Niqqy/kimeru-clir

# Cross-Lingual Information Retrieval for Kimeru The first documented information retrieval experiment for **Kimeru**, a Bantu language spoken by roughly two million people in Kenya's Meru region — a language with no prior NLP resources or retrieval benchmarks. This project investigates whether augmenting an English retrieval corpus with Kimeru-language passages improves cross-lingual retrieval performance for Meru speakers, comparing a classical lexical method (BM25) against two multilingual embedding models (MiniLM and LaBSE). > **Note on data:** The underlying corpus is built from parallel English/Kimeru Bible > translations (Good News Translation, American Bible Society; Kimeru Bible, Biblica), which > are copyrighted works. The raw extracted passage text is **not included** in this repo — > only code, evaluation queries, and result/metric summaries are published. See > Reproducing the corpus below. ## Summary of findings - Adding 311 Kimeru passages to an English-only index (622 passages total) raised BM25's Mean Reciprocal Rank (MRR) on Kimeru queries from **0.17 → 0.73**, with English-query performance essentially unaffected (0.70 → 0.68). - **LaBSE** showed meaningful cross-lingual transfer to Kimeru *without any Kimeru text in the corpus at all*, reaching an MRR of 0.29 on Kimeru queries against an English-only index — likely via indirect transfer from related Bantu languages (e.g. Swahili) in its training data. - **BM25 outperformed both neural embedding models overall**, attributed to strong lexical overlap between LLM-generated synthetic queries and their source passages — a gap that narrows notably on human-written queries. - **MiniLM degraded** on Kimeru queries after corpus augmentation, likely due to weaker cross-lingual alignment routing queries toward superficially similar but semantically unrelated Kimeru text; LaBSE did not show this failure mode. ## Project structure ``` kimeru-clir/ ├── bm25/ # BM25 retrieval implementation + …