BanglaTOD

A multi-domain dataset for intent classification and slot filling in Bangla task-oriented dialogue. Every utterance is written in three parallel forms: Bangla (Bengali script), English, and Banglish (Romanized Bangla), to support cross-lingual and code-mixed NLU research.

Paper: https://arxiv.org/abs/2610.18060

Overview

Property	Value
Total dialogues	209
Languages	3 (Bangla, English, Banglish)
Domains	9
Intents	64

Domain coverage

Domain	Dialogues	Intents
Education	50	9
Goverment Service	32	19
Healthcare	25	5
Mobile Financial Service	25	5
E-commerce	25	10
Restaurant	16	5
Tourist	14	5
Travel	12	4
Hotel	10	2
Total	209	64

Data format

Each file is JSON Lines, one dialogue per line:

Field	Description
dialogue_id	Unique identifier for the dialogue
domains	List of domain(s) of the dialogue
intent	The dominant user goal
turns	Ordered list of conversational turns

Usage

from datasets import load_dataset
ds = load_dataset("afri7here/BanglaTOD", "healthcare")

Citation

@misc{banglatod2026ewu,
      title={A Benchmark for Bangla Task-Oriented Dialogue Understanding},
      author={Afrina Alam Lamia and Sumiaya Sharif and Prottoy Modak and Ohiduzzaman Rayhan and Nafis Sadeq},
      year={2026},
      eprint={2610.18060},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2610.18060},
}