AI-NLP-ML Group, IIT Patna

12/04/2026

We’re excited to share that the following survey paper:

📄 “Bridging the Linguistic Divide: Leveraging Large Language Models for Machine Translation”
has been accepted to Language Resources & Evaluation (Q1 journal)🎉

🔗 Paper: https://arxiv.org/abs/2504.01919

💻 Code: https://github.com/babangain/llm-mt-survey

🔍 What does this work cover?

This paper provides a comprehensive and up-to-date survey of how LLMs are transforming Machine Translation (MT), spanning methods, evaluation, and real-world challenges

💡 Key Takeaways

▪ LLMs are reshaping MT paradigms: From prompting and in-context learning to fine-tuning and RL-based alignment, translation is evolving beyond traditional encoder–decoder systems.

▪ Data quality > model scale: Performance gains increasingly depend on high-quality data, preference signals, and context usage, not just bigger models.

▪ Low-resource MT remains challenging: LLMs help, but performance is still highly dependent on cross-lingual support and data availability.

▪ Prompting vs Fine-tuning is not a competition: They are complementary tools, each useful under different constraints and deployment scenarios.

▪ Evaluation is evolving: LLMs as evaluators offer interpretability but introduce biases and instability, making multi-metric evaluation essential.

▪ Beyond sentence-level translation: Emerging work explores document-level, discourse-aware, and instruction-controlled MT, opening new research directions.

🌍 Why this matters

As LLMs become central to NLP, understanding their role in MT is crucial for building robust, inclusive, and controllable multilingual systems.

This survey aims to serve as a single reference point for researchers and practitioners navigating this rapidly evolving space. Also, the code enables practitioners to rapidly build LLM-based machine translation systems using few-shot prompting, supervised fine-tuning (SFT), and parameter-efficient fine-tuning (PEFT) approaches.

Large Language Models (LLMs) are rapidly reshaping machine translation (MT), particularly by introducing instruction-following, in-context learning, and preference-based alignment into what has traditionally been a supervised encoder-decoder paradigm. This survey provides a comprehensive and up-to-d...

09/04/2026

We are pleased to share that the following papers from our lab have been accepted in ACL 2026:

1. Deepak K., B. Gain, A. Ekbal. 𝗠𝗶𝗻𝗱 𝘁𝗵𝗲 𝗣𝗮𝘂𝘀𝗲: 𝗗𝗶𝘀𝗳𝗹𝘂𝗲𝗻𝗰𝘆-𝗔𝘄𝗮𝗿𝗲 𝗢𝗯𝗷𝗲𝗰𝘁𝗶𝘃𝗲 𝗧𝘂𝗻𝗶𝗻𝗴 𝗳𝗼𝗿 𝗠𝘂𝗹𝘁𝗶𝗹𝗶𝗻𝗴𝘂𝗮𝗹 𝗦𝗽𝗲𝗲𝗰𝗵 𝗖𝗼𝗿𝗿𝗲𝗰𝘁𝗶𝗼𝗻 𝘄𝗶𝘁𝗵 𝗟𝗟𝗠𝘀. In ACL (Main).

𝗦𝘂𝗺𝗺𝗮𝗿𝘆: Speech disfluencies like fillers, repetitions, and false starts in ASR transcripts hurt downstream NLP tasks, yet simply deleting them often breaks grammar and meaning. This paper proposes a multilingual pipeline that detects disfluencies via a sequence tagger and rewrites transcripts using an instruction-tuned LLM, further enhanced by contrastive learning. Experiments across Hindi, Bengali, and Marathi show consistent improvements over strong baselines, offering a practical and scalable solution for disfluency correction.

2. Prajwal V. Kajare, P. Priya, B. Santra, A. Ekbal. 𝗣𝗥𝗜𝗦𝗠𝗔: 𝗣𝗿𝗲𝗳𝗲𝗿𝗲𝗻𝗰𝗲-𝗥𝗲𝗶𝗻𝗳𝗼𝗿𝗰𝗲𝗱 𝗦𝗲𝗹𝗳-𝗧𝗿𝗮𝗶𝗻𝗶𝗻𝗴 𝗔𝗽𝗽𝗿𝗼𝗮𝗰𝗵 𝗳𝗼𝗿 𝗜𝗻𝘁𝗲𝗿𝗽𝗿𝗲𝘁𝗮𝗯𝗹𝗲 𝗘𝗺𝗼𝘁𝗶𝗼𝗻𝗮𝗹𝗹𝘆 𝗜𝗻𝘁𝗲𝗹𝗹𝗶𝗴𝗲𝗻𝘁 𝗡𝗲𝗴𝗼𝘁𝗶𝗮𝘁𝗶𝗼𝗻 𝗗𝗶𝗮𝗹𝗼𝗴𝘂𝗲𝘀. In ACL (Main).

𝗦𝘂𝗺𝗺𝗮𝗿𝘆: We introduce PRISMA, an emotionally intelligent interpretable negotiation dialogue system for job interviews and resource allocation. Its key novelty is Emotion-aware Negotiation Strategy-informed Chain-of-Thought framework, which models how the system perceives, understands, uses, and manages emotions to enable transparent, human-like responses. Building on this framework, we introduce two datasets, JobNego and ResNego, and train PRISMA with a preference-reinforced self-training approach to improve emotional appropriateness and interpretability during negotiation.

3. Sandeep K., Y. Kamdar, A. Hossain, B. Kumari, T. Saikh, A. Ekbal. 𝗪𝗵𝗲𝗻 𝗥𝗲𝘃𝗶𝗲𝘄𝘀 𝗗𝗶𝘀𝗮𝗴𝗿𝗲𝗲: 𝗙𝗶𝗻𝗲-𝗚𝗿𝗮𝗶𝗻𝗲𝗱 𝗖𝗼𝗻𝘁𝗿𝗮𝗱𝗶𝗰𝘁𝗶𝗼𝗻 𝗔𝗻𝗮𝗹𝘆𝘀𝗶𝘀 𝗶𝗻 𝗦𝗰𝗶𝗲𝗻𝘁𝗶𝗳𝗶𝗰 𝗣𝗲𝗲𝗿 𝗥𝗲𝘃𝗶𝗲𝘄𝘀. In ACL (Findings).

𝗦𝘂𝗺𝗺𝗮𝗿𝘆: Detecting contradictions in peer reviews is underexplored, with existing methods treating it as binary and overlooking disagreement severity. This paper presents a fine-grained formulation with graded intensity scores and evidence spans, supported by RevCI, a new expert-annotated benchmark. The multi-agent framework IMPACT combines evidence extraction, reasoning, and adjudication, distilled into TIDE - a compact, cost-efficient model. Results show strong improvements over baselines, offering a practical tool for Area Chairs in large-scale conference reviewing.