Hybrid retrieval slightly beats BM25 on Maharashtra scheme search
Original title: Hybrid Retrieval for Government Scheme Search in English and Marathi
Equal-weight BM25 and embedding fusion lifts MRR from 0.864 to at most 0.879 on curated bilingual queries.
Researchers describe a bilingual search system for Maharashtra government schemes and public services that fuses BM25 with sentence-embedding similarity at equal weight. It indexes aligned English-Marathi descriptions and is evaluated with IndicSBERT and MahaSBERT against BM25 and dense-only retrieval. Hybrid MRR reaches 0.872 and 0.879 versus a BM25 baseline of 0.864, while dense-only systems score 0.663 and 0.655. Gains over lexical search are described as modest and metric-dependent.
- MahaSBERT fusion raises precision from 0.781 to 0.808.
- Hybrid recall matches BM25 at 0.946.
- Dense-only IndicSBERT and MahaSBERT MRR are 0.663 and 0.655.
- Queries were manually curated and bilingual.
From the original article
1 min read at arxiv.orgRetrieving government schemes requires matching citizens' information needs to formal administrative descriptions, often across languages and terminology. We present a bilingual search system for Maharashtra government schemes and public services that combines BM25 retrieval with sentence-embedding similarity through equal-weight score fusion.
By Aishwarya Gunjal · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.