Classification and Generation of Arabic News Titles from Raw Text Based on an Encoder-Decoder Transformer Model (mT5)

Classification and Generation of Arabic News Titles from Raw Text Based on an Encoder-Decoder Transformer Model (mT5) | 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 Classification and Generation of Arabic News Titles from Raw Text Based on an Encoder-Decoder Transformer Model (mT5) ‪Ayedh Abdulaziz Mohsen‬‏, Marwah Yahya Al-Nahari, Akram Alsubari This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-3982909/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 Multilingual Transformer 5 (MT5) is a versatile architecture in natural language processing (NLP) that demonstrates proficiency across various languages. This study aimed to improve the performance of the MT5 model in two key tasks: topic classification and headline generation. The datasets used were 183K and 294K samples. The classification task involved categorizing news articles, while the news generation task aimed to create coherent and contextually relevant Arabic news content. Through careful fine-tuning and rigorous evaluation, the MT5 model significantly advances its ability to address complex challenges in Arabic NLP. This study provides practical insights into real-world applications in processing Arab news. The performance of the MT5 model was evaluated using various online platforms. The mT5small model achieved an accuracy of 0.7858 and an F1 score of 0.7858, while the mT5base model achieved an accuracy of 0.8230 and an F1 score of 0.8230. The generative approach for headline generation yielded Rouge-1, Rouge-2, and Rouge-L scores under the task "Generative of Headlines." These outcomes demonstrate the effectiveness of the fine-tuned MT5 model across various evaluation metrics and tasks, confirming its potential for practical applications in Arabic NLP. Artificial Intelligence and Machine Learning Transformer Decoder Encoder Title Generation and Classification MT5 Arabic NLP Fine-tuning Figures Figure 1 Figure 2 I. Introduction Arabic, as a language, poses unique challenges in NLP tasks due to its rich morphology, diverse dialects, and contextual intricacies ‎[1]. To address these challenges, studies have concentrated on two pivotal tasks: the classification of news articles and the generation of news content ‎[2]. News classification holds relevance in efficiently organizing vast amounts of information, aiding in targeted content retrieval and analysis. Concurrently, news generation entails the creation of coherent and contextually relevant articles, contributing to automated content production in Arabic journalism. Despite the strides made in NLP, the specific intricacies of Arabic language processing remain relatively underexplored, necessitating dedicated research to address the nuances of Arabic text ‎[3]. In the realm of Arabic journalism and news processing, the challenges are further compounded. Efficiently categorizing and comprehending news articles, as well as generating coherent and contextually relevant content, demands a nuanced understanding of the language. The need for automated systems that can accurately classify news articles and generate relevant content is vital in today's rapidly evolving information landscape. By delving into the background of transformer-based models, the challenges posed by Arabic language intricacies, and the need for specialized NLP solutions in the realm of Arabic news, this research endeavors to pave the way for advancements in the field. This unique contribution lies in the thorough exploration of fine-tuning MT5 for Arabic NLP tasks, with a specific focus on classification and news generation, aiming to bridge the gap between transformer-based models and the complexities of the Arabic language. Multilingual Transformer 5 (MT5) ‎[4] presents a compelling solution, given its proficiency across multiple languages. However, for it to excel in Arabic NLP tasks, fine-tuning is imperative. Fine-tuning involves training the model on domain-specific data, adapting it to the intricacies of the target language and tasks. The motivation for this study stems from the need to augment the capabilities of MT5 in handling complex Arabic NLP tasks. Leveraging datasets collected from the internet, the approach used encompasses a multifaceted fine-tuning strategy, focusing on individual tasks for in-depth exploration and optimization. This research not only seeks to enhance the model's performance but also to provide insights into the generalizability and adaptability of transformer models to Arabic language intricacies. The objectives of this study are twofold: first, to rigorously investigate the fine-tuning of MT5 for Arabic NLP tasks, considering the unique language challenges; second, to contribute empirical evidence and practical insights for the classification of news articles and the generation of news content in Arabic, thereby addressing the existing gaps in the literature pertaining to Arabic-focused transformer models.The next section describes related work, offering insight into previous research and relevant work related to the paper's topic. In section III, the methodology is presented, explaining the data collection process, data preprocessing, data analysis, and evolutionary methods utilized. Subsequently, the fourth section addresses the Results & Discussion section, showcasing the obtained results and providing an in-depth discussion, highlighting achievements and improvements. The paper concludes with the fifth section, the conclusion, summarizing the results and offering a final evaluation. Finally, references are included, comprising studies and sources used during the writing of the scientific paper. II. Related Work In the dynamic landscape of natural language processing (NLP), the years spanning from 2020 to 2023 witnessed significant strides in understanding and harnessing the intricacies of the Arabic language. This comprehensive literature review delves into 13 seminal research papers, providing a nuanced exploration of the models employed, the datasets utilized, and the performance metrics and accuracy measures considered in the realm of Arabic NLP. The journey commences with the work of Linting Xue et al. ‎[4], who introduced the mT5 variation of T5, which was pretrained on a groundbreaking common crawling-based dataset. Their study not only emphasized the cutting-edge performance of mT5 on multilingual benchmarks but also addressed the challenges in zero-shot scenarios. The relevance of the T5 recipe in a multilingual environment was underscored, showcasing its versatility across languages. Mehrdad Farahani et al.‎[5] contributed to the landscape by employing mT5 and the ParsBERT model for refining Persian text summarization. Recognizing the dearth of attention given to this facet of NLP, the researchers introduced two pretrained approaches based on the multilingual T5 model and ParsBERT. Simultaneously, they created and released a summary dataset, providing a valuable resource for further exploration. Remi Calizzano et al. ‎[6] navigated the classification of German Facebook comments using GermEval 2021, prioritizing data gathering through data augmentation and task-specific pretraining. Their utilization of 12 datasets for pretraining two multilingual models marked a strategic approach to enhance the classification accuracy. El Moatez Billah Nagoudi et al. ‎[7] presented three potent Arabic text-to-text models, setting the stage for benchmarking with ARGEN—a compilation of seven tasks from 19 datasets. Their work showcased the superior performance of these models compared to mT5, even in machine translation involving four different foreign languages, underlining the adaptability of these models. Aleksandra Chrabrowa et al. ‎[8] introduced a new standard for evaluating Polish text-to-text models, emphasizing the need for standardized datasets in summarization and question answering for the Polish language. The introduction of plT5, a general-purpose Polish text-to-text model, further expanded the scope of language-specific models. Ahlam Fuad et al. ‎[9] leveraged the mT5 multilingual transformer model to present the first full Arabic generative model for task-oriented dialog systems (AraConv). Their work, anchored by the Arabic Dialog Dataset (Arabic-TOD), showcased a hybrid tuning method for mT5-based transformer models, achieving competitive results in polarity detection and gravity prediction subtasks. Table 1 Some previous studies, datasets and results for the classification task. Ref model Dataset f1 scores Accuracy ‎[28] mT5 MSA 0.70 0.72 ‎[29] AraT5 ARGEN 0.84 0.88 ‎[31] AraBERT (Target/BERT) Saaq al-Bambuu 0.989 0.989 ‎[32] AraGPT2 Wikipedia - 98% In the realm of classification tasks, Table 1 showcased a diverse range of models applied to various datasets, each contributing valuable insights and advancements. Notably, the use of the mT5 model on the Modern Standard Arabic (MSA) dataset, as reported in ‎[28], demonstrated moderate yet commendable performance, with an F1-score of 0.70 and an accuracy of 0.72. Moreover, the AraBERT (Target/BERT) model, discussed in ‎[31], exhibited exceptional capabilities when applied to the Saaq al-Bambuu dataset, achieving an impressive F1-score and accuracy of 0.989. Additionally, the application of AraGPT2 to the Wikipedia dataset, as detailed in ‎[32], resulted in a notable accuracy of 98%, although the F1-score was not explicitly provided. These findings underscore the efficacy of state-of-the-art models in handling classification tasks across diverse linguistic and thematic domains. As delve into the specifics of each model's performance on distinct datasets, it becomes evident that the nuanced nature of the tasks and the characteristics of the data play crucial roles in shaping the effectiveness of these models. Table 2 shows some previous studies, datasets and results for the generation task. In the landscape of generation tasks, recent studies have employed diverse models to address the challenges posed by different datasets, showcasing notable advancements in the Table 2 AraBERT (Target/BERT), as examined in ‎[31], and demonstrated its effectiveness in the generation task when applied to the Saaq al-Bambuu dataset, achieving a precision (P) of 0.848, a recall (R) of 0.823, and an F1-score of 0.879. The study in ‎[33] introduced AraT5, utilizing the XSum and OrangeSum datasets, which showed significant improvements, with R1, R2, and BLEU values of 7.5, 18.30, and an undetermined value, respectively. Furthermore, the application of mT5 to the APGC dataset, as discussed in ‎[34], resulted in a precision of 71.6%, an F1-score of 0.820, and a BLEU score of 97.5. The generation task on the Arabic-TOD and Arabic-NTG datasets, as explored by mT5 and AraT5 in ‎[35] and ‎[36], respectively, revealed noteworthy contributions with BLEU scores of 31.05. These findings underscore the diverse applications of state-of-the-art models in addressing generation tasks across different datasets, highlighting the need for tailored approaches based on the characteristics of the specific data at hand. III. Methodology The approach to enhancing Arabic NLP transformer models involves a detailed examination of existing fine-tuning methods, focusing on news classification and generation. Diverse datasets of Arabic news are collected and meticulously curated to capture linguistic variations. The Fig. 1 to depict processes involving text input and processing through a model mT5, resulting in a labeled and title output Comprehensive analysis of the dataset provides insights into its distribution and characteristics. Customized fine-tuning strategies for the multilingual transformer 5 (MT5) model are developed based on dataset insights, addressing language nuances. Evaluation metrics are established to assess model performance in Arabic news processing. The aim is to improve model capabilities and performance for news tasks with greater accuracy and efficiency. a. Data collection The dataset was collected from the websites Almashhadalaraby and Kaggle, which include a large amount of Arabic news. The dataset is a large collection of Arabic news articles and categories to which the classification has been applied. Here, code was applied to the article, and a portion of the text was extracted as a title to be used in generation because during the collection of news data, titles were not pulled from websites, where there are many linguistic programming tasks. Neural techniques can be implemented, such as text generation, text classification, and text summarization. Table 3 dataset Task Total samples Training set Validation set Test set Classification 183955 105K 33K 15K Generations 294307 50K 10K 10K b. Pre-Preprocessing The titles and the articles are preprocessed such that punctuations are separated from the words and Arabic diacritics are removed. The dataset is divided into training, validation, and test sets of sizes, as shown in Table 3 . The training, validation, and test sets were created to ensure the proper evaluation and training of the models. The datasets shown in Table 3 , encompassing 183,955 samples for the classification task and 294,307 samples for the generation task, are further divided into training, validation, and test sets to facilitate comprehensive training and evaluation of the Multilingual Transformer 5 (MT5) model. The classification task involves categorizing news articles, while the generation task focuses on producing coherent and contextually relevant Arabic news content. c. Data Analysis The Generation dataset has a total of 29 million words composed of 300K unique words. The average title length is 6.36 words, while the average article length is 736 characters. Most words are rare, which means that they are seen fewer than 5 times in the entire corpus. The Classification dataset has a total of 6 million words and is composed of 300,000 unique words. The average title length is 9.95 words, while the average article length is 281 characters, as shown in Table 4 . eight classes and produce word embedding models. The data contained nine categories (sports, political, financial, artistic, technical, medical, cultural, and religious), as shown in Fig. 2 . The graph in Fig. 2 shows that circular drawings represent different categories and their proportions. It contains several colorful sections representing categories such as sports, politics, finance, art, technology, medicine, culture, religion and community. The pie chart gives a quick look at the size of each category compared to the next, showing that the Sports category is 32% larger than the pie chart, followed by political, financial, artistic, technical, medical, cultural, religious, and community news. Through comparative analysis with other researchers, the methods utilized for both classification and generation exhibited promising results across various metrics. These favorable outcomes emphasize the effectiveness of the selected methodologies in tackling the challenges associated with news classification and generation. Nonetheless, it is essential to recognize a notable obstacle encountered during insufficient research resources for additional model refinement. Despite the achievement of noteworthy results, the potential for enhancing model accuracy remains constrained by limited resources. This constraint is particularly pronounced when considering the demanding nature of classification and generation tasks, especially within the domain of Arabic news. The intricate linguistic nuances and contextual intricacies in Arabic news present unique challenges that necessitate sophisticated models for optimal performance. In light of these challenges, recommendations for future work emphasize the utilization of the MT5 model. While acknowledging its resource-intensive nature, the MT5 model holds promise for significantly improving the classification and generation processes in the context of Arabic news. The recommendation stems from the understanding that, despite the associated resource demands, leveraging advanced models such as mT5 is instrumental in pushing the boundaries of accuracy and effectiveness. In conclusion, the results of comparisons with other researchers highlight the success of current techniques used in classification and generation. However, the identified resource constraint underscores the importance of continued efforts to secure the necessary resources for further model refinement. The proposed adoption of the MT5 model represents a strategic move, acknowledging its resource-intensive nature while emphasizing its potential to improve the performance of classification and generation tasks in the challenging domain of Arabic news. Table 4 Analytical Exploration Comparison Generation dataset Classification dataset Number of words in the file 2937143 word 6000000 word Number of unique words 300000 word 300000 word Average title length 6.36 Letter 9.95 Letter Average article length 736 Letter 281 Letter Dataset 294307 183955 Number of categories - 8 Table 4 presents a concise summary of two datasets: "Generation" and "Classification." It includes key metrics such as word count, unique words, average title length, average article length, and the number of categories. These metrics provide valuable insights into the size, vocabulary, and structure of the datasets, enabling researchers to better understand and analyze the data. 1) Calculating the number of words : Number of words = text length/word length 2) Counting the number of unique words : Number of unique words = number of words - number of frequent words 3) Calculating average word length : Average word length = sum of word lengths/number of words 4) Calculating the average article length : Average article length = sum of article lengths/number of articles d. Model Fine-Tuning Model fine-tuning entailed the use of monolingual datasets sourced from various news outlets for training and assessing the models. The news datasets, comprising Arabic articles, were gathered from diverse online sources and preprocessed by segregating punctuation from words and eliminating Arabic diacritics. A unified text-to-text modeling framework is adopted to establish a baseline model for a range of tasks supported by a dataset. Specifically, mT5-base ‎[1]was fine-tuned for each task in a single-task setting. mT5 is a large multilingual pretrained language model (PLM) based on the English-only T5 model ‎[10]and consists of an encoder-decoder transformer architecture ‎[11]. However, earlier PLMs consisted of either a standalone encoder for classification and representation tasks ‎[17] or a decoder for generative tasks ‎[18], and the architecture of mT5 offers two major benefits. First, the complexity of multitask NLP pipelines can be reduced by removing the need for specialized task-specific frameworks. Second, it may also lead to improved performance since sharing a unified set of parameters between related tasks can be beneficial ‎[19]. For each of the tasks, a distinct natural-language prefix that is intended to aid the model in learning the task is selected (e.g., “summaries:” for bullet-point summary generation, “generate title:” for headline generation). In the single-task setting, mT5 is fine-tuned on each of the tasks in isolation, resulting in one model per task. The training of the Seq2Seq models was carried out on the training set, while model selection and hyperparameter tuning were performed on the validation set. A consistent experimental setup was maintained to fine-tune and assess the models, adopting an approach akin to subtask C with adjustments tailored for the Arabic language. The hidden size and embedding size were both set at 256. The key hyperparameters included a learning rate of 4e-5, a batch size of 8, 5 epochs, and a weight decay of 0.01. Additionally, FP16 precision was assessed using the fp16 = true setting in Seq2SeqTrainingArguments. e. Evaluation Methods Likewise, most TS evaluation approaches have been driven by other similar NLP research areas. Various evaluation methods have been applied across studies to measure the three main aspects of the newly generated text. These aspects are: i) Fluency : Ensuring grammatical accuracy and coherence in the generated text, maintaining syntactic consistency, and creating a natural flow of language. ii)Adequacy Preserving the intended meaning of the input during generation, capturing the essence of the message, and avoiding distortions or misinterpretations. iii) Simplicity : Enhancing readability and comprehension by using clear and concise language, avoiding overly complex sentence structures, and organizing information in a straightforward manner. Classification approach – automatic evaluation To evaluate the accuracy of the classification approach, we used a standard set of evaluation metrics, including precision, recall, and F1-score ‎[23]. The experimental setup included the use of monolingual datasets from multiple news sources, which were preprocessed by separating punctuation from words and removing Arabic diacritics. The classification approach was evaluated using standard metrics such as precision, recall, and F1-score. The models were trained using diverse Arabic news datasets obtained from online sources. After preprocessing, the obtained dataset was used to train and evaluate the models for the classification approach. Enhanced Computational Approaches Given an input text sequence 𝑥, the classification model can be considered a function, i.e., 𝑓(𝑥) = 𝐶 for measuring the conditional probability distributions over all possible labels in the predefined category set 𝐶 = {𝑐 1 , 𝑐 2 , 𝑐 3 ..., 𝑐 𝐿 }. Text classification works on an instance of space 𝒳 where each instance is an input text sequence 𝑥. Hence, let 𝑋 = {𝑥 1 , 𝑥 2 , 𝑥 3 ..., 𝑥 𝑁 } be the training set, and each input text sequence 𝑥 𝑖 = {𝑡 1 , 𝑡 2 , 𝑡 3 ..., 𝑡𝑃} represents a sequence of tokens. 𝑌 = {𝑦 1 , 𝑦 2 , 𝑦 3 , ..., 𝑦 𝑁 } the set of categories in which each document 𝑥 𝑖 is classified, each y 𝑖 ∈ 𝐶. The model was trained for the set of parameters 𝜃 ∈ \({IR}^{q}\) minimizing the cross-entropy loss in Equation ( \(4\) ): $${\theta }{=}_{{\theta }\in {\text{I}\text{R}}^{\text{q}}} \text{F}\left({\theta }, \text{X}\right)$$ 1 $$\text{F}\left({\theta }, \text{X}\right)= - \sum _{i=1}^{N}\sum _{c=1}^{L}{{\Upsilon }}_{{x}_{i|c}} \text{l}\text{o}\text{g}\left({\text{P}}_{{x}_{i|c}} \left({\theta }\right)\right)$$ 2 $${{\Upsilon }}_{{x}_{i|c}}=\left\{\begin{array}{c}1 if {\text{y}}_{\text{i}} = c\\ 0 otherwise \end{array}\right.$$ 3 where \({\varvec{P}}_{{x}_{i|c}} \left(\varvec{\theta }\right)\) is the predicted conditional probability of 𝑥 𝑖 giving class C. To calculate the F1 score, the equations below were used. TP refers to ‘true positive’, which counts the tokens appearing in both the prediction and the ground truth. FP refers to a ‘false positive’ that counts the tokens that appear only in the prediction. FN refers to ‘False Negative’, which means the number of tokens that appear only in the ground truth. The F1 score of the dataset is an average of the F1 score of every sample, and \(F1\) ᵢ is the individual value of F1 calculated for each i from 1 to N. $$precision=\frac{TP}{TP+FP}$$ 4 $$F1=\frac{2 \times precision \times Recall}{precision+Recall}$$ 5 $$F1=\frac{\sum _{i=1}^{N}{F1}_{i }}{N}$$ 6 Generative approach - automatic evaluation The primary evaluation metric used for the generation task is ROUGE, which measures recall ‎[14]. ROUGE indicates how much of the human-produced titles appear in the model-generated titles. Moreover, the BLEU is also used as a secondary metric to measure precision ‎[15]. BLEU [8], an n-gram-based evaluation metric. These methods were not able to capture two main simplification features: 1) Changing word order as a paraphrasing simplification method. 2) Maintaining the deep structure meaning despite changes in the surface form structure. Specifically, how much of the model-generated titles appear in the human-produced titles? To compute attention, one typically scores a hidden state of the decoder and a corresponding state of the encoder. The attention ratings are then normalized over all encoder states, resulting in an attention score of 1, as shown in Eq. (7). IV. RESULTS & DISCUSSION a. Classification Results The classification approach yielded promising results, with precision and F1-score values indicating the effectiveness of the models in accurately categorizing Arabic news articles. The models exhibited strong performance across the diverse categories present in the dataset, showcasing their ability to handle multiclass classification tasks effectively. These results affirm the suitability of transfer learning and text-to-text models in addressing classification challenges within Arabic NLP, thereby contributing to the advancement of language-specific natural language processing techniques. Table 5 Shows the evaluation of mT5 model Model Accuracy F1 Score mT5small 0.7858 0.7858 mT5base 0.8230 0.8230 mT5large 0.8742 0.8742 Table 5 presents the results of the evaluation dataset for the mT5 model. These metrics, including accuracy and F1 score, provide a comprehensive evaluation of the mT5 model's performance across different sizes (small, base, and large). The results showcase the model's proficiency in handling Arabic natural language processing tasks. b. Generative Results The generative approach produced compelling outcomes, as evidenced by the automatic evaluation metrics. The models demonstrated low perplexity, indicative of their proficient generation of coherent and contextually relevant Arabic text. Furthermore, the high BLEU and ROUGE scores underscored the ability of the models to generate outputs that closely align with reference summaries and maintain semantic coherence. These results emphasize the potential of transfer learning and text-to-text models in facilitating the generation of high-quality Arabic text, thereby enriching the landscape of natural language generation in the language domain. Table 6 Shows the results of generative approach Task Rouge-1 Rouge-2 Rouge-L Generative of Headlines 9.719 1.575 9.315 The outcomes, as depicted in Table 6 , illustrate the performance of the generative approach for the specified task of headline generation. The evaluation metrics include the Rouge-1, Rouge-2, and Rouge-L scores, providing insights into the model's effectiveness in generating coherent and contextually relevant Arabic news headlines. The results indicate notable scores, with Rouge-1 at 9.719, Rouge-2 at 1.575, and Rouge-L at 9.315, confirming the proficiency of the generative model in addressing the complexities of Arabic natural language processing tasks. c. Discussion The presented discussion delves into the comprehensive evaluation of the mT5 model across various scales and its application in a generative task, shedding light on its performance metrics and implications for advancing Arabic natural language processing (NLP). The findings are detailed in Table 5 and Table 6 , providing a nuanced understanding of the model's capabilities in both classification and generation tasks. Comparison Analysis of Classification Results : After comparing the results from Table 1 , which show findings from previous studies, with those presented in Table 5 for the mT5 model, the following findings emerged: For the "MSA" dataset, the mT5 model achieved a competitive accuracy of 0.72, comparable to that of the mT5 model in ‎[28]. Compared with AraT5 on the "ARGEN" dataset, the mT5base model demonstrates similar accuracy but a slightly lower F1-score. AraBERT (Target/BERT) achieves outstanding results on the "Saaq al-Bambuu" dataset, with extremely high F1-scores and accuracies. The mT5base model, while achieving lower scores, still performs well in this context. For the "Wikipedia" dataset, AraGPT2 attains an impressive accuracy of 98%, whereas mT5 models show competitive results, especially mT5large, with an accuracy of 87%. Comparison Analysis of the Generative Results While the descriptive analysis provides insights into the performance of mT5 models across various datasets, the generative approach's results in Table 6 focus on the headline generation task and Rouge scores, limiting direct comparisons for the mentioned datasets. Performance Metrics of the mT5 Model The mT5 model, in its different variants—mT5small, mT5base, and mT5large—underwent evaluation based on accuracy and F1 score across diverse dataset sizes. The accuracy measures the overall correctness of the model's predictions, while the F1 score reflects the balance between precision and recall. The results indicate a progressive improvement in performance with increasing model size. Notably, mT5large outperforms mT5base and mT5small, demonstrating the scalability and efficiency of the model across a range of complexities within the Arabic language. This scalability is crucial in real-world applications where the complexity and size of the datasets can vary. The mT5 model's adaptability across different scales suggests its potential to handle diverse linguistic challenges encountered in Arabic NLP. Generative Task Results Table 6 focuses on the outcomes of a generative task, specifically the generation of headlines, assessed through Rouge-1, Rouge-2, and Rouge-L scores. Rouge metrics evaluate the overlap and similarity between generated and reference texts, providing insights into the quality of the generated content. The Rouge-1, Rouge-2, and Rouge-L scores of 9.719, 1.575, and 9.315, respectively, suggest a notable proficiency in generating headlines that exhibit semantic coherence and similarity to reference headlines. The ability of the mT5 model to perform well in generative tasks underscores its versatility in understanding and reproducing nuanced language structures in Arabic. This is particularly significant in applications requiring content creation, such as summarization, headline generation, and paraphrasing. Transfer Learning and Text-to-Text Models in Arabic NLP The combined results underscore the effectiveness of transfer learning and the text-to-text paradigm in advancing Arabic NLP. Transfer learning enables the model to leverage knowledge from pretraining tasks, showcasing its ability to generalize well across diverse linguistic challenges. The text-to-text approach, where all NLP tasks are framed as text generation tasks, proves to be a powerful and unifying paradigm for various tasks. The robust performance of the mT5 model in both classification and generation tasks suggests its potential for a wide array of applications within Arabic NLP, including text classification, generation, simplification, and sentiment analysis. These findings contribute valuable insights into the development of models tailored for the Arabic language, paving the way for innovative applications and deeper understanding within this domain. In conclusion, the discussion emphasizes the pivotal role of transfer learning and text-to-text models in improving Arabic NLP, providing evidence of their impact on performance across different tasks. The presented results open avenues for future research and development, encouraging the exploration of novel applications and advancements in the field of Arabic language processing. V. Conclusion The results from both the generative and classification approaches highlight the effectiveness of employing transfer learning and text-to-text models to advance Arabic natural language processing. The robust performance observed in classification and generation tasks underscores the adaptability and efficacy of these models in addressing diverse language-specific challenges. Notably, the best results mentioned in the initial tables include an impressive Rouge-1 score of 9.719 and an accuracy of 0.8742 in the classification task using the mT5large model. These outcomes affirm the significance of leveraging such models to drive progress in Arabic NLP, offering promising applications in text classification, generation, simplification, and sentiment analysis. As the field of natural language processing evolves, the utilization of these models is likely to contribute significantly to the development of cutting-edge solutions and advancements in Arabic NLP. It is imperative to acknowledge the pivotal role played by high-quality corpora and specialized tools in data collection and annotation, ensuring the accuracy of the classification and generation processes. References Marie-Sainte, S. L., Alalyani, N., Alotaibi, S., Ghouzali, S., & Abunadi, I. (2018). Arabic natural language processing and machine learning-based systems. IEEE Access, 7, 7011-7020. ‏ Almomani, A., Al-Nawasrah, A., Alauthman, M., Al-Betar, M. A., & Meziane, F. (2021). Botnet detection used fast-flux technique, based on adaptive dynamic evolving spiking neural network algorithm. International Journal of Ad Hoc and Ubiquitous Computing, 36(1), 50-65. ‏ Alkhurayyif, Y., & Sait, A. R. W. (2023). Developing an Open Domain Arabic Question Answering System Using a Deep Learning Technique. IEEE Access. ‏ Xue, L., Constant, N., Roberts, A., Kale, M., Al-Rfou, R., Siddhant, A., ... & Raffel, C. (2020). mT5: A massively multilingual pretrained text-to-text transformer. arXiv preprint arXiv:2010.11934. ‏ Farahani, M., Gharachorloo, M., & Manthouri, M. (2021, March). Leveraging ParsBERT and pretrained mT5 for Persian abstractive text summarization. In 2021 26th International Computer Conference, Computer Society of Iran (CSICC) (pp. 1-6). IEEE. ‏ Calizzano, R., Ostendorff, M., & Rehm, G. (2021, September). DFKI SLT at GermEval 2021: Multilingual Pretraining and Data Augmentation for the Classification of Toxicity in Social Media Comments. In Proceedings of the GermEval 2021 Shared Task on the Identification of Toxic, Engaging, and Fact-Claiming Comments (pp. 25-31). ‏ Nagoudi, E. M. B., Elmadany, A., & Abdul-Mageed, M. (2021). AraT5: Text-to-text transformers for Arabic language generation. arXiv preprint arXiv:2109.12068. ‏ Chrabrowa, A., Dragan, Ł., Grzegorczyk, K., Kajtoch, D., Koszowski, M., Mroczkowski, R., & Rybak, P. (2022). Evaluation of transfer learning for polish with a text-to-text model. arXiv preprint arXiv:2205.08808. ‏ Fuad, A., & Al-Yahya, M. (2022). AraConv: Developing an Arabic task-oriented dialog system using multilingual transformer model mT5. Applied Sciences, 12(4), 1881. ‏ Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., ... & Liu, P. J. (2020). Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1), 5485-5551. ‏ Toledano-López, O. G., Madera, J., González, H., Simón-Cuevas, A., Demeester, T., & Mannens, E. (2022). Finetuning mt5-based transformer via cma-es for sentiment analysis. In Proceedings of the Third Workshop for Iberian Languages Evaluation Forum (IberLEF 2022), CEUR WS Proceedings. Alrowili, S., & Vijay-Shanker, K. (2022, December). Generative approach for gender-rewriting task with ArabicT5. In Proceedings of the The Seventh Arabic Natural Language Processing Workshop (WANLP) (pp. 491-495). ‏ Kale, M., Siddhant, A., Constant, N., Johnson, M., Al-Rfou, R., & Xue, L. (2021). nmT5--Is parallel data still relevant for pretraining massively multilingual language models?. arXiv preprint arXiv:2106.02171. ‏ Khallaf, N., & Sharoff, S. (2022). Toward arabic sentence simplification via classification and generative approaches. arXiv preprint arXiv:2204.09292. ‏ Fuadi, M., Wibawa, A. D., & Sumpeno, S. (2023). idT5: Indonesian Version of Multilingual T5 Transformer. arXiv preprint arXiv:2302.00856. ‏ Štajner, S., Sheang, K. C., & Saggion, H. (2022, June). Sentence simplification capabilities of transfer-based models. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 36, No. 11, pp. 12172-12180). ‏ Cageggi, G., Di Rosa, E., & Uboldi, A. (2023). App2Check at EMit: Large Language Models for Multilabel Emotion Classification. ‏ Lin, C. Y. (2004, July). Rouge: A package for automatic evaluation of summaries. In Text summarization branches out (pp. 74-81). ‏ Papineni, K., Roukos, S., Ward, T., & Zhu, W. J. (2002, July). Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics (pp. 311-318). ‏ Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., ... & Liu, P. J. (2020). Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1), 5485-5551. ‏ Devlin, J., Chang, M. W., Lee, K., & Toutanova, K. (2018). Bert: Pretraining of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805. ‏ Radford, A., Wu, J., Child, R., Luan, D., Amodei, D.,& Sutskever,I.(2019). Language models are unsupervised multitask learners. OpenAI blog, 1(8),9. ‏ Ruder, S. (2017). An overview of multitask learning in deep neural networks. arXiv preprint arXiv:1706.05098. ‏ Papineni, K., Roukos, S., Ward, T., & Zhu, W. J. (2002, July). Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics (pp. 311-318). ‏ Xu, W., Napoles, C., Pavlick, E., Chen, Q., & Callison-Burch, C. (2016). Optimizing statistical machine translation for text simplification. Transactions of the Association for Computational Linguistics, 4, 401-415. ‏ Khorsheed, M. S., & Al-Thubaity, A. O. (2013). Comparative evaluation of text classification techniques using a large diverse Arabic dataset. Language resources and evaluation, 47, 513-538. ‏ Alrowili, S., & Vijay-Shanker, K. (2022, December). Generative approach for gender-rewriting task with ArabicT5. In Proceedings of the The Seventh Arabic Natural Language Processing Workshop (WANLP) (pp. 491-495). ‏ Khallaf, N., & Sharoff, S. (2022). Toward arabic sentence simplification via classification and generative approaches. arXiv preprint arXiv:2204.09292. ‏ Nagoudi, E. M. B., Elmadany, A., & Abdul-Mageed, M. (2021). AraT5: Text-to-text transformers for Arabic language generation. arXiv preprint arXiv:2109.12068. ‏ Luong, T., Brevdo, E., & Zhao, R. (2017). Neural machine translation (seq2seq) tutorial. 2017. URL: https://www. TensorFlow. org/tutorials/seq2seq (дата обращения 17.02. 2018). ‏ Khallaf, N., & Sharoff, S. (2022). Toward arabic sentence simplification via classification and generative approaches. arXiv preprint arXiv:2204.09292. ‏ Antoun, W., Baly, F., & Hajj, H. (2020). AraGPT2: Pretrained transformer for Arabic language generation. arXiv preprint arXiv:2012.15520. ‏ Zmandar, N., El-Haj, M., & Rayson, P. (2023, September). FinAraT5: A text to text model for financial Arabic text understanding and generation. In Proceedings of the 4th Conference on Language, Data and Knowledge (pp. 262-273). ‏ Alrowili, S., & Vijay-Shanker, K. (2022, December). Generative approach for gender-rewriting task with ArabicT5. In Proceedings of the The Seventh Arabic Natural Language Processing Workshop (WANLP) (pp. 491-495). ‏ Fuad, A., & Al-Yahya, M. (2022). AraConv: Developing an Arabic task-oriented dialog system using multilingual transformer model mT5. Applied Sciences, 12(4), 1881. ‏ Elmadany, A., Nagoudi, E. M. B., & Abdul-Mageed, M. (2023). Octopus: A Multitask Model and Toolkit for Arabic Natural Language Generation. arXiv preprint arXiv:2310.16127. Additional Declarations The authors declare no competing interests. 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-3982909","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":274562560,"identity":"ddc600e2-5e28-4501-bab6-6b87fd611ee8","order_by":0,"name":"‪Ayedh Abdulaziz Mohsen‬‏","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA9klEQVRIiWNgGAWjYDACdgY2BsYGBiAJBB+AmI2dkBZmiBYJkBbGGSAtzMRqAbN5ICL4AX8z+7MHH3fY1fExMD/8bPNrmzwfMwPjh485uLVIHOYxN5x5JhnoMDZj6dy+24ZtzAzMkjO34bHmMA+bNG8bM8gvBtK5PbcZgVrYmHnxaJE/zP4MqKUeqIX982/Lntv2BLUYHGYwA2o5DNTCYybN8ON2IkEthod5zCRnnjku2cbMU2bZ23A7uY2ZsRmvX+SOtz+T+Lijml++vX3zjR9/btvOb28++OEjPu/DASg6GNtALFBiIB78IUXxKBgFo2AUjBQAAGYJRSJNBY9wAAAAAElFTkSuQmCC","orcid":"","institution":"Ibb","correspondingAuthor":true,"prefix":"","firstName":"‪Ayedh","middleName":"Abdulaziz","lastName":"Mohsen‬‏","suffix":""},{"id":274562807,"identity":"5dd86002-cb69-4d93-82b6-6e893376a90e","order_by":1,"name":"Marwah Yahya Al-Nahari","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA5ElEQVRIiWNgGAWjYLACxgYbZvv25gNApoQMsVrS2A14jiWAtPAQq+Uwv4GEjwGITViLbvvZZxI/d6RJm0vwfH51o8aCh4H98NEN+LSYnUk3k+w9Y2NsObt3m3XOMaDDeNLSbuDVciCN7QZvW1oyw52z24xz2IBaJHjM8Gs5/4zt5t+2w/UNN3KeGef8I0bLjTS227xth5kNbuQwP85tI0rLM/bfsm1pzJI9x8yYc/skeNgI+uV8GrPh2zYbZn725sefc77VyfGzHz6GVwsyYJMAk8QqBwHmD6SoHgWjYBSMgpEDAPR9SmFeBc2cAAAAAElFTkSuQmCC","orcid":"","institution":"","correspondingAuthor":true,"prefix":"","firstName":"Marwah","middleName":"Yahya","lastName":"Al-Nahari","suffix":""},{"id":274562872,"identity":"4bfa51cb-ef43-4021-81b7-5ba3dde6ea36","order_by":2,"name":"Akram Alsubari","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABGklEQVRIiWNgGAWjYDACHgglA2EZ2NT3g7gJBYS18EC1pDHObABpMSBaC8Nhxg0HQDQeLfw9Zx9+usFgxyM/I/fg54qCNGbj86sTPzwwYJDnFzuAVYvE2XZj6RyGZB6DG3nJkmcMbNjMbrzdLAF0mOHM2QnYrTnPxgDUwsxjIJFjINlgkMZjduPsBpCWBIPb2LXIn2dj/p3DUA90WI7xzwaDwxLGM85u/oFPi8HZNjagLYd5GG7kmAFtOWxgwN+7Da8thmeOsVnnGBznMTjzLs0S6LAEiRu82ywSDCRw+kXuTBrz7ZyKajn59tzDNxv+2CTw95/dfPNHhY08vzQO70Och8yRAKuUwKMcA/AfIEX1KBgFo2AUjAAAALpzWe3RgexuAAAAAElFTkSuQmCC","orcid":"","institution":"","correspondingAuthor":true,"prefix":"","firstName":"Akram","middleName":"","lastName":"Alsubari","suffix":""}],"badges":[],"createdAt":"2024-02-23 19:27:45","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-3982909/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-3982909/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":51629071,"identity":"4436e11c-0eba-4fe1-adbd-7b3b0186e1be","added_by":"auto","created_at":"2024-02-26 08:52:07","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":14942,"visible":true,"origin":"","legend":"\u003cp\u003eSteps of classification and generation\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-3982909/v1/8c782e249cf9b548723162fd.png"},{"id":51629070,"identity":"ec11dc35-688b-4c59-a151-a208d6e79a55","added_by":"auto","created_at":"2024-02-26 08:52:07","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":71176,"visible":true,"origin":"","legend":"\u003cp\u003eTotal Topic Samples at the Classification Task\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-3982909/v1/485a9421ca38e0f5b5a83959.png"},{"id":51629233,"identity":"b0d2d548-7ad6-4184-ba5c-0ee5f9350288","added_by":"auto","created_at":"2024-02-26 09:00:09","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":540887,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-3982909/v1/597e5310-1c4c-4848-8996-7d875b6ed54d.pdf"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003e\u003cstrong\u003eClassification and Generation of Arabic News Titles from Raw Text Based on an Encoder-Decoder Transformer Model (mT5)\u003c/strong\u003e\u003c/p\u003e","fulltext":[{"header":"I. Introduction","content":"\u003cp\u003eArabic, as a language, poses unique challenges in NLP tasks due to its rich morphology, diverse dialects, and contextual intricacies \u0026lrm;[1]. To address these challenges, studies have concentrated on two pivotal tasks: the classification of news articles and the generation of news content \u0026lrm;[2]. News classification holds relevance in efficiently organizing vast amounts of information, aiding in targeted content retrieval and analysis. Concurrently, news generation entails the creation of coherent and contextually relevant articles, contributing to automated content production in Arabic journalism. Despite the strides made in NLP, the specific intricacies of Arabic language processing remain relatively underexplored, necessitating dedicated research to address the nuances of Arabic text \u0026lrm;[3].\u003c/p\u003e \u003cp\u003eIn the realm of Arabic journalism and news processing, the challenges are further compounded. Efficiently categorizing and comprehending news articles, as well as generating coherent and contextually relevant content, demands a nuanced understanding of the language. The need for automated systems that can accurately classify news articles and generate relevant content is vital in today's rapidly evolving information landscape. By delving into the background of transformer-based models, the challenges posed by Arabic language intricacies, and the need for specialized NLP solutions in the realm of Arabic news, this research endeavors to pave the way for advancements in the field. This unique contribution lies in the thorough exploration of fine-tuning MT5 for Arabic NLP tasks, with a specific focus on classification and news generation, aiming to bridge the gap between transformer-based models and the complexities of the Arabic language. Multilingual Transformer 5 (MT5) \u0026lrm;[4] presents a compelling solution, given its proficiency across multiple languages. However, for it to excel in Arabic NLP tasks, fine-tuning is imperative. Fine-tuning involves training the model on domain-specific data, adapting it to the intricacies of the target language and tasks.\u003c/p\u003e \u003cp\u003eThe motivation for this study stems from the need to augment the capabilities of MT5 in handling complex Arabic NLP tasks. Leveraging datasets collected from the internet, the approach used encompasses a multifaceted fine-tuning strategy, focusing on individual tasks for in-depth exploration and optimization. This research not only seeks to enhance the model's performance but also to provide insights into the generalizability and adaptability of transformer models to Arabic language intricacies. The objectives of this study are twofold: first, to rigorously investigate the fine-tuning of MT5 for Arabic NLP tasks, considering the unique language challenges; second, to contribute empirical evidence and practical insights for the classification of news articles and the generation of news content in Arabic, thereby addressing the existing gaps in the literature pertaining to Arabic-focused transformer models.The next section describes related work, offering insight into previous research and relevant work related to the paper's topic. In section III, the methodology is presented, explaining the data collection process, data preprocessing, data analysis, and evolutionary methods utilized. Subsequently, the fourth section addresses the \u003cspan refid=\"Sec4\" class=\"InternalRef\"\u003eResults \u0026amp; Discussion\u003c/span\u003e section, showcasing the obtained results and providing an in-depth discussion, highlighting achievements and improvements. The paper concludes with the fifth section, the conclusion, summarizing the results and offering a final evaluation. Finally, references are included, comprising studies and sources used during the writing of the scientific paper.\u003c/p\u003e"},{"header":"II. Related Work","content":"\u003cp\u003eIn the dynamic landscape of natural language processing (NLP), the years spanning from 2020 to 2023 witnessed significant strides in understanding and harnessing the intricacies of the Arabic language. This comprehensive literature review delves into 13 seminal research papers, providing a nuanced exploration of the models employed, the datasets utilized, and the performance metrics and accuracy measures considered in the realm of Arabic NLP.\u003c/p\u003e\n\u003cp\u003eThe journey commences with the work of Linting Xue et al. \u0026lrm;[4], who introduced the mT5 variation of T5, which was pretrained on a groundbreaking common crawling-based dataset. Their study not only emphasized the cutting-edge performance of mT5 on multilingual benchmarks but also addressed the challenges in zero-shot scenarios. The relevance of the T5 recipe in a multilingual environment was underscored, showcasing its versatility across languages. Mehrdad Farahani et al.\u0026lrm;[5] contributed to the landscape by employing mT5 and the ParsBERT model for refining Persian text summarization. Recognizing the dearth of attention given to this facet of NLP, the researchers introduced two pretrained approaches based on the multilingual T5 model and ParsBERT. Simultaneously, they created and released a summary dataset, providing a valuable resource for further exploration. Remi Calizzano et al. \u0026lrm;[6] navigated the classification of German Facebook comments using GermEval 2021, prioritizing data gathering through data augmentation and task-specific pretraining. Their utilization of 12 datasets for pretraining two multilingual models marked a strategic approach to enhance the classification accuracy.\u003c/p\u003e\n\u003cp\u003eEl Moatez Billah Nagoudi et al. \u0026lrm;[7] presented three potent Arabic text-to-text models, setting the stage for benchmarking with ARGEN\u0026mdash;a compilation of seven tasks from 19 datasets. Their work showcased the superior performance of these models compared to mT5, even in machine translation involving four different foreign languages, underlining the adaptability of these models.\u003c/p\u003e\n\u003cp\u003eAleksandra Chrabrowa et al. \u0026lrm;[8] introduced a new standard for evaluating Polish text-to-text models, emphasizing the need for standardized datasets in summarization and question answering for the Polish language. The introduction of plT5, a general-purpose Polish text-to-text model, further expanded the scope of language-specific models. Ahlam Fuad et al. \u0026lrm;[9] leveraged the mT5 multilingual transformer model to present the first full Arabic generative model for task-oriented dialog systems (AraConv). Their work, anchored by the Arabic Dialog Dataset (Arabic-TOD), showcased a hybrid tuning method for mT5-based transformer models, achieving competitive results in polarity detection and gravity prediction subtasks.\u003c/p\u003e\n\u003cdiv class=\"gridtable\"\u003e\n\u003ctable id=\"Tab1\" border=\"1\"\u003e\u003ccaption\u003e\n\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\n\u003cdiv class=\"CaptionContent\"\u003e\n\u003cp\u003eSome previous studies, datasets and results for the classification task.\u003c/p\u003e\n\u003c/div\u003e\n\u003c/caption\u003e\n\u003ctbody\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e\u003cstrong\u003eRef\u003c/strong\u003e\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e\u003cstrong\u003emodel\u003c/strong\u003e\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e\u003cstrong\u003eDataset\u003c/strong\u003e\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e\u003cstrong\u003ef1 scores\u003c/strong\u003e\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e\u003cstrong\u003eAccuracy\u003c/strong\u003e\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e\u0026lrm;[28]\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003emT5\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eMSA\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e0.70\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e0.72\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e\u0026lrm;[29]\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eAraT5\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eARGEN\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e0.84\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e0.88\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e\u0026lrm;[31]\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eAraBERT (Target/BERT)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eSaaq al-Bambuu\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e0.989\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e0.989\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e\u0026lrm;[32]\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eAraGPT2\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eWikipedia\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e-\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e98%\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003c/tbody\u003e\n\u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003eIn the realm of classification tasks, Table \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e showcased a diverse range of models applied to various datasets, each contributing valuable insights and advancements. Notably, the use of the mT5 model on the Modern Standard Arabic (MSA) dataset, as reported in \u0026lrm;[28], demonstrated moderate yet commendable performance, with an F1-score of 0.70 and an accuracy of 0.72. Moreover, the AraBERT (Target/BERT) model, discussed in \u0026lrm;[31], exhibited exceptional capabilities when applied to the Saaq al-Bambuu dataset, achieving an impressive F1-score and accuracy of 0.989. Additionally, the application of AraGPT2 to the Wikipedia dataset, as detailed in \u0026lrm;[32], resulted in a notable accuracy of 98%, although the F1-score was not explicitly provided. These findings underscore the efficacy of state-of-the-art models in handling classification tasks across diverse linguistic and thematic domains. As delve into the specifics of each model's performance on distinct datasets, it becomes evident that the nuanced nature of the tasks and the characteristics of the data play crucial roles in shaping the effectiveness of these models.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003e2\u003c/strong\u003e shows some previous studies, datasets and results for the generation task.\u003c/p\u003e\n\u003cp\u003e\u003cimg src=\"data:image/png;base64,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\" alt=\"\" width=\"975\" height=\"439\" /\u003e\u003c/p\u003e\n\u003cp\u003eIn the landscape of generation tasks, recent studies have employed diverse models to address the challenges posed by different datasets, showcasing notable advancements in the Table \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e AraBERT (Target/BERT), as examined in \u0026lrm;[31], and demonstrated its effectiveness in the generation task when applied to the Saaq al-Bambuu dataset, achieving a precision (P) of 0.848, a recall (R) of 0.823, and an F1-score of 0.879. The study in \u0026lrm;[33] introduced AraT5, utilizing the XSum and OrangeSum datasets, which showed significant improvements, with R1, R2, and BLEU values of 7.5, 18.30, and an undetermined value, respectively. Furthermore, the application of mT5 to the APGC dataset, as discussed in \u0026lrm;[34], resulted in a precision of 71.6%, an F1-score of 0.820, and a BLEU score of 97.5. The generation task on the Arabic-TOD and Arabic-NTG datasets, as explored by mT5 and AraT5 in \u0026lrm;[35] and \u0026lrm;[36], respectively, revealed noteworthy contributions with BLEU scores of 31.05. These findings underscore the diverse applications of state-of-the-art models in addressing generation tasks across different datasets, highlighting the need for tailored approaches based on the characteristics of the specific data at hand.\u003c/p\u003e"},{"header":"III. Methodology","content":"\u003cp\u003eThe approach to enhancing Arabic NLP transformer models involves a detailed examination of existing fine-tuning methods, focusing on news classification and generation. Diverse datasets of Arabic news are collected and meticulously curated to capture linguistic variations. The Fig.\u0026nbsp;1 to depict processes involving text input and processing through a model mT5, resulting in a labeled and title output\u003c/p\u003e\n\u003cp\u003eComprehensive analysis of the dataset provides insights into its distribution and characteristics. Customized fine-tuning strategies for the multilingual transformer 5 (MT5) model are developed based on dataset insights, addressing language nuances. Evaluation metrics are established to assess model performance in Arabic news processing. The aim is to improve model capabilities and performance for news tasks with greater accuracy and efficiency.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ea. Data collection\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe dataset was collected from the websites Almashhadalaraby and Kaggle, which include a large amount of Arabic news. The dataset is a large collection of Arabic news articles and categories to which the classification has been applied. Here, code was applied to the article, and a portion of the text was extracted as a title to be used in generation because during the collection of news data, titles were not pulled from websites, where there are many linguistic programming tasks. Neural techniques can be implemented, such as text generation, text classification, and text summarization.\u003c/p\u003e\n\u003cdiv class=\"gridtable\"\u003e\n \u003ctable id=\"Tab3\" border=\"1\"\u003e\n \u003ccaption\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003edataset\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eTask\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eTotal samples\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eTraining set\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eValidation set\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eTest set\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eClassification\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e183955\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e105K\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e33K\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e15K\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eGenerations\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e294307\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e50K\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e10K\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e10K\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003e\u003cstrong\u003eb. Pre-Preprocessing\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe titles and the articles are preprocessed such that punctuations are separated from the words and Arabic diacritics are removed. The dataset is divided into training, validation, and test sets of sizes, as shown in Table\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e. The training, validation, and test sets were created to ensure the proper evaluation and training of the models. The datasets shown in Table \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e, encompassing 183,955 samples for the classification task and 294,307 samples for the generation task, are further divided into training, validation, and test sets to facilitate comprehensive training and evaluation of the Multilingual Transformer 5 (MT5) model. The classification task involves categorizing news articles, while the generation task focuses on producing coherent and contextually relevant Arabic news content.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ec. Data Analysis\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe Generation dataset has a total of 29\u0026nbsp;million words composed of 300K unique words. The average title length is 6.36 words, while the average article length is 736 characters. Most words are rare, which means that they are seen fewer than 5 times in the entire corpus. The Classification dataset has a total of 6\u0026nbsp;million words and is composed of 300,000 unique words. The average title length is 9.95 words, while the average article length is 281 characters, as shown in Table \u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e. eight classes and produce word embedding models.\u003c/p\u003e\n\u003cp\u003eThe data contained nine categories (sports, political, financial, artistic, technical, medical, cultural, and religious), as shown in Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e. The graph in Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e shows that circular drawings represent different categories and their proportions. It contains several colorful sections representing categories such as sports, politics, finance, art, technology, medicine, culture, religion and community. The pie chart gives a quick look at the size of each category compared to the next, showing that the Sports category is 32% larger than the pie chart, followed by political, financial, artistic, technical, medical, cultural, religious, and community news. Through comparative analysis with other researchers, the methods utilized for both classification and generation exhibited promising results across various metrics. These favorable outcomes emphasize the effectiveness of the selected methodologies in tackling the challenges associated with news classification and generation. Nonetheless, it is essential to recognize a notable obstacle encountered during insufficient research resources for additional model refinement. Despite the achievement of noteworthy results, the potential for enhancing model accuracy remains constrained by limited resources. This constraint is particularly pronounced when considering the demanding nature of classification and generation tasks, especially within the domain of Arabic news. The intricate linguistic nuances and contextual intricacies in Arabic news present unique challenges that necessitate sophisticated models for optimal performance. In light of these challenges, recommendations for future work emphasize the utilization of the MT5 model. While acknowledging its resource-intensive nature, the MT5 model holds promise for significantly improving the classification and generation processes in the context of Arabic news. The recommendation stems from the understanding that, despite the associated resource demands, leveraging advanced models such as mT5 is instrumental in pushing the boundaries of accuracy and effectiveness. In conclusion, the results of comparisons with other researchers highlight the success of current techniques used in classification and generation. However, the identified resource constraint underscores the importance of continued efforts to secure the necessary resources for further model refinement. The proposed adoption of the MT5 model represents a strategic move, acknowledging its resource-intensive nature while emphasizing its potential to improve the performance of classification and generation tasks in the challenging domain of Arabic news.\u003c/p\u003e\n\u003cdiv class=\"gridtable\"\u003e\n \u003ctable id=\"Tab4\" border=\"1\"\u003e\n \u003ccaption\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eAnalytical Exploration\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eComparison\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eGeneration dataset\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eClassification dataset\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eNumber of words in the file\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2937143 word\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e6000000 word\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eNumber of unique words\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e300000 word\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e300000 word\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAverage title length\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e6.36 Letter\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e9.95 Letter\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAverage article length\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e736 Letter\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e281 Letter\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eDataset\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e294307\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e183955\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eNumber of categories\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e8\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003eTable\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e presents a concise summary of two datasets: \"Generation\" and \"Classification.\" It includes key metrics such as word count, unique words, average title length, average article length, and the number of categories. These metrics provide valuable insights into the size, vocabulary, and structure of the datasets, enabling researchers to better understand and analyze the data.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e1) Calculating the number of words\u003c/strong\u003e:\u003c/p\u003e\n\u003cp\u003eNumber of words = text length/word length\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2) Counting the number of unique words\u003c/strong\u003e:\u003c/p\u003e\n\u003cp\u003eNumber of unique words = number of words - number of frequent words\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3) Calculating average word length\u003c/strong\u003e:\u003c/p\u003e\n\u003cp\u003eAverage word length = sum of word lengths/number of words\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4) Calculating the average article length\u003c/strong\u003e:\u003c/p\u003e\n\u003cp\u003eAverage article length = sum of article lengths/number of articles\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ed. Model Fine-Tuning\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eModel fine-tuning entailed the use of monolingual datasets sourced from various news outlets for training and assessing the models. The news datasets, comprising Arabic articles, were gathered from diverse online sources and preprocessed by segregating punctuation from words and eliminating Arabic diacritics. A unified text-to-text modeling framework is adopted to establish a baseline model for a range of tasks supported by a dataset. Specifically, mT5-base ‎[1]was fine-tuned for each task in a single-task setting. mT5 is a large multilingual pretrained language model (PLM) based on the English-only T5 model ‎[10]and consists of an encoder-decoder transformer architecture ‎[11]. However, earlier PLMs consisted of either a standalone encoder for classification and representation tasks ‎[17] or a decoder for generative tasks ‎[18], and the architecture of mT5 offers two major benefits. First, the complexity of multitask NLP pipelines can be reduced by removing the need for specialized task-specific frameworks. Second, it may also lead to improved performance since sharing a unified set of parameters between related tasks can be beneficial ‎[19]. For each of the tasks, a distinct natural-language prefix that is intended to aid the model in learning the task is selected (e.g., “summaries:” for bullet-point summary generation, “generate title:” for headline generation).\u003c/p\u003e\n\u003cp\u003eIn the single-task setting, mT5 is fine-tuned on each of the tasks in isolation, resulting in one model per task. The training of the Seq2Seq models was carried out on the training set, while model selection and hyperparameter tuning were performed on the validation set. A consistent experimental setup was maintained to fine-tune and assess the models, adopting an approach akin to subtask C with adjustments tailored for the Arabic language. The hidden size and embedding size were both set at 256. The key hyperparameters included a learning rate of 4e-5, a batch size of 8, 5 epochs, and a weight decay of 0.01. Additionally, FP16 precision was assessed using the fp16 = true setting in Seq2SeqTrainingArguments.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ee. Evaluation Methods\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eLikewise, most TS evaluation approaches have been driven by other similar NLP research areas. Various evaluation methods have been applied across studies to measure the three main aspects of the newly generated text. These aspects are:\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ei) Fluency\u003c/strong\u003e: Ensuring grammatical accuracy and coherence in the generated text, maintaining syntactic consistency, and creating a natural flow of language.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eii)Adequacy\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003ePreserving the intended meaning of the input during generation, capturing the essence of the message, and avoiding distortions or misinterpretations.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eiii) Simplicity\u003c/strong\u003e: Enhancing readability and comprehension by using clear and concise language, avoiding overly complex sentence structures, and organizing information in a straightforward manner.\u003c/p\u003e\n\u003cul\u003e\n \u003cli\u003e\n \u003cp\u003e\u003cstrong\u003eClassification approach – automatic evaluation\u003c/strong\u003e\u003c/p\u003e\n \u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eTo evaluate the accuracy of the classification approach, we used a standard set of evaluation metrics, including precision, recall, and F1-score ‎[23]. The experimental setup included the use of monolingual datasets from multiple news sources, which were preprocessed by separating punctuation from words and removing Arabic diacritics. The classification approach was evaluated using standard metrics such as precision, recall, and F1-score. The models were trained using diverse Arabic news datasets obtained from online sources. After preprocessing, the obtained dataset was used to train and evaluate the models for the classification approach.\u003c/p\u003e\n\u003cul\u003e\n \u003cli\u003e\n \u003cp\u003e\u003cstrong\u003eEnhanced Computational Approaches\u003c/strong\u003e\u003c/p\u003e\n \u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eGiven an input text sequence 𝑥, the classification model can be considered a function, i.e.,\u003c/p\u003e\n\u003cp\u003e𝑓(𝑥) = 𝐶 for measuring the conditional probability distributions over all possible labels in the predefined category set 𝐶 = {𝑐\u003csub\u003e1\u003c/sub\u003e, 𝑐\u003csub\u003e2\u003c/sub\u003e, 𝑐\u003csub\u003e3\u003c/sub\u003e..., 𝑐\u003csub\u003e𝐿\u003c/sub\u003e}.\u003c/p\u003e\n\u003cp\u003eText classification works on an instance of space 𝒳 where each instance is an input text sequence 𝑥. Hence, let 𝑋 = {𝑥\u003csub\u003e1\u003c/sub\u003e, 𝑥\u003csub\u003e2\u003c/sub\u003e, 𝑥\u003csub\u003e3\u003c/sub\u003e..., 𝑥\u003csub\u003e𝑁\u003c/sub\u003e } be the training set, and each input text sequence 𝑥\u003csub\u003e𝑖\u003c/sub\u003e = {𝑡\u003csub\u003e1\u003c/sub\u003e, 𝑡\u003csub\u003e2\u003c/sub\u003e, 𝑡\u003csub\u003e3\u003c/sub\u003e..., 𝑡𝑃} represents a sequence of tokens. 𝑌 = {𝑦\u003csub\u003e1\u003c/sub\u003e, 𝑦\u003csub\u003e2\u003c/sub\u003e, 𝑦\u003csub\u003e3\u003c/sub\u003e, ..., 𝑦\u003csub\u003e𝑁\u003c/sub\u003e } the set of categories in which each document 𝑥\u003csub\u003e𝑖\u003c/sub\u003e is classified, each y\u003csub\u003e𝑖\u003c/sub\u003e ∈ 𝐶. The model was trained for the set of parameters 𝜃 ∈ \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({IR}^{q}\\)\u003c/span\u003e\u003c/span\u003e minimizing the cross-entropy loss in Equation (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(4\\)\u003c/span\u003e\u003c/span\u003e):\u003c/p\u003e\n\u003cdiv id=\"Equ1\" class=\"Equation\"\u003e\n \u003cdiv id=\"FileID_Equ1\" class=\"mathdisplay\"\u003e$${\\theta }{=}_{{\\theta }\\in {\\text{I}\\text{R}}^{\\text{q}}} \\text{F}\\left({\\theta }, \\text{X}\\right)$$\u003c/div\u003e\n \u003cdiv class=\"EquationNumber\"\u003e1\u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Equ2\" class=\"Equation\"\u003e\n \u003cdiv id=\"FileID_Equ2\" class=\"mathdisplay\"\u003e$$\\text{F}\\left({\\theta }, \\text{X}\\right)= - \\sum _{i=1}^{N}\\sum _{c=1}^{L}{{\\Upsilon }}_{{x}_{i|c}} \\text{l}\\text{o}\\text{g}\\left({\\text{P}}_{{x}_{i|c}} \\left({\\theta }\\right)\\right)$$\u003c/div\u003e\n \u003cdiv class=\"EquationNumber\"\u003e2\u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Equ3\" class=\"Equation\"\u003e\n \u003cdiv id=\"FileID_Equ3\" class=\"mathdisplay\"\u003e$${{\\Upsilon }}_{{x}_{i|c}}=\\left\\{\\begin{array}{c}1 if {\\text{y}}_{\\text{i}} = c\\\\ 0 otherwise \\end{array}\\right.$$\u003c/div\u003e\n \u003cdiv class=\"EquationNumber\"\u003e3\u003c/div\u003e\n\u003c/div\u003e\n\u003cp\u003ewhere \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({\\varvec{P}}_{{x}_{i|c}} \\left(\\varvec{\\theta }\\right)\\)\u003c/span\u003e\u003c/span\u003eis the predicted conditional probability of 𝑥\u003csub\u003e𝑖\u003c/sub\u003e giving class C. To calculate the F1 score, the equations below were used. TP refers to ‘true positive’, which counts the tokens appearing in both the prediction and the ground truth. FP refers to a ‘false positive’ that counts the tokens that appear only in the prediction. FN refers to ‘False Negative’, which means the number of tokens that appear only in the ground truth. The F1 score of the dataset is an average of the F1 score of every sample, and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(F1\\)\u003c/span\u003e\u003c/span\u003eᵢ is the individual value of F1 calculated for each i from 1 to N.\u003c/p\u003e\n\u003cdiv id=\"Equ4\" class=\"Equation\"\u003e\n \u003cdiv id=\"FileID_Equ4\" class=\"mathdisplay\"\u003e$$precision=\\frac{TP}{TP+FP}$$\u003c/div\u003e\n \u003cdiv class=\"EquationNumber\"\u003e4\u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Equ5\" class=\"Equation\"\u003e\n \u003cdiv id=\"FileID_Equ5\" class=\"mathdisplay\"\u003e$$F1=\\frac{2 \\times precision \\times Recall}{precision+Recall}$$\u003c/div\u003e\n \u003cdiv class=\"EquationNumber\"\u003e5\u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Equ6\" class=\"Equation\"\u003e\n \u003cdiv id=\"FileID_Equ6\" class=\"mathdisplay\"\u003e$$F1=\\frac{\\sum _{i=1}^{N}{F1}_{i }}{N}$$\u003c/div\u003e\n \u003cdiv class=\"EquationNumber\"\u003e6\u003c/div\u003e\n\u003c/div\u003e\n\u003cul\u003e\n \u003cli\u003e\n \u003cp\u003e\u003cstrong\u003eGenerative approach - automatic evaluation\u003c/strong\u003e\u003c/p\u003e\n \u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eThe primary evaluation metric used for the generation task is ROUGE, which measures recall ‎[14]. ROUGE indicates how much of the human-produced titles appear in the model-generated titles. Moreover, the BLEU is also used as a secondary metric to measure precision ‎[15]. BLEU [8], an n-gram-based evaluation metric. These methods were not able to capture two main simplification features:\u003c/p\u003e\n\u003cp\u003e1) Changing word order as a paraphrasing simplification method.\u003c/p\u003e\n\u003cp\u003e2) Maintaining the deep structure meaning despite changes in the surface form structure. Specifically, how much of the model-generated titles appear in the human-produced titles?\u003c/p\u003e\n\u003cp\u003eTo compute attention, one typically scores a hidden state of the decoder and a corresponding state of the encoder. The attention ratings are then normalized over all encoder states, resulting in an attention score of 1, as shown in Eq.\u0026nbsp;(7).\u003c/p\u003e\n\u003cp\u003e\u003cimg src=\"data:image/png;base64,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\" width=\"771\" height=\"213\"\u003e\u003cbr\u003e\u003c/p\u003e"},{"header":"IV. RESULTS \u0026 DISCUSSION","content":"\u003cp\u003e\u003cstrong\u003ea. Classification Results\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe classification approach yielded promising results, with precision and F1-score values indicating the effectiveness of the models in accurately categorizing Arabic news articles. The models exhibited strong performance across the diverse categories present in the dataset, showcasing their ability to handle multiclass classification tasks effectively. These results affirm the suitability of transfer learning and text-to-text models in addressing classification challenges within Arabic NLP, thereby contributing to the advancement of language-specific natural language processing techniques.\u003c/p\u003e\n\u003cdiv class=\"gridtable\"\u003e\n\u003ctable id=\"Tab5\" border=\"1\"\u003e\u003ccaption\u003e\n\u003cdiv class=\"CaptionNumber\"\u003eTable 5\u003c/div\u003e\n\u003cdiv class=\"CaptionContent\"\u003e\n\u003cp\u003eShows the evaluation of mT5 model\u003c/p\u003e\n\u003c/div\u003e\n\u003c/caption\u003e\n\u003cthead\u003e\n\u003ctr\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eModel\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eAccuracy\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eF1 Score\u003c/p\u003e\n\u003c/th\u003e\n\u003c/tr\u003e\n\u003c/thead\u003e\n\u003ctbody\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003emT5small\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.7858\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.7858\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003emT5base\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.8230\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.8230\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003emT5large\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.8742\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.8742\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003c/tbody\u003e\n\u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003eTable\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003e presents the results of the evaluation dataset for the mT5 model. These metrics, including accuracy and F1 score, provide a comprehensive evaluation of the mT5 model's performance across different sizes (small, base, and large). The results showcase the model's proficiency in handling Arabic natural language processing tasks.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eb. Generative Results\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe generative approach produced compelling outcomes, as evidenced by the automatic evaluation metrics. The models demonstrated low perplexity, indicative of their proficient generation of coherent and contextually relevant Arabic text. Furthermore, the high BLEU and ROUGE scores underscored the ability of the models to generate outputs that closely align with reference summaries and maintain semantic coherence. These results emphasize the potential of transfer learning and text-to-text models in facilitating the generation of high-quality Arabic text, thereby enriching the landscape of natural language generation in the language domain.\u003c/p\u003e\n\u003cdiv class=\"gridtable\"\u003e\n\u003ctable id=\"Tab6\" border=\"1\"\u003e\u003ccaption\u003e\n\u003cdiv class=\"CaptionNumber\"\u003eTable 6\u003c/div\u003e\n\u003cdiv class=\"CaptionContent\"\u003e\n\u003cp\u003eShows the results of generative approach\u003c/p\u003e\n\u003c/div\u003e\n\u003c/caption\u003e\n\u003cthead\u003e\n\u003ctr\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eTask\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eRouge-1\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eRouge-2\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eRouge-L\u003c/p\u003e\n\u003c/th\u003e\n\u003c/tr\u003e\n\u003c/thead\u003e\n\u003ctbody\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e\u003cem\u003eGenerative of Headlines\u003c/em\u003e\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e\u003cstrong\u003e9.719\u003c/strong\u003e\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e\u003cstrong\u003e1.575\u003c/strong\u003e\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e\u003cstrong\u003e9.315\u003c/strong\u003e\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003c/tbody\u003e\n\u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003eThe outcomes, as depicted in Table \u003cspan class=\"InternalRef\"\u003e6\u003c/span\u003e, illustrate the performance of the generative approach for the specified task of headline generation. The evaluation metrics include the Rouge-1, Rouge-2, and Rouge-L scores, providing insights into the model's effectiveness in generating coherent and contextually relevant Arabic news headlines. The results indicate notable scores, with Rouge-1 at 9.719, Rouge-2 at 1.575, and Rouge-L at 9.315, confirming the proficiency of the generative model in addressing the complexities of Arabic natural language processing tasks.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ec. Discussion\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe presented discussion delves into the comprehensive evaluation of the mT5 model across various scales and its application in a generative task, shedding light on its performance metrics and implications for advancing Arabic natural language processing (NLP). The findings are detailed in Table \u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003e and Table \u003cspan class=\"InternalRef\"\u003e6\u003c/span\u003e, providing a nuanced understanding of the model's capabilities in both classification and generation tasks.\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003eComparison Analysis of Classification Results\u003c/strong\u003e:\u003c/p\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eAfter comparing the results from Table \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e, which show findings from previous studies, with those presented in Table \u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003e for the mT5 model, the following findings emerged:\u003c/p\u003e\n\u003col\u003e\n\u003cli\u003e\n\u003cp\u003eFor the \"MSA\" dataset, the mT5 model achieved a competitive accuracy of 0.72, comparable to that of the mT5 model in \u0026lrm;[28].\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eCompared with AraT5 on the \"ARGEN\" dataset, the mT5base model demonstrates similar accuracy but a slightly lower F1-score.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eAraBERT (Target/BERT) achieves outstanding results on the \"Saaq al-Bambuu\" dataset, with extremely high F1-scores and accuracies. The mT5base model, while achieving lower scores, still performs well in this context.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eFor the \"Wikipedia\" dataset, AraGPT2 attains an impressive accuracy of 98%, whereas mT5 models show competitive results, especially mT5large, with an accuracy of 87%.\u003c/p\u003e\n\u003c/li\u003e\n\u003c/ol\u003e\n\u003cul\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003eComparison Analysis of the Generative Results\u003c/strong\u003e\u003c/p\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eWhile the descriptive analysis provides insights into the performance of mT5 models across various datasets, the generative approach's results in Table \u003cspan class=\"InternalRef\"\u003e6\u003c/span\u003e focus on the headline generation task and Rouge scores, limiting direct comparisons for the mentioned datasets.\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003ePerformance Metrics of the mT5 Model\u003c/strong\u003e\u003c/p\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eThe mT5 model, in its different variants\u0026mdash;mT5small, mT5base, and mT5large\u0026mdash;underwent evaluation based on accuracy and F1 score across diverse dataset sizes. The accuracy measures the overall correctness of the model's predictions, while the F1 score reflects the balance between precision and recall. The results indicate a progressive improvement in performance with increasing model size. Notably, mT5large outperforms mT5base and mT5small, demonstrating the scalability and efficiency of the model across a range of complexities within the Arabic language. This scalability is crucial in real-world applications where the complexity and size of the datasets can vary. The mT5 model's adaptability across different scales suggests its potential to handle diverse linguistic challenges encountered in Arabic NLP.\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003eGenerative Task Results\u003c/strong\u003e\u003c/p\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eTable\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e6\u003c/span\u003e focuses on the outcomes of a generative task, specifically the generation of headlines, assessed through Rouge-1, Rouge-2, and Rouge-L scores. Rouge metrics evaluate the overlap and similarity between generated and reference texts, providing insights into the quality of the generated content. The Rouge-1, Rouge-2, and Rouge-L scores of 9.719, 1.575, and 9.315, respectively, suggest a notable proficiency in generating headlines that exhibit semantic coherence and similarity to reference headlines. The ability of the mT5 model to perform well in generative tasks underscores its versatility in understanding and reproducing nuanced language structures in Arabic. This is particularly significant in applications requiring content creation, such as summarization, headline generation, and paraphrasing.\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003eTransfer Learning and Text-to-Text Models in Arabic NLP\u003c/strong\u003e\u003c/p\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eThe combined results underscore the effectiveness of transfer learning and the text-to-text paradigm in advancing Arabic NLP. Transfer learning enables the model to leverage knowledge from pretraining tasks, showcasing its ability to generalize well across diverse linguistic challenges. The text-to-text approach, where all NLP tasks are framed as text generation tasks, proves to be a powerful and unifying paradigm for various tasks. The robust performance of the mT5 model in both classification and generation tasks suggests its potential for a wide array of applications within Arabic NLP, including text classification, generation, simplification, and sentiment analysis. These findings contribute valuable insights into the development of models tailored for the Arabic language, paving the way for innovative applications and deeper understanding within this domain. In conclusion, the discussion emphasizes the pivotal role of transfer learning and text-to-text models in improving Arabic NLP, providing evidence of their impact on performance across different tasks. The presented results open avenues for future research and development, encouraging the exploration of novel applications and advancements in the field of Arabic language processing.\u003c/p\u003e"},{"header":"V. Conclusion","content":"\u003cp\u003eThe results from both the generative and classification approaches highlight the effectiveness of employing transfer learning and text-to-text models to advance Arabic natural language processing. The robust performance observed in classification and generation tasks underscores the adaptability and efficacy of these models in addressing diverse language-specific challenges. Notably, the best results mentioned in the initial tables include an impressive Rouge-1 score of 9.719 and an accuracy of 0.8742 in the classification task using the mT5large model. These outcomes affirm the significance of leveraging such models to drive progress in Arabic NLP, offering promising applications in text classification, generation, simplification, and sentiment analysis. As the field of natural language processing evolves, the utilization of these models is likely to contribute significantly to the development of cutting-edge solutions and advancements in Arabic NLP. It is imperative to acknowledge the pivotal role played by high-quality corpora and specialized tools in data collection and annotation, ensuring the accuracy of the classification and generation processes.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eMarie-Sainte, S. L., Alalyani, N., Alotaibi, S., Ghouzali, S., \u0026amp; Abunadi, I. (2018). Arabic natural language processing and machine learning-based systems. IEEE Access, 7, 7011-7020.\u003cspan dir=\"RTL\"\u003e\u0026rlm;\u003c/span\u003e\u003c/li\u003e\n\u003cli\u003eAlmomani, A., Al-Nawasrah, A., Alauthman, M., Al-Betar, M. A., \u0026amp; Meziane, F. (2021). Botnet detection used fast-flux technique, based on adaptive dynamic evolving spiking neural network algorithm. International Journal of Ad Hoc and Ubiquitous Computing, 36(1), 50-65.\u003cspan dir=\"RTL\"\u003e\u0026rlm;\u003c/span\u003e\u003c/li\u003e\n\u003cli\u003eAlkhurayyif, Y., \u0026amp; Sait, A. R. W. (2023). Developing an Open Domain Arabic Question Answering System Using a Deep Learning Technique. IEEE Access.\u003cspan dir=\"RTL\"\u003e\u0026rlm;\u003c/span\u003e\u003c/li\u003e\n\u003cli\u003eXue, L., Constant, N., Roberts, A., Kale, M., Al-Rfou, R., Siddhant, A., ... \u0026amp; Raffel, C. (2020). mT5: A massively multilingual pretrained text-to-text transformer. arXiv preprint arXiv:2010.11934.\u003cspan dir=\"RTL\"\u003e\u0026rlm;\u003c/span\u003e\u003c/li\u003e\n\u003cli\u003eFarahani, M., Gharachorloo, M., \u0026amp; Manthouri, M. (2021, March). Leveraging ParsBERT and pretrained mT5 for Persian abstractive text summarization. In 2021 26th International Computer Conference, Computer Society of Iran (CSICC) (pp. 1-6). IEEE.\u003cspan dir=\"RTL\"\u003e\u0026rlm;\u003c/span\u003e\u003c/li\u003e\n\u003cli\u003eCalizzano, R., Ostendorff, M., \u0026amp; Rehm, G. (2021, September). DFKI SLT at GermEval 2021: Multilingual Pretraining and Data Augmentation for the Classification of Toxicity in Social Media Comments. In Proceedings of the GermEval 2021 Shared Task on the Identification of Toxic, Engaging, and Fact-Claiming Comments (pp. 25-31).\u003cspan dir=\"RTL\"\u003e\u0026rlm;\u003c/span\u003e\u003c/li\u003e\n\u003cli\u003eNagoudi, E. M. B., Elmadany, A., \u0026amp; Abdul-Mageed, M. (2021). AraT5: Text-to-text transformers for Arabic language generation. arXiv preprint arXiv:2109.12068.\u003cspan dir=\"RTL\"\u003e\u0026rlm;\u003c/span\u003e\u003c/li\u003e\n\u003cli\u003eChrabrowa, A., Dragan, Ł., Grzegorczyk, K., Kajtoch, D., Koszowski, M., Mroczkowski, R., \u0026amp; Rybak, P. (2022). Evaluation of transfer learning for polish with a text-to-text model. arXiv preprint arXiv:2205.08808.\u003cspan dir=\"RTL\"\u003e\u0026rlm;\u003c/span\u003e\u003c/li\u003e\n\u003cli\u003eFuad, A., \u0026amp; Al-Yahya, M. (2022). AraConv: Developing an Arabic task-oriented dialog system using multilingual transformer model mT5. Applied Sciences, 12(4), 1881.\u003cspan dir=\"RTL\"\u003e\u0026rlm;\u003c/span\u003e \u003c/li\u003e\n\u003cli\u003eRaffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., ... \u0026amp; Liu, P. J. (2020). Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1), 5485-5551.\u003cspan dir=\"RTL\"\u003e\u0026rlm;\u003c/span\u003e\u003c/li\u003e\n\u003cli\u003eToledano-L\u0026oacute;pez, O. G., Madera, J., Gonz\u0026aacute;lez, H., Sim\u0026oacute;n-Cuevas, A., Demeester, T., \u0026amp; Mannens, E. (2022). Finetuning mt5-based transformer via cma-es for sentiment analysis. In Proceedings of the Third Workshop for Iberian Languages Evaluation Forum (IberLEF 2022), CEUR WS Proceedings.\u003c/li\u003e\n\u003cli\u003eAlrowili, S., \u0026amp; Vijay-Shanker, K. (2022, December). Generative approach for gender-rewriting task with ArabicT5. In Proceedings of the The Seventh Arabic Natural Language Processing Workshop (WANLP) (pp. 491-495).\u003cspan dir=\"RTL\"\u003e\u0026rlm;\u003c/span\u003e\u003c/li\u003e\n\u003cli\u003eKale, M., Siddhant, A., Constant, N., Johnson, M., Al-Rfou, R., \u0026amp; Xue, L. (2021). nmT5--Is parallel data still relevant for pretraining massively multilingual language models?. arXiv preprint arXiv:2106.02171.\u003cspan dir=\"RTL\"\u003e\u0026rlm;\u003c/span\u003e\u003c/li\u003e\n\u003cli\u003eKhallaf, N., \u0026amp; Sharoff, S. (2022). Toward arabic sentence simplification via classification and generative approaches. arXiv preprint arXiv:2204.09292.\u003cspan dir=\"RTL\"\u003e\u0026rlm;\u003c/span\u003e \u003c/li\u003e\n\u003cli\u003eFuadi, M., Wibawa, A. D., \u0026amp; Sumpeno, S. (2023). idT5: Indonesian Version of Multilingual T5 Transformer. arXiv preprint arXiv:2302.00856.\u003cspan dir=\"RTL\"\u003e\u0026rlm;\u003c/span\u003e \u003c/li\u003e\n\u003cli\u003e\u0026Scaron;tajner, S., Sheang, K. C., \u0026amp; Saggion, H. (2022, June). Sentence simplification capabilities of transfer-based models. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 36, No. 11, pp. 12172-12180).\u003cspan dir=\"RTL\"\u003e\u0026rlm;\u003c/span\u003e\u003c/li\u003e\n\u003cli\u003eCageggi, G., Di Rosa, E., \u0026amp; Uboldi, A. (2023). App2Check at EMit: Large Language Models for Multilabel Emotion Classification.\u003cspan dir=\"RTL\"\u003e\u0026rlm;\u003c/span\u003e \u003c/li\u003e\n\u003cli\u003eLin, C. Y. (2004, July). Rouge: A package for automatic evaluation of summaries. In Text summarization branches out (pp. 74-81).\u003cspan dir=\"RTL\"\u003e\u0026rlm;\u003c/span\u003e\u003c/li\u003e\n\u003cli\u003ePapineni, K., Roukos, S., Ward, T., \u0026amp; Zhu, W. J. (2002, July). Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics (pp. 311-318).\u003cspan dir=\"RTL\"\u003e\u0026rlm;\u003c/span\u003e \u003c/li\u003e\n\u003cli\u003eRaffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., ... \u0026amp; Liu, P. J. (2020). Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1), 5485-5551.\u003cspan dir=\"RTL\"\u003e\u0026rlm;\u003c/span\u003e \u003c/li\u003e\n\u003cli\u003eDevlin, J., Chang, M. W., Lee, K., \u0026amp; Toutanova, K. (2018). Bert: Pretraining of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805.\u003cspan dir=\"RTL\"\u003e\u0026rlm;\u003c/span\u003e \u003c/li\u003e\n\u003cli\u003eRadford, A., Wu, J., Child, R., Luan, D., Amodei, D.,\u0026amp; Sutskever,I.(2019). Language models are unsupervised multitask learners. OpenAI blog, 1(8),9.\u003cspan dir=\"RTL\"\u003e\u0026rlm;\u003c/span\u003e \u003c/li\u003e\n\u003cli\u003eRuder, S. (2017). An overview of multitask learning in deep neural networks. arXiv preprint arXiv:1706.05098.\u003cspan dir=\"RTL\"\u003e\u0026rlm;\u003c/span\u003e \u003c/li\u003e\n\u003cli\u003ePapineni, K., Roukos, S., Ward, T., \u0026amp; Zhu, W. J. (2002, July). Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics (pp. 311-318).\u003cspan dir=\"RTL\"\u003e\u0026rlm;\u003c/span\u003e \u003c/li\u003e\n\u003cli\u003eXu, W., Napoles, C., Pavlick, E., Chen, Q., \u0026amp; Callison-Burch, C. (2016). Optimizing statistical machine translation for text simplification. Transactions of the Association for Computational Linguistics, 4, 401-415.\u003cspan dir=\"RTL\"\u003e\u0026rlm;\u003c/span\u003e \u003c/li\u003e\n\u003cli\u003eKhorsheed, M. S., \u0026amp; Al-Thubaity, A. O. (2013). Comparative evaluation of text classification techniques using a large diverse Arabic dataset. Language resources and evaluation, 47, 513-538.\u003cspan dir=\"RTL\"\u003e\u0026rlm;\u003c/span\u003e\u003c/li\u003e\n\u003cli\u003eAlrowili, S., \u0026amp; Vijay-Shanker, K. (2022, December). Generative approach for gender-rewriting task with ArabicT5. In Proceedings of the The Seventh Arabic Natural Language Processing Workshop (WANLP) (pp. 491-495).\u003cspan dir=\"RTL\"\u003e\u0026rlm;\u003c/span\u003e \u003c/li\u003e\n\u003cli\u003eKhallaf, N., \u0026amp; Sharoff, S. (2022). Toward arabic sentence simplification via classification and generative approaches. arXiv preprint arXiv:2204.09292.\u003cspan dir=\"RTL\"\u003e\u0026rlm;\u003c/span\u003e \u003c/li\u003e\n\u003cli\u003eNagoudi, E. M. B., Elmadany, A., \u0026amp; Abdul-Mageed, M. (2021). AraT5: Text-to-text transformers for Arabic language generation. arXiv preprint arXiv:2109.12068.\u003cspan dir=\"RTL\"\u003e\u0026rlm;\u003c/span\u003e \u003c/li\u003e\n\u003cli\u003eLuong, T., Brevdo, E., \u0026amp; Zhao, R. (2017). Neural machine translation (seq2seq) tutorial. 2017. URL: https://www. TensorFlow. org/tutorials/seq2seq (дата обращения 17.02. 2018).\u003cspan dir=\"RTL\"\u003e\u0026rlm;\u003c/span\u003e \u003c/li\u003e\n\u003cli\u003eKhallaf, N., \u0026amp; Sharoff, S. (2022). Toward arabic sentence simplification via classification and generative approaches. arXiv preprint arXiv:2204.09292.\u003cspan dir=\"RTL\"\u003e\u0026rlm;\u003c/span\u003e \u003c/li\u003e\n\u003cli\u003eAntoun, W., Baly, F., \u0026amp; Hajj, H. (2020). AraGPT2: Pretrained transformer for Arabic language generation. arXiv preprint arXiv:2012.15520.\u003cspan dir=\"RTL\"\u003e\u0026rlm;\u003c/span\u003e\u003c/li\u003e\n\u003cli\u003eZmandar, N., El-Haj, M., \u0026amp; Rayson, P. (2023, September). FinAraT5: A text to text model for financial Arabic text understanding and generation. In Proceedings of the 4th Conference on Language, Data and Knowledge (pp. 262-273).\u003cspan dir=\"RTL\"\u003e\u0026rlm;\u003c/span\u003e \u003c/li\u003e\n\u003cli\u003eAlrowili, S., \u0026amp; Vijay-Shanker, K. (2022, December). Generative approach for gender-rewriting task with ArabicT5. In Proceedings of the The Seventh Arabic Natural Language Processing Workshop (WANLP) (pp. 491-495).\u003cspan dir=\"RTL\"\u003e\u0026rlm;\u003c/span\u003e\u003c/li\u003e\n\u003cli\u003eFuad, A., \u0026amp; Al-Yahya, M. (2022). AraConv: Developing an Arabic task-oriented dialog system using multilingual transformer model mT5. Applied Sciences, 12(4), 1881.\u003cspan dir=\"RTL\"\u003e\u0026rlm;\u003c/span\u003e \u003c/li\u003e\n\u003cli\u003eElmadany, A., Nagoudi, E. M. B., \u0026amp; Abdul-Mageed, M. (2023). Octopus: A Multitask Model and Toolkit for Arabic Natural Language Generation. arXiv preprint arXiv:2310.16127.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"Ibb University","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":"Transformer, Decoder, Encoder, Title Generation and Classification, MT5, Arabic NLP, Fine-tuning","lastPublishedDoi":"10.21203/rs.3.rs-3982909/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-3982909/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eMultilingual Transformer 5 (MT5) is a versatile architecture in natural language processing (NLP) that demonstrates proficiency across various languages. This study aimed to improve the performance of the MT5 model in two key tasks: topic classification and headline generation. The datasets used were 183K and 294K samples. The classification task involved categorizing news articles, while the news generation task aimed to create coherent and contextually relevant Arabic news content. Through careful fine-tuning and rigorous evaluation, the MT5 model significantly advances its ability to address complex challenges in Arabic NLP. This study provides practical insights into real-world applications in processing Arab news. The performance of the MT5 model was evaluated using various online platforms. The mT5small model achieved an accuracy of 0.7858 and an F1 score of 0.7858, while the mT5base model achieved an accuracy of 0.8230 and an F1 score of 0.8230. The generative approach for headline generation yielded Rouge-1, Rouge-2, and Rouge-L scores under the task \"Generative of Headlines.\" These outcomes demonstrate the effectiveness of the fine-tuned MT5 model across various evaluation metrics and tasks, confirming its potential for practical applications in Arabic NLP.\u003c/p\u003e","manuscriptTitle":"Classification and Generation of Arabic News Titles from Raw Text Based on an Encoder-Decoder Transformer Model (mT5)","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-02-26 08:52:02","doi":"10.21203/rs.3.rs-3982909/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":"6a8e11d3-c627-4127-8af0-809f176f77f8","owner":[],"postedDate":"February 26th, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":28966800,"name":"Artificial Intelligence and Machine Learning"}],"tags":[],"updatedAt":"2024-02-26T08:52:03+00:00","versionOfRecord":[],"versionCreatedAt":"2024-02-26 08:52:02","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-3982909","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-3982909","identity":"rs-3982909","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}