision, recall, and accuracy across both labels. Khan et al. [ 32 ] conducted a focused case study on COVID-19 misinformation detection, achieving an accuracy of 93%. Their study emphasized the effectiveness of machine learning techniques in identifying false information related to the pandemic. Imran et al. [ 33 ] introduced the CoAID-DEEP framework, an optimized intelligent system for automated detection of COVID-19 misleading information on Twitter. This framework achieved an accuracy of 98.57%, demonstrating its efficacy in combating misinformation on social media platforms. Zhou et al. [ 34 ] addressed the challenge of fake news detection under limited information settings using the ReCOVery dataset. They developed a multimodal repository for COVID-19 news credibility research, achieving an accuracy of 86.82%. Their work highlights the importance of integrating various data modalities to enhance detection performance. 3. Proposed Model This section presents the proposed model for fake news detection, which consists of five key steps: pre-processing, embedding generation using BERT, two-level feature extraction, feature fusion using a Hierarchical Attention Network (HAN), and classification, as shown in Figure 1. 3.1. Pre-Processing Step Pre-processing ensures the input data is clean and compatible with BERT's embedding process. The pre-processing steps are: Basic Cleaning : Remove excessive noise such as repetitive hashtags, irrelevant symbols, or unnecessary special characters [35]. Eliminate URLs, email addresses, or other non-textual elements that do not contribute to the context [36]. Handle emojis by converting them to textual representations (e.g., "😂" to "laughing") if they carry meaning [37]. Preserve Casing : Retain the original casing of the text, as BERT is case-sensitive and uses it for contextual understanding (e.g., "Apple" vs. "apple") [38]. Focusing on these cleaning steps ensures that the input data remains compatible with BERT’s tokenizer and embedding process, providing high-quality, context-aware representations. 3.2. BERT for Embedding Step The pre-processed text is passed through the BERT (Bidirectional Encoder Representations from Transformers) model to generate contextual embeddings. BERT's architecture, which leverages a transformer-based bidirectional encoder, captures the semantic relationships between words within a sentence. Unlike traditional word embeddings that generate static representations, BERT produces dynamic, context-aware embeddings that vary depending on the surrounding words [39]. This bidirectional context allows BERT to understand nuanced meanings, such as polysemy (words with multiple meanings), which is crucial for distinguishing subtle differences in fake news content. The embeddings generated by BERT serve as a robust and rich representation of the textual data, providing a solid foundation for subsequent feature extraction. By fine-tuning BERT on our specific fake news dataset, we further enhance its ability to capture domain-specific linguistic patterns by improving the model's overall performance in detecting fake news. This step ensures the model is equipped with high-quality, contextually rich embedding essential for accurate classification. 3.3. Two-Level Feature Extraction Feature extraction is performed in two stages to capture word-level and sentence-level features, ensuring a comprehensive understanding of textual content. This hierarchical approach allows the model to analyze local patterns (e.g., specific word combinations) and global structures (e.g., sentence coherence and context). 3.3.1. Word-level feature extraction A CNN is employed to extract local and spatial features from the word embeddings generated by BERT. The CNN applies multiple filters to detect n-gram patterns, such as word combinations or phrases, which are critical for understanding the text context [7]. For example, in the sentence "The government announced a new policy to reduce taxes," CNN might identify key n-grams like " reduce taxes " or " new policy ," which could be indicative of the sentence's intent or tone. Focusing on these local features, CNN captures subtle linguistic cues often present in fake news, such as exaggerated language or misleading phrases. 3.3.2. Sentence-Level Feature Extraction: A BiGRU extracts sequential and contextual information at the sentence level. Unlike unidirectional models, the BiGRU processes the text in both forward and backward directions, capturing dependencies between past and future words [40]. This bidirectional approach ensures a more comprehensive understanding of the sentence's structure and meaning. For instance, in the sentence "Despite the economic growth, unemployment rates remain high," the BiGRU can capture the contrast between "economic growth" and "unemployment rates," which is crucial for understanding the overall sentiment and context of the sentence. Analyzing the entire sentence, the BiGRU helps the model identify inconsistencies or illogical sequences common in fake news. For Example, consider the following two sentences: "The new vaccine is 100% effective with no side effects." "Experts warn that the new vaccine may have rare side effects." At the word level, CNN might detect n-grams like "100% effective" in the first sentence, which could be a red flag for exaggeration, and "rare side effects" in the second sentence, which reflects a more cautious tone. At the sentence level, the BiGRU would analyze each sentence's overall structure and context. The first sentence, with its absolute claim ( "100% effective" ), might be flagged as suspicious, while the second sentence, which includes a qualifier ( "may have" ), would be considered more credible. By combining word-level and sentence-level feature extraction, the model gains a multidimensional understanding of the text, enabling it to detect both local anomalies and global inconsistencies characteristic of fake news. This two-level approach ensures that the model is sensitive to specific word patterns and capable of understanding the broader context in which those words appear. 3.4. HAN for Feature Fusion A HAN combines the extracted word-level and sentence-level features. The HAN introduces a dual-level attention mechanism that prioritizes the most relevant features at both the word and sentence levels, as shown in Figure 3. This ensures that the model focuses on the key aspects of the text that contribute to the classification task. This hierarchical approach enables the model to dynamically weigh the importance of different words within a sentence and across different sentences within a document, thereby enhancing its ability to detect subtle patterns in fake news [41]. Figure 2 explains the HAN architecture. 3.4.1 Word-Level Attention: The attention mechanism assigns weights to each word at the word level based on its contribution to the sentence's meaning. Words that are more relevant to the context or have stronger semantic meanings are given higher weights. For example, in the sentence "The politician claimed that the election was rigged," the attention mechanism might assign higher weights to words like "rigged" and "claimed," as they are critical to understanding the sentence's intent and potential for misinformation. This ensures the model focuses on the most informative words while downplaying less relevant ones, such as stop or filler. 3.4.2 Sentence-Level Attention : At the sentence level, the attention mechanism evaluates the importance of each sentence within the broader context of the document. Sentences that contain key claims, evidence, or contradictions are given higher weights. For instance, in a news article, sentences like "Experts have debunked the claim that 5G causes COVID-19" would receive more attention than less informative sentences like "The article was published last week." This allows the model to prioritize sentences more likely to contain critical information to detect fake news. 3.4.3 Feature Fusion : The weighted word-level and sentence-level features are then fused into a unified representation. This fused representation captures both the local nuances (from word-level features) and the global context (from sentence-level features), providing a comprehensive understanding of the text. The fused features are passed to the classification layer, where the model makes its final prediction. Consider a fake news article with the following sentences: "A new study proves that drinking bleach cures COVID-19." "The study was conducted by an unknown group with no peer review." "Health officials have warned against this dangerous practice." At the word level, the attention mechanism might highlight words like "bleach," "cures," and "dangerous" as highly relevant due to their strong semantic meaning and potential for misinformation. At the sentence level, the attention mechanism would likely assign higher weights to the first and third sentences, as they contain strong claims and warnings, respectively, while downplaying the second sentence provides context but is less critical for classification. Using the HAN for feature fusion, the model effectively combines the strengths of both word-level and sentence-level analysis, ensuring that it captures the most relevant information for accurate fake news detection. This hierarchical attention mechanism improves the model's performance. It enhances its interpretability, as the attention weights provide insights into which words and sentences the model considers most important for its decision-making process. 5. Classification Step In the final step, the fused features are fed into a fully connected dense layer, followed by a softmax activation function, to perform classification. The output layer assigns probabilities to the predefined classes (fake or real news), enabling the model to make accurate predictions. The softmax function ensures that the predicted probabilities sum up to one, allowing for a clear interpretation of the classification results [43]. By integrating these steps, the proposed model effectively combines the strengths of deep contextual embeddings, hierarchical feature extraction, and attention mechanisms, resulting in a robust framework for fake news detection. Algorithm: Fake News Detection using two-level feature fusion Input : Labeled news statements Output : Predicted label: Fake or Real Text Preprocessing : Clean and tokenize input using standard NLP techniques to normalize text for modeling. BERT Embeddings : Generate contextualized token representations using a pre-trained BERT model. Feature Extraction : CNN captures local n-gram patterns. Bi-GRU models sequential dependencies across tokens. Feature Fusion : Apply Hierarchical Attention Network (HAN) to integrate CNN and Bi-GRU outputs, emphasizing semantically salient components. Classification : Pass fused features through a dense layer, followed by a softmax function to predict class probabilities. Training and Evaluation : Optimize with cross-entropy loss using Adam. Evaluate using accuracy, precision, recall, and F1-score. Inference : Apply the trained model to predict labels on unseen news statements. 4. Evaluation Specific datasets are recognized as more challenging when evaluating fake news detection models. They may yield lower accuracy due to their complexity, noise, or the subtle nature of the fake news they contain. 4 − 1 datasets To evaluate the performance and robustness of the proposed fake news detection model, we employed four publicly available English-language datasets, each annotated with binary labels (i.e., fake or real). These datasets were chosen to ensure diversity in topical focus, textual style, and data modality. FakeNewsNet, sourced from PolitiFact and GossipCop , includes content and social context, supporting models incorporating user interaction signals [ 44 ]. The ISOT Fake News Dataset provides over 44,000 articles from reliable and unreliable sources, offering balanced coverage of political and general news [ 45 ]. The CoAID dataset contributes COVID-19-specific misinformation data with user engagement metadata, while ReCOVery offers credibility-labeled COVID-19 news annotated by fact-checking services like NewsGuard [ 46 , 47 ]. These datasets vary widely in size and thematic scope from general political discourse to health misinformation, enabling a thorough assessment of the model’s generalizability and real-world applicability. Figure 3 shows the dataset class distribution. 4 − 2 Evaluation metrics We employ four fundamental evaluation metrics to assess the proposed fake news detection model's performance: Accuracy, Precision, Recall, and F1-Measure. These metrics provide a comprehensive assessment of classification effectiveness. 1. Accuracy Accuracy is a fundamental metric that measures the overall correctness of the model’s predictions. It is defined as: $$\:Accuracy\:\left(Acc\right)=\frac{TP+TN}{TP+TN+FP+FN}100\%\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\left(1\right)$$ Where: TP (True Positives): Fake news correctly classified as fake. TN (True Negatives): Real news correctly classified as real. FP (False Positives): Real news incorrectly classified as fake. FN (False Negatives): Fake news incorrectly classified as real [ 48 ]. While accuracy provides an intuitive measure of performance, it can be misleading in cases of imbalanced class distribution. Therefore, additional metrics are required for a more nuanced evaluation. 2. Precision Precision, also known as Positive Predictive Value (PPV), measures the model's reliability in predicting fake news. It is defined as: $$\:Precesion\:\left(Pre\right)=\frac{TN}{TN+FP}100\%\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\left(2\right)$$ A high precision value indicates that the model is likely correct when classifying a news statement as fake. This metric is critical in minimizing false alarms and ensuring that legitimate news is not incorrectly flagged as counterfeit. 3. Recall Recall, called Sensitivity or True Positive Rate (TPR), quantifies the model’s ability to identify fake news correctly. It is given by: $$\:Recall\:\left(Re\right)=\frac{TP}{TP+FN}100\%\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\left(3\right)$$ A high recall value indicates that the model successfully detects a large proportion of actual fake news articles. This metric is particularly critical in applications where missing fake news (false negatives) can lead to misinformation spreading unchecked. 4. F1-Measure : Also called the F1-score, it is the harmonic mean of Precision and Recall, balancing both metrics to provide a single performance measure. F1-Measure is particularly useful in scenarios with class imbalance, where either precision or recall alone might give a misleading picture of performance. A high F1-Measure indicates that the model effectively balances minimizing false positives and negatives [ 49 ]. $$\:F1-Measure=2\frac{\left(Re\right)\left(Pre\right)}{\left(Re\right)+\left(Pre\right)}\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\left(4\right)$$ We employ these four key evaluation metrics to ensure a comprehensive and robust assessment of the proposed model’s effectiveness in detecting fake news. 4 − 3 Model Hyperparameters The selection of optimal hyperparameters influences the performance of the proposed fake news detection model. The model combines BERT embeddings (max token length: 128, embedding dimension: 768, fine-tuning learning rate: 2×10 − 52×10 − 5) with CNN (128 filters, kernel size: 3) for word-level features and Bi-GRU (256 units) for sentence-level features. A HAN (word-level and sentence-level attention dimensions: 128) fuses these features, focusing on essential words and sentences. The classification layer uses a softmax activation, and training employs the Adam optimizer (learning rate: 1×10 − 41×10 − 4, batch size: 32) for up to 20 epochs with early stopping. This architecture is designed for fake news detection, leveraging deep semantic analysis. 5. Results and Analysis This section presents a detailed performance evaluation of BCGH-Net against existing models across four benchmark datasets. It is followed by a discussion of its performance strengths and a critical analysis of its limitations. 5.1 Performance Evaluation To validate the effectiveness of the proposed BCGH-Net model, we conducted extensive experiments using four benchmark datasets: FakeNewsNet, ISOT, CoAID, and ReCOVery. These datasets cover various misinformation challenges, including general fake news, politically biased content, and health-specific disinformation. They are well-suited for evaluating the model’s accuracy, robustness, and domain adaptability. We compared BCGH-Net with a diverse range of competing models, including traditional deep learning architectures such as CNN and LSTM; transformer-based models like RoBERTa and XLNet; recent hybrid methods such as BERT-OPCNN combined with Bi-LSTM and OPCNN-FAKE; and specialized frameworks like FakeBERT and CAF-ODNN. Additionally, we included domain-specific systems tailored for COVID-19 misinformation, such as CoAID-DEEP and the ReCOVery benchmark. This comprehensive comparison enables a rigorous assessment of BCGH-Net’s performance across different model categories and content domains. Table 3 Performance comparison between the models Model FakeNewsNet ISOT CoAID ReCOVery CNN 89.6% 91.2% 86.7% 84.5% LSTM 90.8% 92.5% 87.4% 85.9% RoBERTa 94.1% 94.8% 91.6% 90.1% XLNet 93.9% 95.0% 90.7% 89.4% OPCNN-FAKE [ 14 ] — 94.2% — — FakeBERT [ 23 ] — 98.9% — — BERT-OPCNN + Bi-LSTM [ 20 ] — 98.2% — — CAF-ODNN [ 24 ] 92.0% — — — Graph CNN [ 27 ] 92.7% — — — CoAID-DEEP [ 33 ] — — 98.57% — ReCOVery (Zhou et al.) [ 34 ] — — — 86.82% Proposed Model (BCGH-Net) 95.8% 99.3% 98.7% 92.4% Figure 4 compares FakeNewsNet, ISOT, CoAID, and ReCOVery model Accuracies. BCGH-Net tops every dataset, especially the tougher ISOT and CoAID, showing its strong, generalizable edge in fake-news detection. We include confusion matrices for each dataset to further assess BCGH-Net's classification performance. This highlights the model’s ability to accurately distinguish between fake and real news instances, as shown in Fig. 5 . The matrices reveal high true positive and true negative rates, indicating substantial precision and recall across diverse misinformation domains. This supports the model’s robustness in minimizing false alarms and missed detections. 5.2 Discussion The experimental results presented in Table 3 and Fig. 4 demonstrate the superior performance of the proposed BCGH-Net model across four diverse and challenging fake news detection datasets. Unlike traditional deep learning architectures (CNN, LSTM) and advanced transformer-based models (RoBERTa, XLNet), BCGH-Net consistently achieves the highest accuracy, including 99.3% on ISOT and 98.7% on CoAID. These gains are particularly noteworthy given the varying nature of the datasets—from politically biased news (ISOT) to health-specific misinformation (CoAID and ReCOVery). The robust generalization capability of BCGH-Net can be attributed to its hybrid design that effectively integrates contextualized language representations, enabling a deeper understanding of semantic and structural cues in misinformation. This validates BCGH-Net as a good solution that adapts to domain-specific challenges while maintaining high accuracy. 5.2 Limitation Despite its strong performance, BCGH-Net presents certain limitations that warrant further investigation. First, its computational complexity and model size are significantly higher than other models. BERT fine-tuning increases training time by approximately 1.8x, offering a performance gain of 8.1–8.4% in accuracy, potentially limiting its deployment in real-time or resource-constrained environments. Additionally, while the model generalizes well across the selected benchmark datasets, its performance has yet to be evaluated on non-English or multilingual corpora, where linguistic variations and cultural context may impact detection accuracy. Future work should explore model compression techniques and multilingual training to enhance BCGH-Net's scalability and global applicability. 6. Conclusion This study presents BCGH-Net, an advanced AI-driven framework for fake news detection, leveraging BERT embeddings, multi-level feature extraction, and hierarchical attention mechanisms. The model’s exceptional performance across multiple datasets underscores its effectiveness in identifying misinformation with high accuracy and generalizability. By addressing limitations of traditional methods, such as semantic nuance capture and contextual understanding, BCGH-Net sets a new benchmark in the field. However, challenges remain, including computational demands and the need for multilingual adaptability. Future work should explore model optimization for real-time applications and extend evaluation to non-English corpora. The findings highlight the potential of hybrid deep learning architectures in combating misinformation, contributing to the broader goal of fostering informed and resilient digital communities. This research paves the way for further innovations in automated credibility assessment, ensuring safer and more trustworthy information ecosystems. Declarations Funding: This research received no external funding. Author Contribution The authors collaboratively contributed to the development and completion of this research. Hussain Al-Kaabi designed the study, developed the BCGH-Net model, and led the manuscript writing. Fuqdan Al-Ibrahimi conducted data preprocessing and experimental validation. Ali Kadhim Jasim contributed to model optimization and performance analysis. Zainab S. Idan handled literature review, dataset curation, and manuscript editing. All authors reviewed and approved the final manuscript. References Darwish, O., Tashtoush, Y., Bashayreh, A., Alomar, A., Alkhaza’leh, S., & Darweesh, D. (2023). A survey to uncover misleading and cyberbullying on social media for public health. Cluster computing , 26 (3), 1709-1735.‏ Collins, B., Hoang, D. T., Nguyen, N. T., & Hwang, D. (2021). Trends in combating fake news on social media–a survey. Journal of Information and Telecommunication , 5 (2), 247-266.‏ Grinberg, N., Joseph, K., Friedland, L., Swire-Thompson, B., & Lazer, D. (2019). Fake news on Twitter during the 2016 US presidential election. Science , 363 (6425), 374-378.‏ Al-Kabbi, H. A., Feizi-Derakhshi, M. R., & Pashazadeh, S. (2023). Multi-type feature extraction and early fusion framework for SMS spam detection. IEEE Access .‏ Mridha, M. F., Keya, A. J., Hamid, M. A., Monowar, M. M., & Rahman, M. S. (2021). A comprehensive review on fake news detection with deep learning. IEEE access , 9 , 156151-156170.‏ Al-Tameemi, I. K. S., Feizi-Derakhshi, M. R., Pashazadeh, S., & Asadpour, M. (2024). A comprehensive review of visual–textual sentiment analysis from social media networks. Journal of Computational Social Science , 7 (3), 2767-2838.‏ Derakhshi, M. R. F., Zafarani-Moattar, E., Al-Kabi, H. A. A., & Almarashy, A. H. J. (2024). Pclf: parallel cnn-lstm fusion model for sms spam filtering. In BIO Web of Conferences (Vol. 97, p. 00136). EDP Sciences.‏ Wang, J., Huang, J. X., Tu, X., Wang, J., Huang, A. J., Laskar, M. T. R., & Bhuiyan, A. (2024). Utilizing BERT for Information Retrieval: Survey, Applications, Resources, and Challenges. ACM Computing Surveys , 56 (7), 1-33.‏ Al-Kaabi, H., Darroudi, A. D., & Jasim, A. K. (2024). Survey of SMS Spam Detection Techniques: A Taxonomy. AlKadhim Journal for Computer Science , 2 (4), 23-34.‏ Shaikh, Jasmine, and Rupali Patil. "Fake news detection using machine learning." 2020 IEEE international symposium on sustainable energy, signal processing and cyber security (iSSSC) . IEEE, 2020.‏ Kesarwani, A., Chauhan, S. S., & Nair, A. R. (2020, June). Fake news detection on social media using k-nearest neighbor classifier. In 2020 international conference on advances in computing and communication engineering (ICACCE) (pp. 1-4). IEEE.‏ Elsaeed, E., Ouda, O., Elmogy, M. M., Atwan, A., & El-Daydamony, E. (2021). Detecting fake news in social media using voting classifier. IEEE Access , 9 , 161909-161925.‏ S. A. Alameri and M. Mohd, "Comparison of Fake News Detection using Machine Learning and Deep Learning Techniques," 2021 3rd International Cyber Resilience Conference (CRC) , Langkawi Island, Malaysia, 2021, pp. 1-6, doi: 10.1109/CRC50527.2021.9392458. Saleh, H., Alharbi, A., & Alsamhi, S. H. (2021). OPCNN-FAKE: Optimized convolutional neural network for fake news detection. IEEE Access , 9 , 129471-129489.‏ Lifferth, W. Fake News . Kaggle Shu, K., Mahudeswaran, D., Wang, S., Lee, D., & Liu, H. (2018). FakeNewsNet: A Data Repository with News Content, Social Context, and Spatial-Temporal Information for Studying Fake News on Social Media . arXiv preprint arXiv:1809.01286. Salem, F., Al Feel, H., Elbassuoni, S., Jaber, M., & Farah, M. (2019). FA-KES: A Fake News Dataset around the Syrian War . In Proceedings of the International AAAI Conference on Web and Social Media (Vol. 13, pp. 573-582). Ahmed, H., Traore, I., & Saad, S. (2018). Detecting Opinion Spams and Fake News Using Text Classification . Security and Privacy , 1(1) Kaliyar, Rohit Kumar, et al. "FNDNet–a deep convolutional neural network for fake news detection." Cognitive Systems Research 61 (2020): 32-44.‏ Nithya, K., Krishnamoorthi, M., Easwaramoorthy, S. V., Dhivyaa, C. R., Yoo, S., & Cho, J. (2024). Hybrid approach of deep feature extraction using BERT–OPCNN & FIAC with customized Bi-LSTM for rumor text classification. Alexandria Engineering Journal , 90 , 65-75.‏ Wang, W. Y. (2017). "Liar, Liar Pants on Fire": A New Benchmark Dataset for Fake News Detection. Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) , 422–426. https://doi.org/10.18653/v1/P17-2067 A. H. J. Almarashy, M. -R. Feizi-Derakhshi and P. Salehpour, "Enhancing Fake News Detection by Multi-Feature Classification," in IEEE Access , vol. 11, pp. 139601-139613, 2023, doi: 10.1109/ACCESS.2023.3339621. R. K. Kaliyar, A. Goswami and P. Narang, "FakeBERT: Fake news detection in social media with a BERT-based deep learning approach", Multimedia Tools Appl. , vol. 80, no. 8, pp. 11765-11788, Mar. 2021. Luvembe, A. M., Li, W., Li, S., Liu, F., & Wu, X. (2024). CAF-ODNN: Complementary attention fusion with optimized deep neural network for multimodal fake news detection. Information Processing & Management , 61 (3), 103653.‏ Truică, C. O., Apostol, E. S., & Karras, P. (2024). DANES: Deep neural network ensemble architecture for social and textual context-aware fake news detection. Knowledge-Based Systems, 294, 111715.‏ Thaokar, C., Rout, J. K., Das, H., & Rout, M. (2024). Fake content detection on benchmark dataset using various deep learning models. International Journal of Computational Science and Engineering, 27(5), 570-581.‏ Shu, K., Mahudeswaran, D., Wang, S., Lee, D., Liu, H.: FakeNewsNet: A data repository with news content, social context, and dynamic information for fake news research. Big Data 8, 171–188 (2020). https://doi.org/10.1089/big.2020.0062 Qian, F., Gong, C., Wu, Y., Huang, W.: Neural user response generator: Fake news detection with collective user intelligence. In: Proceedings of the 27th International Joint Conference on Artificial Intelligence (IJCAI), pp. 3834–3840 (2018). https://doi.org/10.24963/ijcai.2018/533 Zhang, L., Cui, L., Fu, Y., Huang, Y., Wei, F.: MCN: Multichannel network for fake news detection. In: Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (ACL), pp. 2341–2351 (2019). https://doi.org/10.18653/v1/P19-1225 Ghenai, A., Mejova, Y.: Fake cures: User-centric modeling of health misinformation in social media. In: Proceedings of the ACM Conference on Computer-Supported Cooperative Work and Social Computing (CSCW), pp. 1–15 (2018). https://doi.org/10.1145/3274314 Wang, Y., McKee, M., Torbica, A., Stuckler, D.: Systematic literature review on the spread of health-related misinformation on social media. Social Science & Medicine 240, 112552 (2019). https://doi.org/10.1016/j.socscimed.2019.112552 Khan, S., Ullah, I., Ahmad, I., et al.: A deep learning-based framework for detecting COVID-19 related misinformation on social media. Computers, Materials & Continua 69(2), 1981–1995 (2021). https://doi.org/10.32604/cmc.2021.015775 Imran, A.S., Daudpota, S.M., Kastrati, Z., Batra, R.: Cross-cultural polarity and emotion detection using transfer learning. Information Processing & Management 58(1), 102361 (2021). https://doi.org/10.1016/j.ipm.2020.102361 Zhou, X., Mulay, A., Ferrara, E., Zafarani, R.: ReCOVery: A multimodal repository for COVID-19 news credibility research. In: Proceedings of the 29th ACM International Conference on Information & Knowledge Management (CIKM), pp. 3205–3212 (2020). https://doi.org/10.1145/3340531.3412880 Dheyaa Radhi, A., Ali Hussein Al Naffakh, H., Fuqdan, A. I., A Hakim, B., & Al-Attar, B. (2024). A comprehensive review of machine learning-based models for fake news detection. In BIO Web of Conferences (Vol. 97, p. 00123). EDP Sciences.‏ Al-Kabbi, H. A., Feizi-Derakhshi, M. R., & Pashazadeh, S. (2024). A Hierarchical Two-Level Feature Fusion Approach for SMS Spam Filtering. Intelligent Automation & Soft Computing, 39(4).‏ Alfreihat, M., Almousa, O., Tashtoush, Y., AlSobeh, A., Mansour, K., & Migdady, H. (2024). Emo-SL Framework: Emoji Sentiment Lexicon Using Text-Based Features and Machine Learning for Sentiment Analysis. IEEE Access.‏ Al-Kaabi, H., Darroudi, A. D., & Jasim, A. K. (2024). Survey of SMS Spam Detection Techniques: A Taxonomy. AlKadhim Journal for Computer Science, 2(4), 23-34.‏ Devlin, J., Chang, M. W., Lee, K., & Toutanova, K. (2019, June). Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers) (pp. 4171-4186).‏ Zhu, Q., Jiang, X., & Ye, R. (2021). Sentiment analysis of review text based on BiGRU-attention and hybrid CNN. IEEE Access , 9 , 149077-149088.‏ Kumar, A. (2022). Contextual semantics using hierarchical attention network for sentiment classification in social internet-of-things. Multimedia Tools and Applications , 81 (26), 36967-36982.‏ Yang T. (2019) Hierarchical Attention Networks. Medium . https://medium.com/analytics-vidhya/hierarchical-attention-networks-d220318cf87e Hassan, A. A. R., & Feizi-Derakhshi, M. R. (2025). ResCBAR-FusionNet: A Hybrid CNN-BiGRU-Attention Model for Human Activity Recognition. Signal, Image and Video Processing , 19 (8), 1-9. Shu, K., Mahudeswaran, D., Wang, S., Lee, D., & Liu, H. (2020). FakeNewsNet: A Data Repository with News Content, Social Context, and Dynamic Information for Studying Fake News on Social Media. Big Data, 8(3), 171–188. https://doi.org/10.1089/big.2020.0062 ISOT Research Lab. (n.d.). Fake News Detection Datasets. University of Victoria. Available at: https://onlineacademiccommunity.uvic.ca/isot/2022/11/27/fake-news-detection-datasets/ Cui, L., & Lee, D. (2020). CoAID: COVID-19 Healthcare Misinformation Dataset. arXiv preprint arXiv:2006.00885. https://arxiv.org/abs/2006.00885 Zhou, X., Mulay, A., Ferrara, E., & Zafarani, R. (2020). ReCOVery: A Multimodal Repository for COVID-19 News Credibility Research. arXiv preprint arXiv:2006.05557. https://arxiv.org/abs/2006.05557 Kamber, A. N. (2020). A variety of electrocardiogram (ECG) signal processing. Solid State Technology , 63 (3), 5370-5377.‏ Sabri, R. F., & Abdullah, N. A. (2025). Extremism Detection in the Iraqi Dialect Based on Machine Learning. Iraqi Journal of Science .‏ 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-6893733","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":472557637,"identity":"f472915b-8c1d-4944-9ba5-3e41edd7ea7b","order_by":0,"name":"Hussein Al-kaabi","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA8ElEQVRIiWNgGAWjYBACAwR58PmPD0CKjZ14LYcNJGeAtDATpQUMmA2kecA0AS3mEumPP/wouMNgzniYwdjm1zZ5PmYGxg8fc3BrsZyRYybZY/CMwbLhMENybt9twzZmBmbJmdvwOOxGDhsDj8FhBoMD5w8czu25zQjUwsbMi1dL+uOPf8BaDjM2W/bctidCSwLQ1xAtzMwMP24nEtZy5o2ZtIzBYR6gFjbG3obbyW3MjM34/XIc6LA3fw7LGdw4zMbw489t2/ntzQc/fMSjBQZ4GCQOMDAwtoHYjA2E1YMBP0jhHyIVj4JRMApGwYgCAKfMUpkvmiIcAAAAAElFTkSuQmCC","orcid":"","institution":"Ministry of Education, General Directorate of Education in Najaf, Iraq","correspondingAuthor":true,"prefix":"","firstName":"Hussein","middleName":"","lastName":"Al-kaabi","suffix":""},{"id":472557638,"identity":"260910e5-e756-4cbe-9e69-eb2ba56c9864","order_by":1,"name":"Fuqdan Al-ibrahimi","email":"","orcid":"","institution":"University of Al-Ameed‎","correspondingAuthor":false,"prefix":"","firstName":"Fuqdan","middleName":"","lastName":"Al-ibrahimi","suffix":""},{"id":472557639,"identity":"827c206e-4f75-4f6c-bc2e-c385f1208ec7","order_by":2,"name":"Ali kadhim jasim","email":"","orcid":"","institution":"Imam Jaafar Al-Sadiq University","correspondingAuthor":false,"prefix":"","firstName":"Ali","middleName":"kadhim","lastName":"jasim","suffix":""},{"id":472557640,"identity":"dbb61f39-8fdf-44b4-a4a2-705fba2baa49","order_by":3,"name":"Zainab S. Idan","email":"","orcid":"","institution":"University of Alkafeel","correspondingAuthor":false,"prefix":"","firstName":"Zainab","middleName":"S.","lastName":"Idan","suffix":""}],"badges":[],"createdAt":"2025-06-14 11:38:19","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6893733/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6893733/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":87511668,"identity":"6c90f6e7-d302-40a5-a788-66521223b681","added_by":"auto","created_at":"2025-07-24 15:42:57","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":81098,"visible":true,"origin":"","legend":"\u003cp\u003eOverall architecture of the proposed fake news detection model\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-6893733/v1/3040481d9cdcaf71484bf4f7.png"},{"id":87510605,"identity":"b9d0982f-9e53-4466-a777-31fdbed9d76c","added_by":"auto","created_at":"2025-07-24 15:34:57","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":144964,"visible":true,"origin":"","legend":"\u003cp\u003eHierarchical Attention Network (HAN) architecture adapted from [42]\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-6893733/v1/0e3f64f088e103ba645be72a.png"},{"id":87511669,"identity":"15be647c-9549-49e5-a8f1-e3b02b769152","added_by":"auto","created_at":"2025-07-24 15:42:57","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":30801,"visible":true,"origin":"","legend":"\u003cp\u003eDistribution of real and fake news articles in the FA-KES and ISOT datasets\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-6893733/v1/58d81ae28c08a7ec790fbcbd.png"},{"id":87510606,"identity":"26b2d631-eec1-4ef4-a16f-f63d99178b13","added_by":"auto","created_at":"2025-07-24 15:34:57","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":124661,"visible":true,"origin":"","legend":"\u003cp\u003eComparison of model accuracies across four benchmark datasets. The proposed BCGH-Net outperforms all baseline and state-of-the-art models in each case.\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-6893733/v1/3d89e45fb22238f52be3cb85.png"},{"id":87512039,"identity":"c5e2f56c-f1c2-47df-b692-057589471934","added_by":"auto","created_at":"2025-07-24 15:50:57","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":51456,"visible":true,"origin":"","legend":"\u003cp\u003eConfusion Matrices for Each Dataset\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-6893733/v1/fa7d5ea1d0e15c274f5470af.png"},{"id":88824436,"identity":"8029519f-4ff3-4673-89b3-12c862448b8f","added_by":"auto","created_at":"2025-08-11 18:31:34","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1205816,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6893733/v1/b12af0f9-465e-44ab-a24f-5dfb0ac0f6db.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"BCGH-Net: A Hierarchical Neural Framework for Fake News Detection Using BERT and Attention Fusion","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eThe spread of misinformation and fake news has been a significant concern for societies throughout history. While the advent of print media in the 15th century marked a considerable revolution in information dissemination, it also posed challenges regarding the accuracy and credibility of the information shared. The mass media's rise in the 20th century deepened these concerns, as the large-scale dissemination of factual and misleading information became more accessible [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. However, the emergence of digital platforms, mainly social media, in the early 21st century dramatically accelerated the spread of fake news. The ease of content creation and distribution on platforms like Facebook, Twitter, and Instagram has made it increasingly difficult to differentiate between legitimate and misleading information [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. The detection of fake news has thus become a pressing issue. If left unchecked, misinformation can erode trust in institutions, exacerbate societal polarization, and shape political outcomes [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. Traditional verification methods are no longer sufficient in the age of social media, where news spreads rapidly and can be subtly manipulated. Consequently, there is an urgent need for automated systems that can efficiently detect fake news and provide the public with accurate and reliable information [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eArtificial Intelligence (AI), specifically Natural Language Processing (NLP), has emerged as a powerful tool in the fight against fake news. With advances in machine learning algorithms and deep learning architectures, AI can now analyze vast amounts of textual data, identify patterns, and discern subtle cues that distinguish real news from fake news [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. Techniques such as sentiment analysis [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e] and spam detection [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e] have enabled AI models to evaluate news content unimaginably. This paper proposes a novel approach for fake news detection that leverages BERT (Bidirectional Encoder Representations from Transformers), a state-of-the-art contextual embedding model [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. Our method incorporates a multi-level architecture that encompasses domain-specific pre-processing, feature extraction utilizing Convolutional Neural Networks (CNN) and Bidirectional Gated Recurrent Units (Bi-GRU), and fusion via a Hierarchical Attention Network (HAN). We aim to enhance the performance of fake news detection models by incorporating social media-specific features, such as hashtags, emojis, and user mentions, and by fine-tuning BERT to improve its understanding of fake news in different contexts. Our work contributes to developing more accurate and efficient systems for detecting fake news, which can be deployed on real-world social media platforms.\u003c/p\u003e\u003cp\u003eThe key contributions of this paper are as follows:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eMulti-Level Feature Extraction: We develop a two-stage feature extraction approach that combines CNN for local word-level features and Bi-GRU for sequential sentence-level dependencies.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eHierarchical Attention Fusion: We employ a Hierarchical Attention Network (HAN) to fuse multi-granular features effectively, allowing the model to focus on the most informative words and sentences.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eSuperior Performance on Benchmark Datasets: The proposed model significantly improves fake news detection accuracy, precision, recall, and F1-score compared to state-of-the-art methods on the FakeNewsNet, ISOT, CoAID, and ReCOVery datasets.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eThe remainder of this paper is organized as follows: Section \u003cspan refid=\"Sec24\" class=\"InternalRef\"\u003e2\u003c/span\u003e reviews the related work in fake news detection, particularly the role of AI and NLP. Section \u003cspan refid=\"Sec25\" class=\"InternalRef\"\u003e3\u003c/span\u003e describes the methodology employed in this study, including the data pre-processing, feature extraction, and model architecture. Section \u003cspan refid=\"Sec21\" class=\"InternalRef\"\u003e4\u003c/span\u003e presents our experiments' experimental setup, dataset, and results. Section \u003cspan refid=\"Sec19\" class=\"InternalRef\"\u003e5\u003c/span\u003e discusses the implications of our findings, and Section 6 concludes the paper with suggestions for future work in the field.\u003c/p\u003e"},{"header":"2. Related work","content":"\u003cp\u003eThe growing prevalence of fake news has led to extensive research into automated detection methods, with machine learning (ML) and deep learning (DL) emerging as dominant approaches. ML techniques rely on handcrafted features and traditional classifiers, while DL models leverage neural networks for feature extraction and classification [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. This section reviews recent advancements in ML- and DL-based fake news detection methodologies.\u003c/p\u003e\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003e2.1 \u003cb\u003eMachine Learning-Based Approaches\u003c/b\u003e\u003c/h2\u003e\u003cp\u003eRecent studies have explored machine learning (ML) techniques for fake news detection using various feature extraction and classification methods. Traditional ML models, such as SVM, LR, and KNN, have been employed with TF-IDF, chi-square, and ANOVA for feature selection [\u003cspan additionalcitationids=\"CR11\" citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. However, ML-based approaches have limitations, including reliance on handcrafted features that may miss semantic nuances, sensitivity to training data quality, and difficulty capturing deep contextual relationships. These challenges have driven the shift toward deep learning, which offers automatic feature extraction and improved accuracy [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e].\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e\u003ch2\u003e2.2 Deep Learning-Based Approaches\u003c/h2\u003e\u003cp\u003eDeep learning (DL) techniques have significantly enhanced text classification tasks by integrating CNN, RNN, and transformer-based models. Saleh et al. [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e] proposed an Optimized Convolutional Neural Network (OPCNN-FAKE) model to enhance fake news detection through deep learning. The study evaluated its performance against traditional machine learning models (Decision Tree, Logistic Regression, K-Nearest Neighbor, Random Forest, Support Vector Machine, and Na\u0026iuml;ve Bayes) as well as deep learning models (Recurrent Neural Network and Long Short-Term Memory) across four benchmark datasets: Kaggle Fake News Dataset [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e], FakeNewsNet [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e], FA-KES [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e], and ISOT [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e], demonstrating its effectiveness in improving detection accuracy. Kaliyar et al. [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e] proposed FNDNet, a deep Convolutional Neural Network (CNN) model explicitly designed for fake news detection. Unlike traditional approaches that rely on hand-crafted features, FNDNet can automatically learn discriminatory features for fake news classification through its multiple hidden layers by leveraging CNN\u0026rsquo;s hierarchical feature extraction capabilities.\u003c/p\u003e\u003cp\u003eNithya et al. [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e] proposed a two-phase feature extraction approach, combining BERT-OPCNN for deep feature extraction and fastText with ant colony optimization (FIAC) for word embedding. A customized Bi-LSTM classifies the resulting vectors, achieving 55.69% accuracy on the LIAR [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e] dataset and 98.24% on the ISOT dataset, demonstrating its effectiveness in fake news detection. Almarashy et al. [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e] suggest a new model to detect fake news by extracting and combining global features (using TF-IDF), spatial features (using CNN), and temporal features (using BiLSTM). These features are then classified using a Fast Learning Network (FLN). The model was tested on two publicly available fake news datasets: ISOT and FA-KES. Kaliyar et al. [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e] proposed FakeBERT, a deep learning model to tackle ambiguity in natural language processing. By integrating BERT with advanced deep learning methods, FakeBERT achieves an impressive accuracy of 98.90% for the fake news dataset, surpassing the performance of existing models.\u003c/p\u003e\u003cp\u003eLuvembee et al. [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e] proposed Complementary Attention Fusion with an Optimized Deep Neural Network (CAF-ODNN), which captures subtle cross-modal relationships using image captioning and bidirectional attention. CAF-ODNN improves feature extraction and fusion, outperforming existing methods on real-world datasets containing tweets and fact-checked news from PolitiFact and GossipCop websites. Truică et al. [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e] proposed a Deep Neural Network Ensemble Architecture for social and textual context-aware fake news detection (DANES). DANES integrates a Text Branch for textual content and a Social Branch for social context, creating a novel Network Embedding. Preliminary ablation studies on three real-world datasets, such as Twitter15 and Twitter16, demonstrate its effectiveness. Thaokar et al. [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e] proposed an approach comprising basic natural language processing techniques with different deep learning models for categorizing content as real or fake. LSTM, Bi-LSTM, and Bi-LSTM with an attention mechanism are employed. The study compared results using one-hot word embedding and a pre-trained GloVe technique. On the benchmark LIAR dataset, the LSTM achieved an accuracy of 67.2%, while Bi-LSTM with GloVe embedding reached 67%. An accuracy of 98.22% was achieved using Bi-LSTM, and 97.98% using LSTM on the Real-Fake dataset. Shu et al. [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e] introduced FakeNewsNet, a comprehensive data repository integrating news content, social context, and dynamic information, achieving approximately 65% accuracy with baseline models. In a comparative analysis, Qian et al. [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e] demonstrated that Graph Neural Networks and Transformer-based architectures can achieve up to 86.16% accuracy on the same dataset. Zhang et al. [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e] proposed MCNNet, a multichannel neural network model that surpassed 90% accuracy by effectively integrating multiple modalities. Similarly, Ghenai et al. [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e] improved fake news detection by converting text into image representations, achieving 99.3% accuracy on the ISOT dataset. While highly effective, this method requires significant computational resources, limiting its practicality in low-resource or real-time settings. Addressing COVID-19 misinformation, Wang et al. [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e] developed a machine-learned solution utilizing the CoAID dataset, employing a Bi-LSTM model trained on 1,257 entries and tested on 419 entries. This approach achieved a weighted average F1-score of 0.93, with equal precision, recall, and accuracy across both labels. Khan et al. [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e] conducted a focused case study on COVID-19 misinformation detection, achieving an accuracy of 93%. Their study emphasized the effectiveness of machine learning techniques in identifying false information related to the pandemic. Imran et al. [\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e] introduced the CoAID-DEEP framework, an optimized intelligent system for automated detection of COVID-19 misleading information on Twitter. This framework achieved an accuracy of 98.57%, demonstrating its efficacy in combating misinformation on social media platforms. Zhou et al. [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e] addressed the challenge of fake news detection under limited information settings using the ReCOVery dataset. They developed a multimodal repository for COVID-19 news credibility research, achieving an accuracy of 86.82%. Their work highlights the importance of integrating various data modalities to enhance detection performance.\u003c/p\u003e\u003c/div\u003e"},{"header":"3. Proposed Model","content":"\u003cp\u003eThis section presents the proposed model for fake news detection, which consists of five key steps: pre-processing, embedding generation using BERT, two-level feature extraction, feature fusion using a Hierarchical Attention Network (HAN), and classification, as shown in Figure 1.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.1. Pre-Processing Step\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003ePre-processing ensures the input data is clean and compatible with BERT\u0026apos;s embedding process. The pre-processing steps are:\u0026nbsp;\u003c/p\u003e\n\u003col\u003e\n \u003cli\u003e\u003cstrong\u003eBasic Cleaning\u003c/strong\u003e:\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003eRemove excessive noise such as repetitive hashtags, irrelevant symbols, or unnecessary special characters [35]. Eliminate URLs, email addresses, or other non-textual elements that do not contribute to the context [36]. \u0026nbsp;Handle emojis by converting them to textual representations (e.g., \u0026quot;😂\u0026quot; to \u0026quot;laughing\u0026quot;) if they carry meaning [37].\u0026nbsp;\u003c/p\u003e\n\u003col start=\"2\"\u003e\n \u003cli\u003e\u003cstrong\u003ePreserve Casing\u003c/strong\u003e:\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003eRetain the original casing of the text, as BERT is case-sensitive and uses it for contextual understanding (e.g., \u0026quot;Apple\u0026quot; vs. \u0026quot;apple\u0026quot;) [38]. Focusing on these cleaning steps ensures that the input data remains compatible with BERT\u0026rsquo;s tokenizer and embedding process, providing high-quality, context-aware representations.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.2. BERT for Embedding Step\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe pre-processed text is passed through the BERT (Bidirectional Encoder Representations from Transformers) model to generate contextual embeddings. BERT\u0026apos;s architecture, which leverages a transformer-based bidirectional encoder, captures the semantic relationships between words within a sentence. Unlike traditional word embeddings that generate static representations, BERT produces dynamic, context-aware embeddings that vary depending on the surrounding words [39]. This bidirectional context allows BERT to understand nuanced meanings, such as polysemy (words with multiple meanings), which is crucial for distinguishing subtle differences in fake news content. The embeddings generated by BERT serve as a robust and rich representation of the textual data, providing a solid foundation for subsequent feature extraction. By fine-tuning BERT on our specific fake news dataset, we further enhance its ability to capture domain-specific linguistic patterns by improving the model\u0026apos;s overall performance in detecting fake news. This step ensures the model is equipped with high-quality, contextually rich embedding essential for accurate classification. \u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.3. Two-Level Feature Extraction\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFeature extraction is performed in two stages to capture word-level and sentence-level features, ensuring a comprehensive understanding of textual content. This hierarchical approach allows the model to analyze local patterns (e.g., specific word combinations) and global structures (e.g., sentence coherence and context). \u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.3.1. Word-level feature extraction\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eA\u0026nbsp;CNN\u0026nbsp;is employed to extract local and spatial features from the word embeddings generated by BERT. The CNN applies multiple filters to detect n-gram patterns, such as word combinations or phrases, which are critical for understanding the text context [7]. For example, in the sentence\u0026nbsp;\u0026quot;The government announced a new policy to reduce taxes,\u0026quot;\u0026nbsp;CNN might identify key n-grams like\u0026nbsp;\u0026quot;\u003cem\u003ereduce taxes\u003c/em\u003e\u0026quot;\u0026nbsp;or\u0026nbsp;\u0026quot;\u003cem\u003enew policy\u003c/em\u003e,\u0026quot;\u0026nbsp;which could be indicative of the sentence\u0026apos;s intent or tone. Focusing on these local features, CNN captures subtle linguistic cues often present in fake news, such as exaggerated language or misleading phrases.\u003cstrong\u003e\u003cbr\u003e\u0026nbsp;3.3.2. Sentence-Level Feature Extraction:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eA BiGRU extracts sequential and contextual information at the sentence level. Unlike unidirectional models, the BiGRU processes the text in both forward and backward directions, capturing dependencies between past and future words [40]. This bidirectional approach ensures a more comprehensive understanding of the sentence\u0026apos;s structure and meaning. For instance, in the sentence \u003cem\u003e\u0026quot;Despite the economic growth, unemployment rates remain high,\u0026quot;\u003c/em\u003e the BiGRU can capture the contrast between \u003cem\u003e\u0026quot;economic growth\u0026quot;\u003c/em\u003e and \u003cem\u003e\u0026quot;unemployment rates,\u0026quot;\u003c/em\u003e which is crucial for understanding the overall sentiment and context of the sentence. Analyzing the entire sentence, the BiGRU helps the model identify inconsistencies or illogical sequences common in fake news. For Example, consider the following two sentences:\u003c/p\u003e\n\u003col start=\"1\" type=\"1\"\u003e\n \u003cli\u003e\u003cem\u003e\u0026quot;The new vaccine is 100% effective with no side effects.\u0026quot;\u003c/em\u003e\u003c/li\u003e\n \u003cli\u003e\u003cem\u003e\u0026quot;Experts warn that the new vaccine may have rare side effects.\u0026quot;\u003c/em\u003e\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003eAt the word level, CNN might detect n-grams like \u003cem\u003e\u0026quot;100% effective\u0026quot;\u003c/em\u003e in the first sentence, which could be a red flag for exaggeration, and \u003cem\u003e\u0026quot;rare side effects\u0026quot;\u003c/em\u003e in the second sentence, which reflects a more cautious tone. At the sentence level, the BiGRU would analyze each sentence\u0026apos;s overall structure and context. The first sentence, with its absolute claim (\u003cem\u003e\u0026quot;100% effective\u0026quot;\u003c/em\u003e), might be flagged as suspicious, while the second sentence, which includes a qualifier (\u003cem\u003e\u0026quot;may have\u0026quot;\u003c/em\u003e), would be considered more credible. By combining\u0026nbsp;word-level\u0026nbsp;and\u0026nbsp;sentence-level\u0026nbsp;feature extraction, the model gains a multidimensional understanding of the text, enabling it to detect both local anomalies and global inconsistencies characteristic of fake news. This two-level approach ensures that the model is sensitive to specific word patterns and capable of understanding the broader context in which those words appear.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.4. HAN for Feature Fusion\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eA\u0026nbsp;HAN combines the extracted word-level and sentence-level features. The HAN introduces a dual-level attention mechanism that prioritizes the most relevant features at both the word and sentence levels, as shown in Figure 3. This ensures that the model focuses on the key aspects of the text that contribute to the classification task. This hierarchical approach enables the model to dynamically weigh the importance of different words within a sentence and across different sentences within a document, thereby enhancing its ability to detect subtle patterns in fake news [41]. Figure 2 explains the HAN architecture.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.4.1 Word-Level Attention:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe attention mechanism assigns weights to each word at the word level based on its contribution to the sentence\u0026apos;s meaning. Words that are more relevant to the context or have stronger semantic meanings are given higher weights. For example, in the sentence \u0026quot;The politician claimed that the election was rigged,\u0026quot; the attention mechanism might assign higher weights to words like \u0026quot;rigged\u0026quot; and \u0026quot;claimed,\u0026quot; as they are critical to understanding the sentence\u0026apos;s intent and potential for misinformation. This ensures the model focuses on the most informative words while downplaying less relevant ones, such as stop or filler.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.4.2 Sentence-Level Attention\u003c/strong\u003e:\u003c/p\u003e\n\u003cp\u003eAt the sentence level, the attention mechanism evaluates the importance of each sentence within the broader context of the document. Sentences that contain key claims, evidence, or contradictions are given higher weights. For instance, in a news article, sentences like \u003cem\u003e\u0026quot;Experts have debunked the claim that 5G causes COVID-19\u0026quot;\u003c/em\u003e would receive more attention than less informative sentences like \u003cem\u003e\u0026quot;The article was published last week.\u0026quot;\u003c/em\u003e This allows the model to prioritize sentences more likely to contain critical information to detect fake news.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.4.3 Feature Fusion\u003c/strong\u003e:\u003c/p\u003e\n\u003cp\u003eThe weighted word-level and sentence-level features are then fused into a unified representation. This fused representation captures both the local nuances (from word-level features) and the global context (from sentence-level features), providing a comprehensive understanding of the text. The fused features are passed to the classification layer, where the model makes its final prediction. Consider a fake news article with the following sentences:\u003c/p\u003e\n\u003col start=\"1\" type=\"1\"\u003e\n \u003cli\u003e\u003cem\u003e\u0026quot;A new study proves that drinking bleach cures COVID-19.\u0026quot;\u003c/em\u003e\u003c/li\u003e\n \u003cli\u003e\u003cem\u003e\u0026quot;The study was conducted by an unknown group with no peer review.\u0026quot;\u003c/em\u003e\u003c/li\u003e\n \u003cli\u003e\u003cem\u003e\u0026quot;Health officials have warned against this dangerous practice.\u0026quot;\u003c/em\u003e\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp;At the word level, the attention mechanism might highlight words like \u003cem\u003e\u0026quot;bleach,\u0026quot;\u003c/em\u003e \u003cem\u003e\u0026quot;cures,\u0026quot;\u003c/em\u003e and \u003cem\u003e\u0026quot;dangerous\u0026quot;\u003c/em\u003e as highly relevant due to their strong semantic meaning and potential for misinformation. At the sentence level, the attention mechanism would likely assign higher weights to the first and third sentences, as they contain strong claims and warnings, respectively, while downplaying the second sentence provides context but is less critical for classification. Using the HAN for feature fusion, the model effectively combines the strengths of both word-level and sentence-level analysis, ensuring that it captures the most relevant information for accurate fake news detection. This hierarchical attention mechanism improves the model\u0026apos;s performance. It enhances its interpretability, as the attention weights provide insights into which words and sentences the model considers most important for its decision-making process.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e5. Classification Step\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eIn the final step, the fused features are fed into a fully connected dense layer, followed by a softmax activation function, to perform classification. The output layer assigns probabilities to the predefined classes (fake or real news), enabling the model to make accurate predictions. The softmax function ensures that the predicted probabilities sum up to one, allowing for a clear interpretation of the classification results [43]. \u0026nbsp;By integrating these steps, the proposed model effectively combines the strengths of deep contextual embeddings, hierarchical feature extraction, and attention mechanisms, resulting in a robust framework for fake news detection.\u0026nbsp;\u003c/p\u003e\n\u003cdiv align=\"\"\u003e\n \u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" class=\"fr-table-selection-hover\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eAlgorithm: Fake News Detection using two-level feature fusion\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eInput\u003c/strong\u003e: Labeled news statements\u003cbr\u003e\u003cstrong\u003eOutput\u003c/strong\u003e: Predicted label: \u003cem\u003eFake\u003c/em\u003e or \u003cem\u003eReal\u003c/em\u003e\u003c/p\u003e\n \u003col start=\"1\" type=\"1\"\u003e\n \u003cli\u003e\u003cstrong\u003eText Preprocessing\u003c/strong\u003e: Clean and tokenize input using standard NLP techniques to normalize text for modeling.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eBERT Embeddings\u003c/strong\u003e: Generate contextualized token representations using a pre-trained BERT model.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eFeature Extraction\u003c/strong\u003e:\u003cul type=\"circle\"\u003e\n \u003cli\u003e\u003cstrong\u003eCNN\u003c/strong\u003e captures local n-gram patterns.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eBi-GRU\u003c/strong\u003e models sequential dependencies across tokens.\u003c/li\u003e\n \u003c/ul\u003e\n \u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eFeature Fusion\u003c/strong\u003e: Apply Hierarchical Attention Network (HAN) to integrate CNN and Bi-GRU outputs, emphasizing semantically salient components.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eClassification\u003c/strong\u003e: Pass fused features through a dense layer, followed by a softmax function to predict class probabilities.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eTraining and Evaluation\u003c/strong\u003e: Optimize with cross-entropy loss using Adam. Evaluate using accuracy, precision, recall, and F1-score.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eInference\u003c/strong\u003e: Apply the trained model to predict labels on unseen news statements.\u003c/li\u003e\n \u003c/ol\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003e\u003cspan dir=\"RTL\"\u003e\u0026nbsp;\u003c/span\u003e\u003c/p\u003e"},{"header":"4. Evaluation","content":"\u003cp\u003eSpecific datasets are recognized as more challenging when evaluating fake news detection models. They may yield lower accuracy due to their complexity, noise, or the subtle nature of the fake news they contain.\u003c/p\u003e\n\u003ch3\u003e4 − 1 datasets\u003c/h3\u003e\n\u003cp\u003eTo evaluate the performance and robustness of the proposed fake news detection model, we employed four publicly available English-language datasets, each annotated with binary labels (i.e., fake or real). These datasets were chosen to ensure diversity in topical focus, textual style, and data modality. FakeNewsNet, sourced from \u003cem\u003ePolitiFact\u003c/em\u003e and \u003cem\u003eGossipCop\u003c/em\u003e, includes content and social context, supporting models incorporating user interaction signals [\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e]. The ISOT Fake News Dataset provides over 44,000 articles from reliable and unreliable sources, offering balanced coverage of political and general news [\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e]. The CoAID dataset contributes COVID-19-specific misinformation data with user engagement metadata, while ReCOVery offers credibility-labeled COVID-19 news annotated by fact-checking services like NewsGuard [\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e, \u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e]. These datasets vary widely in size and thematic scope from general political discourse to health misinformation, enabling a thorough assessment of the model\u0026rsquo;s generalizability and real-world applicability. Figure\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e shows the dataset class distribution.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\n\u003ch3\u003e4 − 2 Evaluation metrics\u003c/h3\u003e\n\u003cp\u003eWe employ four fundamental evaluation metrics to assess the proposed fake news detection model's performance: Accuracy, Precision, Recall, and F1-Measure. These metrics provide a comprehensive assessment of classification effectiveness.\u003c/p\u003e\n\u003ch3\u003e1. Accuracy\u003c/h3\u003e\n\u003cp\u003eAccuracy is a fundamental metric that measures the overall correctness of the model\u0026rsquo;s predictions. It is defined as:\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e\n$$\\:Accuracy\\:\\left(Acc\\right)=\\frac{TP+TN}{TP+TN+FP+FN}100\\%\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\left(1\\right)$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eWhere:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eTP (True Positives): Fake news correctly classified as fake.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eTN (True Negatives): Real news correctly classified as real.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eFP (False Positives): Real news incorrectly classified as fake.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eFN (False Negatives): Fake news incorrectly classified as real [\u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e].\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eWhile accuracy provides an intuitive measure of performance, it can be misleading in cases of imbalanced class distribution. Therefore, additional metrics are required for a more nuanced evaluation.\u003c/p\u003e\n\u003ch3\u003e2. Precision\u003c/h3\u003e\n\u003cp\u003ePrecision, also known as Positive Predictive Value (PPV), measures the model's reliability in predicting fake news. It is defined as:\u003cdiv id=\"Equb\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equb\" name=\"EquationSource\"\u003e\n$$\\:Precesion\\:\\left(Pre\\right)=\\frac{TN}{TN+FP}100\\%\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\left(2\\right)$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eA high precision value indicates that the model is likely correct when classifying a news statement as fake. This metric is critical in minimizing false alarms and ensuring that legitimate news is not incorrectly flagged as counterfeit.\u003c/p\u003e\n\u003ch3\u003e3. Recall\u003c/h3\u003e\n\u003cp\u003eRecall, called Sensitivity or True Positive Rate (TPR), quantifies the model\u0026rsquo;s ability to identify fake news correctly. It is given by:\u003cdiv id=\"Equc\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equc\" name=\"EquationSource\"\u003e\n$$\\:Recall\\:\\left(Re\\right)=\\frac{TP}{TP+FN}100\\%\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\left(3\\right)$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eA high recall value indicates that the model successfully detects a large proportion of actual fake news articles. This metric is particularly critical in applications where missing fake news (false negatives) can lead to misinformation spreading unchecked.\u003c/p\u003e\u003cp\u003e\u003cb\u003e4. F1-Measure\u003c/b\u003e: Also called the F1-score, it is the harmonic mean of Precision and Recall, balancing both metrics to provide a single performance measure. F1-Measure is particularly useful in scenarios with class imbalance, where either precision or recall alone might give a misleading picture of performance. A high F1-Measure indicates that the model effectively balances minimizing false positives and negatives [\u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e].\u003cdiv id=\"Equd\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equd\" name=\"EquationSource\"\u003e\n$$\\:F1-Measure=2\\frac{\\left(Re\\right)\\left(Pre\\right)}{\\left(Re\\right)+\\left(Pre\\right)}\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\left(4\\right)$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eWe employ these four key evaluation metrics to ensure a comprehensive and robust assessment of the proposed model\u0026rsquo;s effectiveness in detecting fake news.\u003c/p\u003e\n\u003ch3\u003e4 − 3 Model Hyperparameters\u003c/h3\u003e\n\u003cp\u003eThe selection of optimal hyperparameters influences the performance of the proposed fake news detection model. The model combines BERT embeddings (max token length: 128, embedding dimension: 768, fine-tuning learning rate: 2\u0026times;10\u0026thinsp;\u0026minus;\u0026thinsp;52\u0026times;10\u0026thinsp;\u0026minus;\u0026thinsp;5) with CNN (128 filters, kernel size: 3) for word-level features and Bi-GRU (256 units) for sentence-level features. A HAN (word-level and sentence-level attention dimensions: 128) fuses these features, focusing on essential words and sentences. The classification layer uses a softmax activation, and training employs the Adam optimizer (learning rate: 1\u0026times;10\u0026thinsp;\u0026minus;\u0026thinsp;41\u0026times;10\u0026thinsp;\u0026minus;\u0026thinsp;4, batch size: 32) for up to 20 epochs with early stopping. This architecture is designed for fake news detection, leveraging deep semantic analysis.\u003c/p\u003e"},{"header":"5. Results and Analysis","content":"\u003cp\u003eThis section presents a detailed performance evaluation of BCGH-Net against existing models across four benchmark datasets. It is followed by a discussion of its performance strengths and a critical analysis of its limitations.\u003c/p\u003e\u003cdiv id=\"Sec28\" class=\"Section2\"\u003e\u003ch2\u003e5.1 Performance Evaluation\u003c/h2\u003e\u003cp\u003eTo validate the effectiveness of the proposed BCGH-Net model, we conducted extensive experiments using four benchmark datasets: FakeNewsNet, ISOT, CoAID, and ReCOVery. These datasets cover various misinformation challenges, including general fake news, politically biased content, and health-specific disinformation. They are well-suited for evaluating the model\u0026rsquo;s accuracy, robustness, and domain adaptability. We compared BCGH-Net with a diverse range of competing models, including traditional deep learning architectures such as CNN and LSTM; transformer-based models like RoBERTa and XLNet; recent hybrid methods such as BERT-OPCNN combined with Bi-LSTM and OPCNN-FAKE; and specialized frameworks like FakeBERT and CAF-ODNN. Additionally, we included domain-specific systems tailored for COVID-19 misinformation, such as CoAID-DEEP and the ReCOVery benchmark. This comprehensive comparison enables a rigorous assessment of BCGH-Net\u0026rsquo;s performance across different model categories and content domains.\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 3\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003ePerformance comparison between the models\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\u003eModel\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eFakeNewsNet\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eISOT\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eCoAID\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eReCOVery\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eCNN\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e89.6%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e91.2%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e86.7%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e84.5%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLSTM\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e90.8%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e92.5%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e87.4%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e85.9%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRoBERTa\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e94.1%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e94.8%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e91.6%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e90.1%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eXLNet\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e93.9%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e95.0%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e90.7%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e89.4%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eOPCNN-FAKE [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u0026mdash;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e94.2%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e\u0026mdash;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e\u0026mdash;\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eFakeBERT [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u0026mdash;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e98.9%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e\u0026mdash;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e\u0026mdash;\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eBERT-OPCNN\u0026thinsp;+\u0026thinsp;Bi-LSTM [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u0026mdash;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e98.2%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e\u0026mdash;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e\u0026mdash;\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eCAF-ODNN [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e92.0%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u0026mdash;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e\u0026mdash;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e\u0026mdash;\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eGraph CNN [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e92.7%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u0026mdash;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e\u0026mdash;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e\u0026mdash;\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eCoAID-DEEP [\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u0026mdash;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u0026mdash;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e98.57%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e\u0026mdash;\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eReCOVery (Zhou et al.) [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u0026mdash;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u0026mdash;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e\u0026mdash;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e86.82%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eProposed Model (BCGH-Net)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u003cb\u003e95.8%\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cb\u003e99.3%\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e\u003cb\u003e98.7%\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e\u003cb\u003e92.4%\u003c/b\u003e\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\u003eFigure \u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e compares FakeNewsNet, ISOT, CoAID, and ReCOVery model Accuracies. BCGH-Net tops every dataset, especially the tougher ISOT and CoAID, showing its strong, generalizable edge in fake-news detection.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eWe include confusion matrices for each dataset to further assess BCGH-Net's classification performance. This highlights the model\u0026rsquo;s ability to accurately distinguish between fake and real news instances, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e. The matrices reveal high true positive and true negative rates, indicating substantial precision and recall across diverse misinformation domains. This supports the model\u0026rsquo;s robustness in minimizing false alarms and missed detections.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec29\" class=\"Section2\"\u003e\u003ch2\u003e5.2 Discussion\u003c/h2\u003e\u003cp\u003eThe experimental results presented in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e3\u003c/span\u003e and Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e demonstrate the superior performance of the proposed BCGH-Net model across four diverse and challenging fake news detection datasets. Unlike traditional deep learning architectures (CNN, LSTM) and advanced transformer-based models (RoBERTa, XLNet), BCGH-Net consistently achieves the highest accuracy, including 99.3% on ISOT and 98.7% on CoAID. These gains are particularly noteworthy given the varying nature of the datasets\u0026mdash;from politically biased news (ISOT) to health-specific misinformation (CoAID and ReCOVery). The robust generalization capability of BCGH-Net can be attributed to its hybrid design that effectively integrates contextualized language representations, enabling a deeper understanding of semantic and structural cues in misinformation. This validates BCGH-Net as a good solution that adapts to domain-specific challenges while maintaining high accuracy.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec30\" class=\"Section2\"\u003e\u003ch2\u003e5.2 Limitation\u003c/h2\u003e\u003cp\u003eDespite its strong performance, BCGH-Net presents certain limitations that warrant further investigation. First, its computational complexity and model size are significantly higher than other models. BERT fine-tuning increases training time by approximately 1.8x, offering a performance gain of 8.1\u0026ndash;8.4% in accuracy, potentially limiting its deployment in real-time or resource-constrained environments. Additionally, while the model generalizes well across the selected benchmark datasets, its performance has yet to be evaluated on non-English or multilingual corpora, where linguistic variations and cultural context may impact detection accuracy. Future work should explore model compression techniques and multilingual training to enhance BCGH-Net's scalability and global applicability.\u003c/p\u003e\u003c/div\u003e"},{"header":"6. Conclusion","content":"\u003cp\u003eThis study presents BCGH-Net, an advanced AI-driven framework for fake news detection, leveraging BERT embeddings, multi-level feature extraction, and hierarchical attention mechanisms. The model\u0026rsquo;s exceptional performance across multiple datasets underscores its effectiveness in identifying misinformation with high accuracy and generalizability. By addressing limitations of traditional methods, such as semantic nuance capture and contextual understanding, BCGH-Net sets a new benchmark in the field. However, challenges remain, including computational demands and the need for multilingual adaptability. Future work should explore model optimization for real-time applications and extend evaluation to non-English corpora. The findings highlight the potential of hybrid deep learning architectures in combating misinformation, contributing to the broader goal of fostering informed and resilient digital communities. This research paves the way for further innovations in automated credibility assessment, ensuring safer and more trustworthy information ecosystems.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eFunding:\u003c/h2\u003e\u003cp\u003eThis research received no external funding.\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eThe authors collaboratively contributed to the development and completion of this research. Hussain Al-Kaabi designed the study, developed the BCGH-Net model, and led the manuscript writing. Fuqdan Al-Ibrahimi conducted data preprocessing and experimental validation. Ali Kadhim Jasim contributed to model optimization and performance analysis. Zainab S. Idan handled literature review, dataset curation, and manuscript editing. All authors reviewed and approved the final manuscript.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eDarwish, O., Tashtoush, Y., Bashayreh, A., Alomar, A., Alkhaza\u0026rsquo;leh, S., \u0026amp; Darweesh, D. (2023). A survey to uncover misleading and cyberbullying on social media for public health. \u003cem\u003eCluster computing\u003c/em\u003e, \u003cem\u003e26\u003c/em\u003e(3), 1709-1735.\u0026rlm;\u003c/li\u003e\n \u003cli\u003eCollins, B., Hoang, D. T., Nguyen, N. T., \u0026amp; Hwang, D. (2021). Trends in combating fake news on social media\u0026ndash;a survey. \u003cem\u003eJournal of Information and Telecommunication\u003c/em\u003e, \u003cem\u003e5\u003c/em\u003e(2), 247-266.\u0026rlm;\u003c/li\u003e\n \u003cli\u003eGrinberg, N., Joseph, K., Friedland, L., Swire-Thompson, B., \u0026amp; Lazer, D. (2019). Fake news on Twitter during the 2016 US presidential election. \u003cem\u003eScience\u003c/em\u003e, \u003cem\u003e363\u003c/em\u003e(6425), 374-378.\u0026rlm;\u003c/li\u003e\n \u003cli\u003eAl-Kabbi, H. A., Feizi-Derakhshi, M. R., \u0026amp; Pashazadeh, S. (2023). Multi-type feature extraction and early fusion framework for SMS spam detection. \u003cem\u003eIEEE Access\u003c/em\u003e.\u0026rlm;\u003c/li\u003e\n \u003cli\u003eMridha, M. F., Keya, A. J., Hamid, M. A., Monowar, M. M., \u0026amp; Rahman, M. S. (2021). A comprehensive review on fake news detection with deep learning. \u003cem\u003eIEEE access\u003c/em\u003e, \u003cem\u003e9\u003c/em\u003e, 156151-156170.\u0026rlm;\u003c/li\u003e\n \u003cli\u003eAl-Tameemi, I. K. S., Feizi-Derakhshi, M. R., Pashazadeh, S., \u0026amp; Asadpour, M. (2024). A comprehensive review of visual\u0026ndash;textual sentiment analysis from social media networks. \u003cem\u003eJournal of Computational Social Science\u003c/em\u003e, \u003cem\u003e7\u003c/em\u003e(3), 2767-2838.\u0026rlm;\u003c/li\u003e\n \u003cli\u003eDerakhshi, M. R. F., Zafarani-Moattar, E., Al-Kabi, H. A. A., \u0026amp; Almarashy, A. H. J. (2024). Pclf: parallel cnn-lstm fusion model for sms spam filtering. In \u003cem\u003eBIO Web of Conferences\u003c/em\u003e (Vol. 97, p. 00136). EDP Sciences.\u0026rlm;\u003c/li\u003e\n \u003cli\u003eWang, J., Huang, J. X., Tu, X., Wang, J., Huang, A. J., Laskar, M. T. R., \u0026amp; Bhuiyan, A. (2024). Utilizing BERT for Information Retrieval: Survey, Applications, Resources, and Challenges. \u003cem\u003eACM Computing Surveys\u003c/em\u003e, \u003cem\u003e56\u003c/em\u003e(7), 1-33.\u0026rlm;\u003c/li\u003e\n \u003cli\u003eAl-Kaabi, H., Darroudi, A. D., \u0026amp; Jasim, A. K. (2024). Survey of SMS Spam Detection Techniques: A Taxonomy. \u003cem\u003eAlKadhim Journal for Computer Science\u003c/em\u003e, \u003cem\u003e2\u003c/em\u003e(4), 23-34.\u0026rlm;\u003c/li\u003e\n \u003cli\u003eShaikh, Jasmine, and Rupali Patil. \u0026quot;Fake news detection using machine learning.\u0026quot; \u003cem\u003e2020 IEEE international symposium on sustainable energy, signal processing and cyber security (iSSSC)\u003c/em\u003e. IEEE, 2020.\u0026rlm;\u003c/li\u003e\n \u003cli\u003eKesarwani, A., Chauhan, S. S., \u0026amp; Nair, A. R. (2020, June). Fake news detection on social media using k-nearest neighbor classifier. In \u003cem\u003e2020 international conference on advances in computing and communication engineering (ICACCE)\u003c/em\u003e (pp. 1-4). IEEE.\u0026rlm;\u003c/li\u003e\n \u003cli\u003eElsaeed, E., Ouda, O., Elmogy, M. M., Atwan, A., \u0026amp; El-Daydamony, E. (2021). Detecting fake news in social media using voting classifier. \u003cem\u003eIEEE Access\u003c/em\u003e, \u003cem\u003e9\u003c/em\u003e, 161909-161925.\u0026rlm;\u003c/li\u003e\n \u003cli\u003eS. A. Alameri and M. Mohd, \u0026quot;Comparison of Fake News Detection using Machine Learning and Deep Learning Techniques,\u0026quot; \u003cem\u003e2021 3rd International Cyber Resilience Conference (CRC)\u003c/em\u003e, Langkawi Island, Malaysia, 2021, pp. 1-6, doi: 10.1109/CRC50527.2021.9392458.\u003c/li\u003e\n \u003cli\u003eSaleh, H., Alharbi, A., \u0026amp; Alsamhi, S. H. (2021). OPCNN-FAKE: Optimized convolutional neural network for fake news detection. \u003cem\u003eIEEE Access\u003c/em\u003e, \u003cem\u003e9\u003c/em\u003e, 129471-129489.\u0026rlm;\u003c/li\u003e\n \u003cli\u003eLifferth, W. \u003cem\u003eFake News\u003c/em\u003e. Kaggle\u003c/li\u003e\n \u003cli\u003eShu, K., Mahudeswaran, D., Wang, S., Lee, D., \u0026amp; Liu, H. (2018). \u003cem\u003eFakeNewsNet: A Data Repository with News Content, Social Context, and Spatial-Temporal Information for Studying Fake News on Social Media\u003c/em\u003e. arXiv preprint arXiv:1809.01286.\u003c/li\u003e\n \u003cli\u003eSalem, F., Al Feel, H., Elbassuoni, S., Jaber, M., \u0026amp; Farah, M. (2019). \u003cem\u003eFA-KES: A Fake News Dataset around the Syrian War\u003c/em\u003e. In \u003cem\u003eProceedings of the International AAAI Conference on Web and Social Media\u003c/em\u003e (Vol. 13, pp. 573-582).\u003c/li\u003e\n \u003cli\u003eAhmed, H., Traore, I., \u0026amp; Saad, S. (2018). \u003cem\u003eDetecting Opinion Spams and Fake News Using Text Classification\u003c/em\u003e. \u003cem\u003eSecurity and Privacy\u003c/em\u003e, 1(1)\u003c/li\u003e\n \u003cli\u003eKaliyar, Rohit Kumar, et al. \u0026quot;FNDNet\u0026ndash;a deep convolutional neural network for fake news detection.\u0026quot; \u003cem\u003eCognitive Systems Research\u003c/em\u003e 61 (2020): 32-44.\u0026rlm;\u003c/li\u003e\n \u003cli\u003eNithya, K., Krishnamoorthi, M., Easwaramoorthy, S. V., Dhivyaa, C. R., Yoo, S., \u0026amp; Cho, J. (2024). Hybrid approach of deep feature extraction using BERT\u0026ndash;OPCNN \u0026amp; FIAC with customized Bi-LSTM for rumor text classification. \u003cem\u003eAlexandria Engineering Journal\u003c/em\u003e, \u003cem\u003e90\u003c/em\u003e, 65-75.\u0026rlm;\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eWang, W. Y. (2017). \u0026quot;Liar, Liar Pants on Fire\u0026quot;: A New Benchmark Dataset for Fake News Detection. \u003cem\u003eProceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers)\u003c/em\u003e, 422\u0026ndash;426. https://doi.org/10.18653/v1/P17-2067\u003c/li\u003e\n \u003cli\u003eA. H. J. Almarashy, M. -R. Feizi-Derakhshi and P. Salehpour, \u0026quot;Enhancing Fake News Detection by Multi-Feature Classification,\u0026quot; in \u003cem\u003eIEEE Access\u003c/em\u003e, vol. 11, pp. 139601-139613, 2023, doi: 10.1109/ACCESS.2023.3339621.\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eR. K. Kaliyar, A. Goswami and P. Narang, \u0026quot;FakeBERT: Fake news detection in social media with a BERT-based deep learning approach\u0026quot;, \u003cem\u003eMultimedia Tools Appl.\u003c/em\u003e, vol. 80, no. 8, pp. 11765-11788, Mar. 2021.\u003c/li\u003e\n \u003cli\u003eLuvembe, A. M., Li, W., Li, S., Liu, F., \u0026amp; Wu, X. (2024). CAF-ODNN: Complementary attention fusion with optimized deep neural network for multimodal fake news detection. \u003cem\u003eInformation Processing \u0026amp; Management\u003c/em\u003e, \u003cem\u003e61\u003c/em\u003e(3), 103653.\u0026rlm;\u003c/li\u003e\n \u003cli\u003eTruică, C. O., Apostol, E. S., \u0026amp; Karras, P. (2024). DANES: Deep neural network ensemble architecture for social and textual context-aware fake news detection. Knowledge-Based Systems, 294, 111715.\u0026rlm;\u003c/li\u003e\n \u003cli\u003eThaokar, C., Rout, J. K., Das, H., \u0026amp; Rout, M. (2024). Fake content detection on benchmark dataset using various deep learning models. International Journal of Computational Science and Engineering, 27(5), 570-581.\u0026rlm;\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eShu, K., Mahudeswaran, D., Wang, S., Lee, D., Liu, H.: FakeNewsNet: A data repository with news content, social context, and dynamic information for fake news research. Big Data 8, 171\u0026ndash;188 (2020). https://doi.org/10.1089/big.2020.0062\u003c/li\u003e\n \u003cli\u003eQian, F., Gong, C., Wu, Y., Huang, W.: Neural user response generator: Fake news detection with collective user intelligence. In: Proceedings of the 27th International Joint Conference on Artificial Intelligence (IJCAI), pp. 3834\u0026ndash;3840 (2018). https://doi.org/10.24963/ijcai.2018/533\u003c/li\u003e\n \u003cli\u003eZhang, L., Cui, L., Fu, Y., Huang, Y., Wei, F.: MCN: Multichannel network for fake news detection. In: Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (ACL), pp. 2341\u0026ndash;2351 (2019). https://doi.org/10.18653/v1/P19-1225\u003c/li\u003e\n \u003cli\u003eGhenai, A., Mejova, Y.: Fake cures: User-centric modeling of health misinformation in social media. In: Proceedings of the ACM Conference on Computer-Supported Cooperative Work and Social Computing (CSCW), pp. 1\u0026ndash;15 (2018). https://doi.org/10.1145/3274314\u003c/li\u003e\n \u003cli\u003eWang, Y., McKee, M., Torbica, A., Stuckler, D.: Systematic literature review on the spread of health-related misinformation on social media. Social Science \u0026amp; Medicine 240, 112552 (2019). https://doi.org/10.1016/j.socscimed.2019.112552\u003c/li\u003e\n \u003cli\u003eKhan, S., Ullah, I., Ahmad, I., et al.: A deep learning-based framework for detecting COVID-19 related misinformation on social media. Computers, Materials \u0026amp; Continua 69(2), 1981\u0026ndash;1995 (2021). https://doi.org/10.32604/cmc.2021.015775\u003c/li\u003e\n \u003cli\u003eImran, A.S., Daudpota, S.M., Kastrati, Z., Batra, R.: Cross-cultural polarity and emotion detection using transfer learning. Information Processing \u0026amp; Management 58(1), 102361 (2021). https://doi.org/10.1016/j.ipm.2020.102361\u003c/li\u003e\n \u003cli\u003eZhou, X., Mulay, A., Ferrara, E., Zafarani, R.: ReCOVery: A multimodal repository for COVID-19 news credibility research. In: Proceedings of the 29th ACM International Conference on Information \u0026amp; Knowledge Management (CIKM), pp. 3205\u0026ndash;3212 (2020).\u0026nbsp;\u003ca href=\"https://doi.org/10.1145/3340531.3412880\"\u003ehttps://doi.org/10.1145/3340531.3412880\u003c/a\u003e\u003c/li\u003e\n \u003cli\u003eDheyaa Radhi, A., Ali Hussein Al Naffakh, H., Fuqdan, A. I., A Hakim, B., \u0026amp; Al-Attar, B. (2024). A comprehensive review of machine learning-based models for fake news detection. In BIO Web of Conferences (Vol. 97, p. 00123). EDP Sciences.\u0026rlm;\u003c/li\u003e\n \u003cli\u003eAl-Kabbi, H. A., Feizi-Derakhshi, M. R., \u0026amp; Pashazadeh, S. (2024). A Hierarchical Two-Level Feature Fusion Approach for SMS Spam Filtering. Intelligent Automation \u0026amp; Soft Computing, 39(4).\u0026rlm;\u003c/li\u003e\n \u003cli\u003eAlfreihat, M., Almousa, O., Tashtoush, Y., AlSobeh, A., Mansour, K., \u0026amp; Migdady, H. (2024). Emo-SL Framework: Emoji Sentiment Lexicon Using Text-Based Features and Machine Learning for Sentiment Analysis. IEEE Access.\u0026rlm;\u003c/li\u003e\n \u003cli\u003eAl-Kaabi, H., Darroudi, A. D., \u0026amp; Jasim, A. K. (2024). Survey of SMS Spam Detection Techniques: A Taxonomy. AlKadhim Journal for Computer Science, 2(4), 23-34.\u0026rlm;\u003c/li\u003e\n \u003cli\u003eDevlin, J., Chang, M. W., Lee, K., \u0026amp; Toutanova, K. (2019, June). Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers) (pp. 4171-4186).\u0026rlm;\u003c/li\u003e\n \u003cli\u003eZhu, Q., Jiang, X., \u0026amp; Ye, R. (2021). Sentiment analysis of review text based on BiGRU-attention and hybrid CNN. \u003cem\u003eIEEE Access\u003c/em\u003e, \u003cem\u003e9\u003c/em\u003e, 149077-149088.\u0026rlm;\u003c/li\u003e\n \u003cli\u003eKumar, A. (2022). Contextual semantics using hierarchical attention network for sentiment classification in social internet-of-things. \u003cem\u003eMultimedia Tools and Applications\u003c/em\u003e, \u003cem\u003e81\u003c/em\u003e(26), 36967-36982.\u0026rlm;\u003c/li\u003e\n \u003cli\u003eYang T. (2019) Hierarchical Attention Networks. \u003cem\u003eMedium\u003c/em\u003e. https://medium.com/analytics-vidhya/hierarchical-attention-networks-d220318cf87e\u003c/li\u003e\n \u003cli\u003eHassan, A. A. R., \u0026amp; Feizi-Derakhshi, M. R. (2025). ResCBAR-FusionNet: A Hybrid CNN-BiGRU-Attention Model for Human Activity Recognition. \u003cem\u003eSignal, Image and Video Processing\u003c/em\u003e, \u003cem\u003e19\u003c/em\u003e(8), 1-9.\u003c/li\u003e\n \u003cli\u003eShu, K., Mahudeswaran, D., Wang, S., Lee, D., \u0026amp; Liu, H. (2020). FakeNewsNet: A Data Repository with News Content, Social Context, and Dynamic Information for Studying Fake News on Social Media. Big Data, 8(3), 171\u0026ndash;188. https://doi.org/10.1089/big.2020.0062\u003c/li\u003e\n \u003cli\u003eISOT Research Lab. (n.d.). Fake News Detection Datasets. University of Victoria. Available at: https://onlineacademiccommunity.uvic.ca/isot/2022/11/27/fake-news-detection-datasets/\u003c/li\u003e\n \u003cli\u003eCui, L., \u0026amp; Lee, D. (2020). CoAID: COVID-19 Healthcare Misinformation Dataset. arXiv preprint arXiv:2006.00885. https://arxiv.org/abs/2006.00885\u003c/li\u003e\n \u003cli\u003eZhou, X., Mulay, A., Ferrara, E., \u0026amp; Zafarani, R. (2020). ReCOVery: A Multimodal Repository for COVID-19 News Credibility Research. arXiv preprint arXiv:2006.05557. https://arxiv.org/abs/2006.05557\u003c/li\u003e\n \u003cli\u003eKamber, A. N. (2020). A variety of electrocardiogram (ECG) signal processing. \u003cem\u003eSolid State Technology\u003c/em\u003e, \u003cem\u003e63\u003c/em\u003e(3), 5370-5377.\u0026rlm;\u003c/li\u003e\n \u003cli\u003eSabri, R. F., \u0026amp; Abdullah, N. A. (2025). Extremism Detection in the Iraqi Dialect Based on Machine Learning. \u003cem\u003eIraqi Journal of Science\u003c/em\u003e.\u0026rlm;\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"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":"Fake News Detection, BERT Embedding, Hierarchical Attention Network (HAN), Deep Learning","lastPublishedDoi":"10.21203/rs.3.rs-6893733/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6893733/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThe proliferation of misinformation and fake news on digital platforms has emerged as a critical challenge, undermining public trust and exacerbating societal polarization. Traditional verification methods are inadequate for rapid information dissemination on social media, necessitating advanced automated solutions. This paper proposes a novel fake news detection model, BCGH-Net, which integrates BERT embeddings with a multi-level feature extraction framework. The model employs Convolutional Neural Networks (CNN) for word-level feature extraction and Bidirectional Gated Recurrent Units (Bi-GRU) for sequential sentence-level dependencies, fused via a Hierarchical Attention Network (HAN) to prioritize salient textual elements. Evaluated on four benchmark datasets (FakeNewsNet, ISOT, CoAID, and ReCOVery), BCGH-Net achieves superior performance across four benchmark datasets, demonstrating its robustness and generalizability, with accuracies of 95.8%, 99.3%, 98.7%, and 92.4%, respectively. The results demonstrate its robustness across diverse domains, including political and health-related misinformation. Key contributions include a two-stage feature extraction approach, hierarchical attention fusion, and superior performance metrics (precision, recall, F1-score) compared to existing models. Despite its computational complexity, BCGH-Net offers a scalable solution for real-world deployment, addressing the urgent need for reliable fake news detection systems.\u003c/p\u003e","manuscriptTitle":"BCGH-Net: A Hierarchical Neural Framework for Fake News Detection Using BERT and Attention Fusion","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-07-24 15:34:52","doi":"10.21203/rs.3.rs-6893733/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":"781acd7c-5173-48a9-9602-7b25e158acda","owner":[],"postedDate":"July 24th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-08-11T18:23:28+00:00","versionOfRecord":[],"versionCreatedAt":"2025-07-24 15:34:52","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6893733","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6893733","identity":"rs-6893733","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}