ViolenceBanglaBERT improves Bangla violence-text classification in controlled benchmark

A fine-tuned language model called ViolenceBanglaBERT achieved a macro F1-score of 0.75 on a held-out test set for identifying violence-instigating Bangla social-media texts, outperforming the other deep-learning and transformer classifiers evaluated under a unified experimental framework. The study used the Vio-Lens dataset, which contains 6,046 annotated texts associated with nine communal-violence incidents in Bangladesh between 2012 and 2022. Each message was assigned to one of three dataset categories: non-violence, passive violence or direct violence.

The result is an offline classification advance, not evidence that the system has prevented violence or has been deployed by a social-media platform. ViolenceBanglaBERT combines the pretrained BanglaBERT encoder with a lightweight multilayer perceptron, or MLP, classifier. BanglaBERT converts each cleaned Bangla sequence into contextual representations for its tokens, from which the model extracts the vector attached to the special classification token, known as the [CLS] token.

This 768-dimensional representation summarizes the input sequence and is passed through fully connected layers with rectified-linear-unit activation, dropout and a three-unit output layer. The resulting logits correspond to the three predefined categories, and the index with the highest score is selected as the predicted class. Transformer classifiers were trained with AdamW at a reported learning rate of 1.29 × 10⁻⁵, while the convolutional and recurrent architectures used Adam at a learning rate of 0.0001 and categorical cross-entropy loss.

The researchers treated the task as a three-class classification problem rather than a simple binary division between harmful and harmless text. Existing work had often addressed abuse, aggression or hate speech, whereas this study focused specifically on violence-instigating Bangla text and compared methods under common preprocessing, dataset-split and evaluation conditions. Prior systems for the same task included traditional machine-learning methods, convolutional and recurrent networks, hybrid CNN–recurrent architectures, BanglaBERT variants, multilingual transformers and ensembles.

Because preprocessing, augmentation and implementation differed among publications, the authors also reimplemented strong baselines where code and sufficient details were available. Before training, the pipeline removed non-Bangla text, punctuation, special characters, stopwords, URLs, mentions and hashtags, as well as texts shorter than two words. To address the original class imbalance, each training example was translated from Bangla into English with Google Translate and then translated back into Bangla.

This back-translation pipeline generated a paraphrase intended to introduce lexical and syntactic variation while preserving the original meaning. The researchers manually inspected all generated samples to check semantic consistency and retained their labels. Augmentation changed the training distribution to 34.5% non-violence, 35.5% passive violence and 30% direct violence.

The development and test sets remained unchanged, so final evaluation used the original annotated examples rather than generated translations. Several textual representations were tested for the conventional classifiers. These included Word2Vec and FastText embeddings trained with either skip-gram or continuous-bag-of-words configurations, as well as model-specific BERT tokenization.

Word2Vec and FastText vectors had a size of 200, a window size of 3 and a minimum count of 2. The deep-learning comparison included CNN, LSTM, bidirectional LSTM, bidirectional GRU and several CNN–recurrent hybrids. The transformer comparison covered multilingual BERT, XLM-R, BanglaBERT-Base and BanglaBERT.

In the controlled benchmark, transformer-based systems generally outperformed the deep-learning architectures. ViolenceBanglaBERT obtained a macro F1-score of 0.75 and accuracy of 0.78 in its best test run, compared with macro F1 scores of 0.71 and accuracy of 0.76 for the BanglaBERT baseline. Across the controlled runs, ViolenceBanglaBERT achieved the highest mean macro F1-score, 0.744, and lower variation than the competing baselines.

Back-translation augmentation was the strongest imbalance-handling strategy tested, producing a macro F1-score of 0.744 ± 0.005 and a direct-violence F1-score of 0.654 ± 0.017. Without any balancing method, direct-violence F1 fell to 0.054. Upsampling improved that category somewhat but duplicated training examples and performed less well on other categories.

Class-weighted loss and weighted focal loss provided moderate but inconsistent gains across the classes. Hyperparameters for ViolenceBanglaBERT were selected through 100 trials of Bayesian optimization using Optuna’s tree-structured Parzen Estimator sampler, with development-set macro F1 as the optimization target. Training terminated after three consecutive epochs without an improvement in validation loss.

In the reported development run, training loss decreased from 0.9523 in the first epoch to 0.4998 in the third, while validation loss declined from 0.7529 to 0.6128. Accuracy increased from 77.22% to 79.25%, precision rose from 0.7969 to 0.8072, and recall ranged from 0.6859 to 0.7461. The researchers also examined ROC and precision–recall curves, reporting class-specific ROC AUC values of 0.88 for non-violence, 0.85 for passive violence and 0.90 for direct violence.

The corresponding average precision values were 0.87, 0.78 and 0.67, with a macro-average average precision of 0.77. An ablation study indicated that the reported performance depended on more than the MLP classifier alone. Removing data augmentation and preprocessing caused the proposed configuration to perform worse than baseline BanglaBERT, indicating that the added classifier head was sensitive to noisy and imbalanced inputs.

Combining the head with cleaned text and back-translated examples produced the strongest overall results, while preprocessing also benefited the baseline. The analysis therefore did not show that one component could be removed without loss, nor did it isolate the contribution of every design choice independently. Instead, it indicated that the strongest reported configuration depended on the combination of contextual BanglaBERT representations, the classification head, preprocessing and augmentation.

The experiments were conducted in TensorFlow and PyTorch with the Hugging Face Transformers library on a system equipped with an NVIDIA Tesla T4 GPU with 15 GB of memory, a multicore Intel Xeon processor and 12.7 GB of system RAM. Error analysis showed where the system remained unreliable. In the displayed confusion matrix, it correctly assigned 927 texts to non-violence, 510 to passive violence and 142 to direct violence.

It nevertheless confused 136 non-violence examples with passive violence and classified 41 passive-violence examples as direct violence, while 168 passive-violence examples were labeled non-violence. A total of 30 direct-violence examples were assigned to non-violence and 29 to passive violence. The authors linked recurring failures to theological masking, hypothetical or conditional threats, political and religious debate vocabulary, dialect and slang, spelling inconsistencies and emotionally intense figurative expressions.

Inconsistent annotation in borderline cases introduced additional noise. The model also showed limited ability to capture deeper discourse-level context, particularly where implicit intent depended on information beyond the individual text. The study’s central finding is therefore a controlled benchmark showing that a BanglaBERT encoder with a compact classification head, combined with text cleaning and back-translation augmentation, can improve recognition of violence-instigating Bangla text on the Vio-Lens corpus.

Its scope remains constrained by a relatively small and imbalanced dataset drawn from a defined set of Bangladeshi communal-violence incidents between 2012 and 2022. The authors identify limited training data, annotation ambiguity, implicit and code-switched language, figurative expression and disagreement between passive and direct violence as major constraints. Future work is proposed to involve more diverse datasets, improved annotation quality and context-enhanced models, but the reported 0.75 macro F1 score does not establish readiness for autonomous enforcement or successful real-world violence prevention.

Subject of Research: Benchmarking a BanglaBERT-based classifier for violence-instigating Bangla social-media text

Article Title: Stop the spark detecting violence instigating Bangla text using ViolenceBanglaBERT

Article References: Chakraborty, P., Mynoddin, M., Hossain, M. S., Golam Sarwar Apu, S. M., & Biplob, U. M. S. M. (2026). Stop the spark detecting violence instigating Bangla text using ViolenceBanglaBERT. Discover Artificial Intelligence, 6(1), Article 1050. https://doi.org/10.1007/s44163-026-01696-0

Image Credits: AI Generated

DOI: 10.1007/s44163-026-01696-0

Keywords: ViolenceBanglaBERT, BanglaBERT, Bangla language, social-media text, violence detection, natural language processing, transformer models, back translation, class imbalance, machine learning, Stop, spark