DocTalkBN: A Novel Dataset of Expert Telemedicine Conversations in Bengali

Anik Saha; Fahmida Sultana Naznin; Sadatul Islam Sadi; Ananya Shahrin Promi; Wahid Al Azad Navid; Rifat Shahriyar

cs.CLAug 27, 2026

DocTalkBN: A Novel Dataset of Expert Telemedicine Conversations in Bengali

Authors: Anik Saha, Fahmida Sultana Naznin, Sadatul Islam Sadi, Ananya Shahrin Promi, Wahid Al Azad Navid, Rifat Shahriyar

Organizations: Bangladesh University of Engineering and Technology (BUET)

Abstract

Reliable medical conversational AI requires authentic expert--patient interaction data, yet such datasets remain scarce, especially for low-resource languages such as Bengali. We present DocTalkBN, a large-scale multimodal dataset of real-world expert telemedicine conversations in Bengali, collected from nationally broadcast telemedicine programs featuring board-certified physicians. DocTalkBN contains 557.63 hours of paired audio and text, 1,515 multi-turn patient calls, 10,274 host--doctor question--answer exchanges, totaling 1.7M tokens, spanning 26 medical specialties. Unlike prior resources derived from medical forums, written health content, or synthetic data, our dataset preserves the spontaneity, contextual richness, and spoken characteristics of authentic medical interactions in a low-resource setting. To support benchmark-driven research, we further construct three downstream tasks from the corpus, medical triage classification, advice safety evaluation, and medical named entity recognition, and benchmark a diverse set of large language models and encoder-based baselines. Our results show that DocTalkBN is a practically useful resource, particularly for clinically grounded reasoning tasks. We release this resource to facilitate future research on reliable medical NLP and safer, more culturally grounded healthcare systems for low-resource languages. Our source codes and dataset are publicly available at https://anonymous.4open.science/r/doctalk.

Figures & tables

Figure 1: Dialogue between doctor and patient annotated with metadata tags.

Figure 2: Overview for dataset curation and processing workflow.

Dataset Name	Source Type	Language	#Dialogues	#Utterances	#Diseases	Department
MZ Wei et al. (2018)	Online Text	Chinese	710	-	4	Pediatrics
DX Xu et al. (2019)	Online Text	Chinese	527	2,186	5	Pediatrics
CMDD Lin et al. (2019)	Medical Forum	Chinese	2,067	87,005	4	Pediatrics
MedDG Liu et al. (2022)	Online Text	Chinese	17,864	385,951	12	Gastroenterology
MedDialog-CN Pal and others (2020)	Online Text	Chinese	3,407,494	11,260,564	172	51 Departments
RealMedDial Xu et al. (2022)	Video Clips	Chinese	2,637	24,255	55	17 Departments

Table 1: Comparison between our dataset and existing medical dialogue datasets.

Overall Dataset Statistics
Dataset Duration (Hours)	557.63
Unique Medical Specialties	26
Patient Calls Proportion (%)	12.9
Host-Doctor QA Proportion (%)	87.1
Patient-Doctor Calls
Total Patient-Doctor Calls	1,515

Table 2: Detailed Dataset Statistics

Downstream Task	#Videos	Samples	Total Labels	Class	Count
Advice Safety	1,744	8,689	29,957	SAFE	22,513
				HARM	7,444
Medical NER	1,842	11,239	105,697	DIS	31,680
				SYM	23,385
				ANA	21,624
				TRT	13,802

Table 3: Downstream Benchmark Datasets & Class Distributions

# Dialogues	11,789
# Utterances	25,410
# Tokens	1,671,210
Avg. # Utterances in a Dialogue	2.16
Max # Utterances in a Dialogue	14
Min # Utterances in a Dialogue	1
Avg. # Tokens in an Utterance	65.8

Table 4: Statistics of the Bangla DocTalkBN dataset

Model	Setting	Macro F1
DeepSeek V3.2	Zero-shot	0.320
DeepSeek V3.2	Few-shot	0.383
Gemma 3 27B Instruct	Zero-shot	0.412
Gemma 3 27B Instruct	Few-shot	0.422
GPT-4o	Zero-shot	0.425
GPT-4o	Few-shot	0.448

Table 5: Triage classification performance across models and settings. Best values are shown in bold and second best values are underlined .

Model	Setting	Macro F1
DeepSeek V3.2	Zero-shot	0.937
DeepSeek V3.2	Few-shot	0.949
Gemma 3 27B Instruct	Zero-shot	0.951
Gemma 3 27B Instruct	Few-shot	0.955
GPT-4o	Zero-shot	0.969
GPT-4o	Few-shot	0.976

Table 6: Advice safety classification performance across models and settings. Best values are shown in bold and second best values are underlined .

Model	Setting	Strict F1 (Macro Avg.)
BanglaBERT	Fine-tuned	0.742
mmBERT	Fine-tuned	0.707
DeepSeek V3.2	Zero-shot	0.694
Gemma 3 27B Instruct	Zero-shot	0.585
GPT-4o	Zero-shot	0.581
GPT-4o	Few-shot	0.575

Table 7: Performance on Medical NER across models and prompting/training settings. Best values are shown in bold and second best values are underlined .

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Hyperparameter	BanglaBERT	mMBERT
Learning rate	$3\times 10^{-5}$	$2\times 10^{-5}$
Epochs	100	100
Train batch size	8	4
Evaluation batch size	16	8
Gradient accumulation	2	4
Warmup steps	200	200

Appendix

Table 8: Hyperparameter settings used for fine-tuning the BanglaBERT and mMBERT models.

Clinical NLP

Clinical natural language processing (NLP) develops methods to extract, interpret, and generate information from healthcare text and conversations.

See all 142 papers

IndicMedDialog: A Parallel Multi-Turn Medical Dialogue Dataset for Accessible Healthcare in Indic Languages
May 13, 2026Shubham Kumar Nigam, Suparnojit Sarkar, Piyush PatelMultilingual Language Models Healthcare
University of Birmingham, Dubai, United Arab Emirates · Heritage Institute of Technology, Kolkata, India · Madan Mohan Malaviya University of Technology, India
MedAidDialog: A Multilingual Multi-Turn Medical Dialogue Dataset for Accessible Healthcare
Mar 25, 2026Shubham Kumar Nigam, Suparnojit Sarkar, Piyush PatelMedical LLMs Text Generation
University of Birmingham, Dubai, United Arab Emirates · Heritage Institute of Technology, Kolkata, India · Madan Mohan Malaviya University of Technology, India
A Dataset of Robot-Patient and Doctor-Patient Medical Dialogues for Spoken Language Processing Tasks
May 26, 2026Heriberto Cuayahuitl, Grace JangHealthcare Speech Processing
School of Engineering and Physical Sciences University of Lincoln, UK · Lincoln Medical School Universities of Lincoln and Nothingham, UK

May 13, 2026J/K move · Enter open · S save

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

Shubham Kumar Nigam, Suparnojit Sarkar, Piyush Patel

University of Birmingham, Dubai, United Arab Emirates · Heritage Institute of Technology, Kolkata, India · Madan Mohan Malaviya University of Technology, India

Multilingual Language Models Healthcare Clinical NLP Medical LLMs

Open paper AskPDF arXiv

Most existing medical dialogue systems operate in a single-turn question--answering paradigm or rely on template-based datasets, limiting conversational realism and multilingual applicability. We introduce IndicMedDialog, a parallel multi-turn medical dialogue dataset spanning English and nine Indic languages: Assamese, Bengali, Gujarati, Hindi, Marathi, Punjabi, Tamil, Telugu, and Urdu. The dataset extends MDDial with LLM-generated synthetic consultations, translated using TranslateGemma, verified by native speakers, and refined through a script-aware post-processing pipeline to correct phonetic, lexical, and character-spacing errors. Building on this dataset, we fine-tune IndicMedLM via parameter-efficient adaptation of a quantized small language model, incorporating optional patient pre-context to personalise multi-turn symptom elicitation. We evaluate against zero-shot multilingual baselines, conduct systematic error analysis across ten languages, and validate clinical plausibility through medical expert evaluation.

Abstract

Clinical NLP

Explore similar work

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

MedAidDialog: A Multilingual Multi-Turn Medical Dialogue Dataset for Accessible Healthcare

A Dataset of Robot-Patient and Doctor-Patient Medical Dialogues for Spoken Language Processing Tasks