papersSEP 10 04:00 UTC
5-Dialects-BN Benchmark Probes How Transliteration Affects Bangla Dialect LLMs
A new arXiv paper introduces 5-Dialects-BN, a resource covering five Bangla dialects used to examine how transliteration choices influence large language model performance. The work targets the sharp accuracy drops LLMs show on low-resource, dialectally diverse languages, with Bangla being the world's sixth most spoken language. The authors position script and transliteration practices as an underexamined factor in this degradation.