•

The dataset brings together more than 70,000 comments in the Bengali language, making it one of the most significant available public sources. About 50,000 comments have been sourced from Kaggle, while the other 20,000 comments are collected from various online Bengali text sources, providing an extensive mix of content.

Various researchers have applied machine learning (ML) and deep learning (DL) techniques to combat cyberbullying, often relying on small, domain-specific datasets [1]. LIME has been widely used for model interpretability. For instance, [2] used LIME with classifiers like SVM, RF, Bi-LSTM, and BERT, offering interpretability but falling short in model-specific explanations. Similarly, [3] and [4] incorporated LIME in hate speech detection but encountered limitations in capturing individual model behaviors. In sentiment analysis, particularly for Bengali, [5] existing research has shown promising results but remains heavily dependent on limited datasets. [6] used FastText and Skipgram for feature extraction, highlighting the robustness of BanglaBERT. However, the study emphasized the need for domain adaptation and multi-class learning. [7] proposed a hybrid ensemble model achieving 85% accuracy, yet it lacked transformer-based architectures crucial for capturing deeper language features. Similarly, [8] demonstrated the superiority of transformer-based embeddings such as BERT and GPT-J-6B over traditional methods like TF-IDF and Word2Vec, reaching up to 98% accuracy. Multilingual efforts like those by [9], using WordPiece tokenization with AfriBERTa and RoBERTa, faced challenges of overfitting and data quality, despite achieving 80.88% accuracy. In Bengali-specific settings, [10] applied BERT with 97% accuracy but was limited by a small and imbalanced dataset. Likewise, [11] integrated TF-IDF with BERT, CNN, and MLP, achieving 92.3% accuracy, though the dataset size raised concerns about overfitting. [12] experimented with pre-trained models like BERT, XLM-R, and mBERT, finding BanglaBERT to be the most effective. These studies underline the necessity of fine-tuned transformer models for Bengali, but again highlight the consistent issue of limited, non-diverse datasets. Hybrid and ensemble models are also gaining attention; [13] reached 87.54% using DistilBERT and XLM-R. However, dataset limitations continue to restrict broader applicability and generalization. Despite technological progress, the lack of large, diverse, and representative Bengali datasets remains a significant barrier to the development of reliable and generalizable models. Most existing works depend on narrow-scope datasets, often sourced from single platforms like Facebook or Twitter, which limits cross-platform adaptability and model robustness. Furthermore, the need for domain adaptation, multi-class learning, and model-specific interpretability frameworks remains largely unattended.

To bridge these gaps, we propose an end-to-end system built upon a large-scale, richly diverse Bengali dataset, integrating ML, DL, transformer-based models (TBM), and explainable AI (XAI) techniques. This comprehensive approach aims to enhance model performance, interpretability, and generalizability for effective cyberbullying detection in the Bengali language.

This section presents a detailed overview of the dataset used in the study, including its collection sources, preprocessing techniques, and preparation steps. We describe the origin and characteristics of the data, followed by the methods applied to clean and tokenize the Bengali textual content to ensure it is suitable for further analysis.

Cyberbullying is one of the most serious problems in today’s highly digital environment, with social media being the prime platform for online abuse [15]. It includes hate speech that targets ethnicity, gender, age, religion, and other such bases [16], [17]. Cyberbullying has been studied widely in many languages; however, much work is yet to be done on underrepresented languages like Bengali [18]. Hence, this study proposes to find a way around this gap by studying cyberbullying in Bengali social media, which will provide useful insights for academic research and real-life applications. So, a systematic detection and mitigation of online abuse would be a burning need with more than 250 million speakers and their ever-increasing presence in social media. The work has implemented a systematic approach to the challenges of cyberbullying detection in the Bengali language. A rich dataset of Bengali texts is compiled. The Sentiment Analysis model extracts the positive, negative comments from the dataset [19], [20]. [Fig. 7 a] shows the original sentiment distribution and [Fig. 7 b] the final annotated class distribution The model distinguishes negative sentiments from positive ones, with around 30,000 comments expressing positive emotions and the remaining 40,000 comments reflecting negative feelings on social media platforms [Table 3]. We note this moderate class imbalance (approximately equal to 57% negative vs. 43% positive comments). To mitigate bias, we used stratified train-test splits and applied class weights in the loss function so that misclassifications of the smaller class are penalized more. These steps ensured balanced learning despite the imbalance. The model identifies sentiment as positive or negative using the BERT transformer architecture, which captures efficient context for this text. This is achieved by fine-tuning the model on a labeled dataset, allowing it to adjust its parameters and minimize classification errors, resulting in a more accurate sentiment analysis for Bengali text [Fig. 5]. We have applied LDA methods afterward for latent topic extraction from the negative comments and got useful ideas regarding prevalent themes in cyberbullying speech [21]. [Fig. 4] displays the top words (scores) for each of the five topics. For instance, Topic 1’s top terms relate to age, Topic 2 to gender, etc., confirming that LDA captured coherent thematic groups aligned with our label categories. We use Latent Dirichlet Allocation (LDA) as a feature extraction technique in a supervised learning setting for classifying Bengali cyberbullying text. Once we have trained the LDA model with the topics, we extract the topic distribution for each document; the distribution provides the probability of each document belonging to each of the 5 topics. The document-topic probabilities are useful as features for supervised classification. It creates a list of topic probabilities for each document that shows how much that document is affiliated with each topic. For example, if there are 5 topics, then it would have 5 values corresponding to the likelihood of that document to belong to a particular topic. Later, several ML, TBMs, and DL are adopted for analyzing and classifying the data [22], [23], [24]. The many-faceted approach, indeed, allows us to evaluate several modeling techniques in terms of their performance with respect to the detection of harmful online behavior [25]. We have introduced a hybrid model named Bengali Cyberbullying Detection(BCD), which integrates CNN, BiLSTM, and GRU architectures to effectively capture both local and sequential features for improved text classification performance. We trained all classification models for 5 epochs using the Adam optimizer with a learning rate of 0.0001 and a batch size of 32. Hyperparameters were chosen based on initial tuning experiments. We used 10-fold cross-validation on the training set to evaluate each model’s average performance. Stratified splits ensured proportional class representation in each fold. These settings were chosen because in preliminary tests, accuracy plateaued by epoch 5, and further training yielded diminishing returns. A learning rate of 0.0001 allowed stable convergence without divergence or vanishing gradients, as observed during hyperparameter tuning. To explain better the understanding of how our model has derived at a particular prediction [26], we have integrated LIME into our models [27], [28]. The study also uses BERT embeddings, experimenting a bit more with CNN, ANN, and other hybrid models to enhance the detection. This methodology not only tries to increase the detection accuracy of cyberbullying but also ensures that the inferences drawn would be interpretable and actionable [29], helpful towards the needs of the platform moderators and policymakers [Fig. 3 a]. In this way, the critical aspects addressed by our approach have distinguished it from the existing works and have contributed significantly to low-resource language cyberbullying research, such as Bengali (Algorithm 1).Fig. 7

Sentiment and Annotated Dataset Distributions.

Our data, while large, comes from particular Bengali social media platforms and periods and therefore can be subject to demographic or topical bias. For example, particular offline situations or some dialects are not represented, potentially hindering generalizability. Also, while annotators were experienced, manual detection of cyberbullying is not objective; certain sarcasm or implicit comments might have been interpreted inconsistently. Finally, pretrained models over general corpora (e.g., mBERT) may still miss nuanced domain-specific vocabulary. We understand these shortcomings and propose future research in diversifying data sources and annotators.

The data were sourced from publicly available platforms, including Kaggle and other Bengali text corpora. As the data were publicly accessible, no consent was required. The dataset was anonymized, and the study complies with the redistribution policies of the respective platforms, including Kaggle, Facebook, and Twitter.

Gazi Tahsina Sharmin Jahin: Conceptualization, Methodology, Writing – original draft, Writing – review & editing, Visualization. Firoze Maliha: Conceptualization, Software, Data curation. Nuren Nafisa: Validation, Supervision. Mohammad Mahadi Hassan: Project administration. M. Jamshed Alam Patwary: Writing – review & editing, Validation.