Reproduce “IndicIRSuite: Multilingual Dataset and Neural Information Models for Indian Languages” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Saiful Haq, Ashutosh Sharma, Pushpak Bhattacharyya
IIT Bombay researchers introduce IndicIRSuite, a collection of large-scale neural information retrieval resources for 11 Indian languages, featuring the translated INDIC-MARCO dataset and language-specific Indic-ColBERT models. These models demonstrate an average 47.47% MRR@10 improvement over existing multilingual baselines, advancing IR capabilities for low-resource Indian languages.
Key insights
- High-quality machine translation (via NLLB) can effectively generate large-scale, general-domain datasets for low-resource languages, proving crucial for training modern neural IR systems.
- Training language-specific neural IR models on substantial translated datasets significantly outperforms general multilingual models that lack specific training on the target Indian languages.
- Even for languages not directly supported by the base multilingual encoder (e.g., Assamese by mBERT), linguistic similarity with supported languages combined with high-quality translated data can lead to notable performance improvements.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.