The development of robust transliteration techniques to enhance the effectiveness of transforming Romanized scripts into native scripts is crucial for Natural Language Processing tasks, including sentiment analysis, speech recognition, information retrieval, and intelligent personal assistants. Despite significant advancements, state-of-the-art multilingual models still face challenges in handling Romanized script, where the Roman alphabet is adopted to represent the phonetic structure of diverse languages. Within the South Asian context, where the use of Romanized script for Indo-Aryan languages is widespread across social media and digital communication platforms, such usage continues to pose significant challenges for cutting-edge multilingual models. While a limited number of transliteration datasets and models are available for Indo-Aryan languages, they generally lack sufficient diversity in pronunciation and spelling variations, adequate code-mixed data for large language model (LLM) training, and low-resource adaptation. To address this research gap, we introduce a novel transliteration dataset for two popular Indo-Aryan languages, Hindi and Bengali, which are ranked as the 3rd and 7th most spoken languages worldwide. Our dataset comprises nearly 1.8 million Hindi and 1 million Bengali transliteration pairs. In addition to that, we pre-train a custom multilingual seq2seq LLM based on Marian architecture using the developed dataset. Experimental results demonstrate significant improvements compared to existing relevant models in terms of BLEU and CER metrics.

This study investigates the attribution patterns underlying Chain-of-Thought (CoT) reasoning in multilingual LLMs. While prior works demonstrate the role of CoT prompting in improving task performance, there are concerns regarding the faithfulness and interpretability of the generated reasoning chains. To assess these properties across languages, we applied two complementary attribution methods--ContextCite for step-level attribution and Inseq for token-level attribution--to the Qwen2.5 1.5B-Instruct model using the MGSM benchmark. Our experimental results highlight key findings such as: (1) attribution scores excessively emphasize the final reasoning step, particularly in incorrect generations; (2) structured CoT prompting significantly improves accuracy primarily for high-resource Latin-script languages; and (3) controlled perturbations via negation and distractor sentences reduce model accuracy and attribution coherence. These findings highlight the limitations of CoT prompting, particularly in terms of multilingual robustness and interpretive transparency.

The BLP-2025 Task 1A requires Bengali hate speech classification into six categories. Traditional supervised approaches need extensive labeled datasets that are expensive for low-resource languages. We developed PromptGuard, a few-shot framework combining chi-square statistical analysis for keyword extraction with adaptive majority voting for decision-making. We explore statistical keyword selection versus random approaches and adaptive voting mechanisms that extend classification based on consensus quality. Chi-square keywords provide consistent improvements across categories, while adaptive voting benefits ambiguous cases requiring extended classification rounds. PromptGuard achieves a micro-F1 of 67.61, outperforming n-gram baselines (60.75) and random approaches (14.65). Ablation studies confirm chi-square-based keywords show the most consistent impact across all categories.

Over the past few years, improving LLM code generation capabilities has been a key focus in NLP research. Despite Bengali having 242 million native speakers worldwide, it receives little attention when it comes to training LLMs. More recently, various fine-tuning and augmented generation techniques have been employed to significantly enhance code generation performance. However, they require considerable expertise and resources to utilize effectively as an end user. The goal of our work is to democratize access to powerful code generation tools in resource-constrained emerging markets, enabling users to leverage them in their native language. We introduce a novel approach that combines Test-Driven Development (TDD) and Code Interpreter (CI), utilizing open-weight models, which improves the baseline accuracy for code generation with Bengali prompts and achieves an overall accuracy of 85%. Our approach requires no finetuning and proves that even the smallest models in the same family can attain up to 98% accuracy compared to the largest models. All of our results are publicly shared in GitHub for validation and reproducibility.

In [4], the authors present the DisCoCirc (Distributed Compositional Circuits) formalism for the English language, a grammar-based framework derived from the production rules that incorporates circuit-like representations in order to give a precise categorical theoretical structure to the language. In this paper, we extend this approach to develop a similar framework for Bengali and apply it to translation tasks between English and Bengali. A central focus of our work lies in reassessing the effectiveness of DisCoCirc in reducing language bureaucracy. Unlike the result suggested in [5], our findings indicate that although it works well for a large part of the language, it still faces limitations due to the structural variation of the two languages. We discuss the possible methods that might handle these shortcomings and show that, in practice, DisCoCirc still struggles even with relatively simple sentences. This divergence from prior claims not only highlights the framework's constraints in translation but also suggest scope for future improvement. Apart from our primary focus on English-Bengali translation, we also take a short detour to examine English conjunctions, following [1], showing a connection between conjunctions and Boolean logic.

In this paper, we describe our system under the team name BLEU Monday for the English-to-Indic Multimodal Translation Task at W AT 2025. We participate in the text-only translation tasks for English-Hindi, English-Bengali, English-Malayalam, and English-Odia language pairs. We present a two-stage approach that addresses quality issues in the training data through automated error detection and correction, followed by parameter-efficient model fine-tuning. Our methodology introduces a vision-augmented judge-corrector pipeline that leverages multimodal language models to systematically identify and correct translation errors in the training data. The judge component classifies translations into three categories: correct, visually ambiguous (requiring image context), or mistranslated (poor translation quality). Identified errors are routed to specialized correctors: GPT-4o-mini regenerates captions requiring visual disambiguation, while IndicTrans2 retranslates cases with pure translation quality issues. This automated pipeline processes 28,928 training examples across four languages, correcting an average of 17.1% of captions per language. We then apply Low-Rank Adaptation (LoRA) to fine-tune the IndicTrans2 en-indic 200M distilled model on both original and corrected datasets.

WORK Although the findings highlight the effectiveness of fine - tuned transformer models for Bengali - Sylheti translation, several limitations remain. The dataset size (5,002 parallel sentences) restricts the models' capacity to generalize across diverse syntactic structures, stylistic variations, and domain - specific expressions. In addition, orthographic inconsistencies in Sylheti introduce noise, leading to training instability, particularly in models like mBART - 50. Another limitation is the reliance on automatic evaluation metrics such as BLEU and chrF, which may not fully capture the linguistic richness or cultural nuance of Sylheti. Future research should therefore focus on expanding the datas et through community - driven contributions and data augmentation strategies. Incorporating orthographic normalization could improve consistency and reduce variability during training. Hybrid approaches that combine the strengths of pre - trained LLMs with fin e - tuned NMT models may also enhance translation robustness in low - resource settings. Finally, incorporating human evaluation will provide a more comprehensive assessment of translation adequacy, fluency, and cultural alignment.

Bengali social media platforms have witnessed a sharp increase in hate speech, disproportionately affecting women and adolescents. While datasets such as BD-SHS provide a basis for structured evaluation, most prior approaches rely on either computationally costly full-model fine-tuning or proprietary APIs. This paper presents the first application of Parameter-Efficient Fine-Tuning (PEFT) for Bengali hate speech detection using LoRA and QLoRA. Three instruction-tuned large language models - Gemma-3-4B, Llama-3.2-3B, and Mistral-7B - were fine-tuned on the BD-SHS dataset of 50,281 annotated comments. Each model was adapted by training fewer than 1% of its parameters, enabling experiments on a single consumer-grade GPU. The results show that Llama-3.2-3B achieved the highest F1-score of 92.23%, followed by Mistral-7B at 88.94% and Gemma-3-4B at 80.25%. These findings establish PEFT as a practical and replicable strategy for Bengali and related low-resource languages.

In multilingual healthcare environments, automatic disease diagnosis from clinical text remains a challenging task due to the scarcity of annotated medical data in low-resource languages and the linguistic variability across populations. This paper proposes SwasthLLM, a unified, zero-shot, cross-lingual, and multi-task learning framework for medical diagnosis that operates effectively across English, Hindi, and Bengali without requiring language-specific fine-tuning. At its core, SwasthLLM leverages the multilingual XLM-RoBERTa encoder augmented with a language-aware attention mechanism and a disease classification head, enabling the model to extract medically relevant information regardless of the language structure. To align semantic representations across languages, a Siamese contrastive learning module is introduced, ensuring that equivalent medical texts in different languages produce similar embeddings. Further, a translation consistency module and a contrastive projection head reinforce language-invariant representation learning. SwasthLLM is trained using a multi-task learning strategy, jointly optimizing disease classification, translation alignment, and contrastive learning objectives. Additionally, we employ Model-Agnostic Meta-Learning (MAML) to equip the model with rapid adaptation capabilities for unseen languages or tasks with minimal data. Our phased training pipeline emphasizes robust representation alignment before task-specific fine-tuning. Extensive evaluation shows that SwasthLLM achieves high diagnostic performance, with a test accuracy of 97.22% and an F1-score of 97.17% in supervised settings. Crucially, in zero-shot scenarios, it attains 92.78% accuracy on Hindi and 73.33% accuracy on Bengali medical text, demonstrating strong generalization in low-resource contexts.

Grounding vision--language models in low-resource languages remains challenging, as they often produce fluent text about the wrong objects. This stems from scarce paired data, translation pivots that break alignment, and English-centric pretraining that ignores target-language semantics. We address this with a compute-aware Bengali captioning pipeline trained on LaBSE-verified EN--BN pairs and 110k bilingual-prompted synthetic images. A frozen MaxViT yields stable visual patches, a Bengali-native mBART-50 decodes, and a lightweight bridge links the modalities. Our core novelty is a tri-loss objective: Patch-Alignment Loss (PAL) aligns real and synthetic patch descriptors using decoder cross-attention, InfoNCE enforces global real--synthetic separation, and Sinkhorn-based OT ensures balanced fine-grained patch correspondence. This PAL+InfoNCE+OT synergy improves grounding, reduces spurious matches, and drives strong gains on Flickr30k-1k (BLEU-4 12.29, METEOR 27.98, BERTScore-F1 71.20) and MSCOCO-1k (BLEU-4 12.00, METEOR 28.14, BERTScore-F1 75.40), outperforming strong CE baselines and narrowing the real--synthetic centroid gap by 41%.