IndicMedDialog: A Parallel Multi-Turn Medical Dialogue Dataset for Accessible Healthcare in Indic Languages

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Natural Language Processing · Depth: Expert, medium

Summary

IndicMedDialog is a parallel multi-turn medical dialogue dataset covering English and nine Indic languages: Assamese, Bengali, Gujarati, Hindi, Marathi, Punjabi, Tamil, Telugu, and Urdu. It expands the MDDial corpus using LLM-generated synthetic consultations, translated by TranslateGemma, verified by native speakers, and refined with a script-aware post-processing pipeline to correct translation errors. Building on this dataset, IndicMedLM was fine-tuned using LoRA on a quantized small language model, incorporating patient pre-context for personalized symptom elicitation. Evaluation against zero-shot baselines identified five failure modes, including Tokenization Failure. Results showed strong diagnostic accuracy in Hindi, Marathi, and Bengali, but Assamese, Tamil, and Telugu exhibited extreme failure due to base-model tokenizer gaps, raising patient safety concerns. Medical experts confirmed the clinical plausibility and safety of the generated consultations.

Key takeaway

For NLP engineers developing medical dialogue systems for Indic languages, you must prioritize robust tokenizer development and language-specific validation. While IndicMedLM shows promise in Hindi, Marathi, and Bengali, its extreme failure in Assamese, Tamil, and Telugu due to base-model tokenizer gaps highlights critical patient safety risks. Ensure your models are thoroughly evaluated across all target languages, especially those with limited resources, to prevent diagnostic inaccuracies and ensure equitable, safe healthcare access.

Key insights

IndicMedDialog provides a multi-turn medical dialogue dataset for nine Indic languages, enabling personalized symptom elicitation.

Principles

• Synthetic data generation requires native speaker verification.
• Base-model tokenizer gaps can severely impact diagnostic accuracy.

Method

Extend MDDial with LLM-generated consultations, translate via TranslateGemma, verify with native speakers, and apply script-aware post-processing to correct errors before LoRA fine-tuning a quantized LLM.

In practice

• Use LoRA for efficient fine-tuning of medical LLMs.
• Incorporate patient pre-context for personalized dialogue.

Topics

• Medical Dialogue Systems
• Indic Languages
• LLM Fine-tuning
• Synthetic Data Generation
• Tokenizer Limitations
• Healthcare AI

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• EMSDialog: Synthetic Multi-person Emergency Medical Service Dialogue Generation from Electronic Patient Care Reports via Multi-LLM Agents — Synthetic multi-speaker EMS dialogues generated from ePCRs enhance conversational diagnosis prediction accuracy and stability.
• IndicDB -- Benchmarking Multilingual Text-to-SQL Capabilities in Indian Languages — IndicDB reveals a significant "Indic Gap" in LLM Text-to-SQL performance for Indian languages compared to English.
• Dial HEALTHDIAL for Advice: A Multilingual and Multi-Parallel Spoken Dialogue Dataset for Knowledge-Grounded Information Seeking — Creating large-scale, multilingual, multi-parallel spoken dialogue datasets is methodologically challenging, revealing performance disparities.
• ArogyaSutra: A Multi-Agent Framework for Multimodal Medical Reasoning in Indic Languages — A multi-agent framework and multilingual multimodal dataset significantly improve medical reasoning in low-resource Indic language healthcare.
• A Pipeline for Generating Longitudinal Synthetic Clinical Notes Using Large Language Models — A modular LLM-powered pipeline generates consistent, diverse longitudinal synthetic clinical notes to enable privacy-preserving healthcare AI development.
• OpenMedQ: Broad Open Pretraining for Medical Vision-Language Models — Broad, open medical pretraining data enables state-of-the-art VLM performance with fewer parameters.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.