Lab
Bengali language model for verification
Training a Bengali language model for the same verification tasks — so news and public claims in Bangla can be transcribed, analysed and checked as well as English ones.
Bangla is one of the world’s most widely spoken languages, but tools for analysing and checking claims in it lag far behind English. This experiment trains a Bengali model for the verification tasks in the video NLP work.
What it explores
- Data — sourcing Bengali news text and transcripts with clear provenance and permission.
- Tasks — detecting checkable claims and linking them to relevant sources.
- Evaluation — measuring where the model works and where it fails, including dialect and code-switching between Bangla and English.
It builds on published Bengali NLP research, credited in the library. No performance figures are claimed until they have been measured and can be reproduced.