Information Retrieval · NLP · Benchmark

Code-Mixed Product Search Relevance Benchmark

How well does product search handle Hinglish — code-mixed Hindi–English queries like lal running shoes and gym ke liye shoes? A reproducible benchmark on Amazon ESCI comparing BM25, multilingual dense retrieval, and reranking.

On code-mixed queries, lexical BM25 with reranking beats multilingual dense retrieval (nDCG@10 0.30 vs 0.16) — the opposite of the naive expectation. Dense is the best method on clean English but the worst on Hinglish.

348
hand-authored queries
348
mapped to ESCI answer key (100%)
6,991
products in corpus
3
retrieval paradigms
33
passing tests

The vernacular gap

nDCG@10 for each method, split by query language. Every method drops sharply from English to Hinglish — and dense retrieval collapses hardest.

English Hinglish / Hindi
Dense scores 0.61 on English but 0.16 on Hinglish — a 0.45-point collapse. BM25 + cross-encoder is the most robust to code-mixing.

Overall ranking quality

Mean nDCG@10 over all 348 queries. BM25 + cross-encoder is the best overall pipeline (0.496).

Where it breaks — nDCG@10 by query category

Colour intensity encodes nDCG@10 — vivid = strong, faded = near-zero. Dense craters on roman_hindi, voice_search, and pure_hindi while staying strong on English-shaped categories.

CategorynBM25BM25+CEDenseDense+CE
Cell colour encodes nDCG@10 (faint → near-zero, vivid → ~0.9). Devanagari (pure_hindi) scores 0.00 on BM25 — zero character overlap with the Latin-script catalog.

Failure modes (selected)

Method & reproducibility

348 hand-authored Hinglish queries (10 domains, 13 categories) are mapped — via exact + fuzzy string matching, then a multilingual-embedding (E5) semantic fallback — onto the US subset of Amazon ESCI (1.8M graded judgments), inheriting Exact/Substitute/Complement/Irrelevant labels as graded relevance. All 348 queries map to an answer key, each with a recorded confidence tier. Three systems are compared: BM25 (from scratch), a multilingual E5 bi-encoder with FAISS, and a cross-encoder / LLM reranker. Metrics (nDCG, MRR) are implemented from scratch. Everything is reproducible from a fixed seed, pinned dependencies, and numbered scripts, guarded by a unit-test suite.

Honest limitations: answer keys are product-level (fine colour/size/price constraints aren't separately scored); ~24% of mapped keys are fuzzy; Indian-vernacular items are under-represented in the US catalog; the dense model is CPU-sized (e5-small). See the technical report for full detail.