Detecting hateful memes is a complex task due to the interplay of text and visuals, with subtle cultural cues often determining whether content is harmful. This challenge is amplified in Bangla, a low-resource language where existing resources provide only binary labels or single dimensions of hate. To bridge this gap, we introduce BanHateME, a comprehensive Bangla hateful meme dataset with hierarchical annotations across three levels: binary hate, hate categories, and targeted groups. The dataset comprises 3,819 culturally grounded memes, annotated with substantial inter-annotator agreement. We further propose a hierarchical loss function that balances predictions across levels, preventing bias toward binary detection at the expense of fine-grained classification. To assess performance, we pair pretrained language and vision models and systematically evaluate three multimodal fusion strategies: summation, concatenation, and co-attention, demonstrating the effectiveness of hierarchical learning and cross-modal alignment. Our work establishes BanHateME as a foundational resource for fine-grained multimodal hate detection in Bangla and contributes key insights for content moderation in low-resource settings.

Understanding scientific concepts in native languages is crucial for educational accessibility and knowledge transfer. In this work, we present a comprehensive evaluation of Large Language Models (LLMs) on Bangla physics questions, introducing P6Jiggasha, a novel dataset of 1,500 multiple-choice questions compiled from HSC physics textbooks, supplementary guides, admission preparation books, and past examination papers from various educational boards. We evaluate three state-of-the-art models—GPT-4.1, Gemini-2.5 Pro, and DeepSeek-R1-Distill-Llama-70B—on both native Bangla questions and their English translations. Our results reveal significant performance variations, with GPT-4.1 achieving 86.67% accuracy on Bangla questions in a single inference, while other models show substantial improvement through multiple inference attempts, with Gemini-2.5 Pro reaching 89.52% after four iterations. We introduce a Cumulative Accuracy@k metric to evaluate iterative reasoning capabilities and provide comprehensive analysis across six physics topics and six question types. Our error analysis reveals systematic cross-lingual inconsistencies where models produce contradictory answers for identical questions across languages. This study provides valuable insights into the capabilities and limitations of current LLMs for low-resource scientific question answering and establishes benchmarks for future research in Bangla natural language processing.

Adapting large pre-trained language models (LLMs) to downstream tasks typically requires fine-tuning, but fully updating all parameters is computationally prohibitive. Parameter-Efficient Fine-Tuning (PEFT) methods like Low-Rank Adaptation (LoRA) reduce this cost by updating a small subset of parameters. However, the standard approach of jointly training LoRA adapters and a new classifier head from a cold start can lead to training instability, as the classifier chases shifting feature representations. To address this, we propose LP-FT-LoRA, a novel three-stage training framework that decouples head alignment from representation learning to enhance stability and performance. Our framework first aligns the classifier head with the frozen backbone via linear probing, then trains only the LoRA adapters to learn task-specific features, and finally performs a brief joint refinement of the head and adapters. We conduct extensive experiments on five Bangla NLP benchmarks across four open-weight compact transformer models. The results demonstrate that LP-FT-LoRA consistently outperforms standard LoRA fine-tuning and other baselines, achieving state-of-the-art average performance and showing improved generalization on out-of-distribution datasets.

Millions in Bangladesh speak Sylheti and Chittagonian (Chatgaiyya) dialects, yet most public health guidance exists only in Standard Bangla, which creates barriers and safety risks. Ad-hoc translation further harms comprehension, while challenges such as scarce data, non-standard spelling, medical terms, numerals, and idioms make accurate translation difficult. We present BanglaCHQ-Prantik, the first benchmark for this setting, extending BanglaCHQ-Summ with human gold references from 17 native translators. We evaluate Qwen 2.5 3B, Gemma 3 1B, GPT-4o mini, and Gemini 2.5 Flash under zero-shot, one-shot, five-shot, and chain-of-thought prompts, using BLEU, ROUGE-1/2/L, and METEOR. Closed-source models (GPT-4o, Gemini 2.5) lead overall, with Gemini 2.5 Flash being strongest. Few-shot prompting helps especially for Sylheti, though errors persist with terminology, numerals, and idioms. The dataset is designed to support both NLP research and public health communication by enabling reliable translation across regional Bangla dialects. To our knowledge, this is the first medical-domain dataset for Sylheti/Chittagonian.

Online safety in low-resource languages relies on effective hate speech detection, yet Bangla remains critically underexplored. Existing resources focus narrowly on binary classification and fail to capture the evolving, implicit nature of online hate. To address this, we introduce BanHate, a large-scale Bangla hate speech dataset, comprising 19,203 YouTube comments collected between April 2024 and June 2025. Each comment is annotated for binary hate labels, seven fine-grained categories, and seven target groups, reflecting diverse forms of abuse in contemporary Bangla discourse. We develop a tailored pipeline for data collection, filtering, and annotation with majority voting to ensure reliability. To benchmark BanHate, we evaluate a diverse set of open- and closed-source large language models under prompting and LoRA fine-tuning. We find that LoRA substantially improves open-source models, while closed-source models, such as GPT-4o and Gemini, achieve strong performance in binary hate classification, but face challenges in detecting implicit and fine-grained hate. BanHate sets a new benchmark for Bangla hate speech research, providing a foundation for safer moderation in low-resource languages. Our dataset is available at: https://huggingface.co/datasets/aplycaebous/BanHate.

Bangla text on the internet often appears in mixed scripts that combine native Bangla characters with their Romanized transliterations. To ensure practical usability, language models should be robust to naturally occurring script mixing. Our work investigates the robustness of current LLMs and Bangla language models under various transliteration-based textual perturbations, i.e., we augment portions of existing Bangla datasets using transliteration. Specifically, we replace words and sentences with their transliterated text to emulate realistic script mixing, and similarly, replace the top k salient words to emulate adversarial script mixing. Our experiments reveal interesting behavioral insights and vulnerabilities to robustness in language models for Bangla, which can be crucial for deploying such models in real-world scenarios and enhancing their overall robustness.

This paper presents our approach for the BLP Shared Task 1, where we implemented Linear Probing of Pre-trained Transformer-based Models for Bangla Hate Speech Detection. The goal of the task was to customize the existing models so that they’re capable of automatically identifying hate speech in Bangla social media text, with a focus on YouTube comments. Our approach relied on fine-tuning several pre-trained BERT models, adapting them to the shared task dataset for improved classification accuracy. To further enhance performance, we applied linear probing on three of the fine-tuned models, enabling more effective utilization of the learned representations. The combination of these strategies resulted in a consistent top-15 ranking across all subtasks of the competition. Our findings highlight the effectiveness of linear probing as a lightweight yet impactful technique for enhancing hate speech detection in low-resource languages like Bangla.

This paper explores various prompting strategies in the BLP-2025 Shared Task 2, utilizing a pipeline that first translates Bangla problem descriptions into English with GPT-4o,then applies techniques like zero-shot, few-shot,chain of thought, synthetic test case integration, and a self-repair loop. We evaluated fourLLMs (GPT-4o, Grok-3, Claude 3.7 Sonnet,and Qwen2.5-Coder 14B). Our findings revealthat while traditional methods like few-shotand chain-of-thought prompting provided inconsistent gains, the integration of explicit unittests delivered a substantial performance boostacross all models. The most effective strategycombined zero-shot prompting with these synthetic tests and a self-repair loop, leading GPT4o to achieve a top Pass@1 score of 72.2%.These results represent the value of using explicit constraints and iterative feedback in codegeneration, offering a solid framework that improves the model’s code generation capabilities.

We present Team BD’s submission to the BEA 2025 Shared Task on Pedagogical Ability Assessment of AI-powered Tutors, under Track 1 (Mistake Identification) and Track 2 (Mistake Location). Both tracks involve three-class classification of tutor responses in educational dialogues – determining if a tutor correctly recognizes a student’s mistake (Track 1) and whether the tutor pinpoints the mistake’s location (Track 2). Our system is built on MPNet, a Transformer-based language modelthat combines BERT and XLNet’s pre-training advantages. We fine-tuned MPNet on the task data using a class-weighted cross-entropy loss to handle class imbalance, and leveraged grouped cross-validation (10 folds) to maximize the use of limited data while avoiding dialogue overlap between training and validation. We then performed a hard-voting ensemble of the best models from each fold, which improves robustness and generalization by combining multiple classifiers. Ourapproach achieved strong results on both tracks, with exact-match macro-F1 scores of approximately 0.7110 for Mistake Identification and 0.5543 for Mistake Location on the official test set. We include comprehensive analysis of our system’s performance, including confusion matrices and t-SNE visualizations to interpret classifier behavior, as well as a taxonomy of common errors with examples. We hope our ensemble-based approach and findings provide useful insights for designing reliable tutor response evaluation systems in educational dialogue settings.

Detecting misogyny in multimodal content remains a notable challenge, particularly in culturally conservative and low-resource contexts like Bangladesh. While existing research has explored hate speech and general meme classification, the nuanced identification of misogyny in Bangla memes, rich in metaphor, humor, and visual-textual interplay, remains severely underexplored. To address this gap, we introduce BanMiMe, the first comprehensive Bangla misogynistic meme dataset comprising 2,000 culturally grounded samples where each meme includes misogyny labels, humor categories, metaphor localization, and detailed human-written explanations. We benchmark the various performance of open and closed-source vision-language models (VLMs) under zero-shot and prompt-based settings and evaluate their capacity for both classification and explanation generation. Furthermore, we systematically explore multiple fine-tuning strategies, including standard, data-augmented, and Chain-of-Thought (CoT) supervision. Our results demonstrate that CoT-based fine-tuning consistently enhances model performance, both in terms of accuracy and in generating meaningful explanations. We envision BanMiMe as a foundational resource for advancing explainable multimodal moderation systems in low-resource and culturally sensitive settings.

The proliferation of transliterated texts in digital spaces has emphasized the need for detecting and classifying hate speech in languages beyond English, particularly in low-resource languages. As online discourse can perpetuate discrimination based on target groups, e.g. gender, religion, and origin, multi-label classification of hateful content can help in understanding hate motivation and enhance content moderation. While previous efforts have focused on monolingual or binary hate classification tasks, no work has yet addressed the challenge of multi-label hate speech classification in transliterated Bangla. We introduce BanTH, the first multi-label transliterated Bangla hate speech dataset. The samples are sourced from YouTube comments, where each instance is labeled with one or more target groups, reflecting the regional demographic. We propose a novel translation-based LLM prompting strategy that translates or transliterates under-resourced text to higher-resourced text before classifying the hate group(s). Experiments reveal further pre-trained encoders achieving state-of-the-art performance on the BanTH dataset while translation-based prompting outperforms other strategies in the zero-shot setting. We address a critical gap in Bangla hate speech and set the stage for further exploration into code-mixed and multi-label classification in underrepresented languages.

Recent advancements in speech-language models have yielded significant improvements in speech tokenization and synthesis. However, effectively mapping the complex, multidimensional attributes of speech into discrete tokens remains challenging. This process demands acoustic, semantic, and contextual information for precise speech representations. Existing speech representations generally fall into two categories: acoustic tokens from audio codecs and semantic tokens from speech self-supervised learning models. Although recent efforts have unified acoustic and semantic tokens for improved performance, they overlook the crucial role of contextual representation in comprehensive speech modeling. Our empirical investigations reveal that the absence of contextual representations results in elevated Word Error Rate (WER) and Word Information Lost (WIL) scores in speech transcriptions. To address these limitations, we propose two novel distillation approaches: (1) a language model (LM)-guided distillation method that incorporates contextual information, and (2) a combined LM and self-supervised speech model (SM)-guided distillation technique that effectively distills multimodal representations (acoustic, semantic, and contextual) into a comprehensive speech tokenizer, termed DM-Codec. The DM-Codec architecture adopts a streamlined encoder-decoder framework with a Residual Vector Quantizer (RVQ) and incorporates the LM and SM during the training process. Experiments show DM-Codec significantly outperforms state-of-the-art speech tokenization models, reducing WER by up to 13.46%, WIL by 9.82%, and improving speech quality by 5.84% and intelligibility by 1.85% on the LibriSpeech benchmark dataset.

Despite significant progress in large language models (LLMs), their knowledge and evaluation continue to be centered around high-resource languages, leaving critical gaps in low-resource settings. This raises questions about how effectively LLMs handle subjects that require locally relevant knowledge. To address this challenge, we need a robust dataset that reflects the knowledge of underrepresented regions such as Bangladesh. In this paper, we present ***SOMAJGYAAN***, a Bangla multiple-choice dataset consisting of 4,234 questions, annotated across five levels of difficulty. The questions are drawn from Bangladesh’s National Curriculum and Global Studies textbooks, covering a wide range of domains including History, Geography, Economics, Social Studies, Politics and Law, and Miscellaneous topics. Difficulty levels were assigned by four expert annotators to minimize annotation bias. The experiments reveal that closed-source LLMs perform better than open-source LLMs. While fine-tuning open-source models on improves their performance, they still fall short of matching closed-source LLMs. Our findings highlight the importance of culturally grounded evaluation datasets and task-specific adaptation to improve LLM performance in low-resource language settings.

Social networks extensively feature memes, particularly cartoon images, as a prevalent form of communication often conveying complex sentiments or harmful content. Detecting such content, particularly when it involves Bengali and English text, remains a multimodal challenge. This paper introduces ***CMBan***, a novel and culturally relevant dataset of 2,641 annotated cartoon memes. It addresses meme classification based on their sentiment across five key categories: Humor, Sarcasm, Offensiveness, Motivational Content, and Overall Sentiment, incorporating both image and text features. Our curated dataset specifically aids in detecting nuanced offensive content and navigating complexities of pure Bengali, English, or code-mixed Bengali-English languages. Through rigorous experimentation involving over 12 multimodal models, including monolingual, multilingual, and proprietary architectures, and utilizing prompting methods like Chain-Of-Thought (CoT), findings suggest this cartoon-based, code-mixed meme content poses substantial understanding challenges. Experimental results demonstrate that closed models excel over open models. While the LoRA fine-tuning strategy equalizes performance across model architectures and improves classification of challenging aspects in multilingual meme contexts, this work advances meme classification by providing effective solution for detecting harmful content in multilingual meme contexts.

Low-resource languages like Bangla are severely limited by the lack of datasets. Romanized Bangla texts are ubiquitous on the internet, offering a rich source of data for Bangla NLP tasks and extending the available data sources. However, due to the informal nature of romanized text, they often lack the structure and consistency needed to provide insights. We address these challenges by proposing: (1) BanglaTLit, the large-scale Bangla transliteration dataset consisting of 42.7k samples, (2) BanglaTLit-PT, a pre-training corpus on romanized Bangla with 245.7k samples, (3) encoders further-pretrained on BanglaTLit-PT achieving state-of-the-art performance in several romanized Bangla classification tasks, and (4) multiple back-transliteration baseline methods, including a novel encoder-decoder architecture using further pre-trained encoders. Our results show the potential of automated Bangla back-transliteration in utilizing the untapped sources of romanized Bangla to enrich this language. The code and datasets are publicly available: https://github.com/farhanishmam/BanglaTLit.

Stemmers are commonly used in NLP to reduce words to their root form. However, this process may discard important information and yield incorrect root forms, affecting the accuracy of NLP tasks. To address these limitations, we propose a Contextual Bangla Neural Stemmer for Bangla language to enhance word representations. Our method involves splitting words into characters within the Neural Stemming Block, obtaining vector representations for both stem words and unknown vocabulary words. A loss function aligns these representations with Word2Vec representations, followed by contextual word representations from a Universal Transformer encoder. Mean Pooling generates sentence-level representations that are aligned with BanglaBERT’s representations using a MLP layer. The proposed model also tries to build good representations for out-of-vocabulary (OOV) words. Experiments with our model on five Bangla datasets shows around 5% average improvement over the vanilla approach. Notably, our method avoids BERT retraining, focusing on root word detection and addressing OOV and sub-word issues. By incorporating our approach into a large corpus-based Language Model, we expect further improvements in aspects like explainability.

In this study, we examine and analyze the behavior of several graph-based models for Bangla text classification tasks. Graph-based algorithms create heterogeneous graphs from text data. Each node represents either a word or a document, and each edge indicates relationship between any two words or word and document. We applied the BERT model and different graph-based models including TextGCN, GAT, BertGAT, and BertGCN on five different datasets including SentNoB, Sarcasm detection, BanFakeNews, Hate speech detection, and Emotion detection datasets for Bangla text. BERT’s model bested the TextGCN and the GAT models by a large difference in terms of accuracy, Macro F1 score, and weighted F1 score. BertGCN and BertGAT are shown to outperform standalone graph models and BERT model. BertGAT excelled in the Emotion detection dataset and achieved a 1%-2% performance boost in Sarcasm detection, Hate speech detection, and BanFakeNews datasets from BERT’s performance. Whereas, BertGCN outperformed BertGAT by 1% for SetNoB, and BanFakeNews datasets while beating BertGAT by 2% for Sarcasm detection, Hate Speech, and Emotion detection datasets. We also examined different variations in graph structure and analyzed their effects.

In the context of the dynamic realm of Bangla communication, online users are often prone to bending the language or making errors due to various factors. We attempt to detect, categorize, and correct those errors by employing several machine learning and deep learning models. To contribute to the preservation and authenticity of the Bangla language, we introduce a meticulously categorized organic dataset encompassing 10,000 authentic Bangla comments from a commonly used social media platform. Through rigorous comparative analysis of distinct models, our study highlights BanglaBERT’s superiority in error-category classification and underscores the effectiveness of BanglaT5 for text correction. BanglaBERT achieves accuracy of 79.1% and 74.1% for binary and multiclass error-category classification while the BanglaBERT is fine-tuned and tested with our proposed dataset. Moreover, BanglaT5 achieves the best Rouge-L score (0.8459) when BanglaT5 is fine-tuned and tested with our corrected ground truths. Beyond algorithmic exploration, this endeavor represents a significant stride in enhancing the quality of digital discourse in the Bangla-speaking community, fostering linguistic precision and coherence in online interactions. The dataset and code is available at https://github.com/SyedT1/BaTEClaCor.

The BLP-2023 Task 1 aims to develop a Natural Language Inference system tailored for detecting and analyzing threats from Bangla YouTube comments. Bangla language models like BanglaBERT have demonstrated remarkable performance in various Bangla natural language processing tasks across different domains. We utilized BanglaBERT for the violence detection task, employing three different classification heads. As BanglaBERT’s vocabulary lacks certain crucial words, our model incorporates some of them as new special tokens, based on their frequency in the dataset, and their embeddings are learned during training. The model achieved the 2nd position on the leaderboard, boasting an impressive macro-F1 Score of 76.04% on the official test set. With the addition of new tokens, we achieved a 76.90% macro-F1 score, surpassing the top score (76.044%) on the test set.

This paper introduces the top-performing approachof “Aambela” for the BLP-2023 Task2: “Sentiment Analysis of Bangla Social MediaPosts”. The objective of the task was tocreate systems capable of automatically detectingsentiment in Bangla text from diverse socialmedia posts. My approach comprised finetuninga Bangla Language Model with threedistinct classification heads. To enhance performance,we employed two robust text classificationtechniques. To arrive at a final prediction,we employed a mode-based ensemble approachof various predictions from different models,which ultimately resulted in the 1st place in thecompetition.