Morphology beats multilingual embeddings for Kazakh retrieval — a 300-query benchmark
TL;DR. I built an open, reproducible retrieval benchmark for Kazakh (300 queries × 3
categories, 8,370 Kazakh Wikipedia passages, BEIR format). The headline: a simple
BM25 + morphological stemmer — no GPU, no embedding model — beats two of three
multilingual dense models (Google LaBSE, IBM Granite) outright, and is the most robust
system on the hardest query category, where the dense models collapse. Dataset, code,
and a preprint with DOI are all public.
Why Kazakh breaks out-of-the-box retrieval
Kazakh is agglutinative: one word appears in dozens of inflected forms.
бала (child) → балалар, баланың, балама, балалардың, балаларымызға …
A lexical index treats each surface form as a different token, so a query in one form
misses documents written in another. Multilingual embeddings are supposed to paper over
this — but how well do they actually do it on Kazakh? Nobody had measured it on an open,
reproducible benchmark. So I did.
The benchmark
- Corpus: 8,370 passages from Kazakh Wikipedia (CC BY-SA 4.0)
- Queries: 300, split into three categories:
natural— everyday phrasinginflected— query and answer use different grammatical forms (the morphology stress test)vocabulary-gap— query uses a synonym/paraphrase not present in the gold passage (the semantic stress test)
- Metrics: Recall@k, MRR@10, nDCG@10, with paired-bootstrap significance
- Format: BEIR-compatible (
corpus,queries,qrels)
Results — nDCG@10 (n=300)
| System | inflected | natural | vocab-gap | ALL |
|---|---|---|---|---|
| BM25 (baseline) | 0.627 | 0.703 | 0.741 | 0.690 |
| BM25 + Stemmer | 0.727 | 0.772 | 0.764 | 0.754 |
| Dense LaBSE (Google) | 0.477 | 0.546 | 0.419 | 0.481 |
| Dense Granite (IBM) | 0.791 | 0.923 | 0.303 | 0.672 |
| Dense E5 (multilingual-e5-base) | 0.845 | 0.947 | 0.562 | 0.785 |
Three things jump out:
1. A stemmer alone closes most of the gap. BM25 → BM25+Stemmer lifts overall nDCG@10
from 0.690 to 0.754 (+9%, p=0.0001), and on the morphology-heavy inflected queries
from 0.627 to 0.727 (+16%, p=0.0017). No neural network, no GPU — just morphological
normalization of corpus and queries.
2. BM25+Stemmer beats two of three multilingual dense models. It outscores LaBSE
(0.481) by a wide margin and IBM Granite (0.672) overall — while running on a CPU in
milliseconds. The "just use embeddings" reflex is not free, and on Kazakh it isn't even
clearly better.
3. On the hardest category, dense models collapse — the stemmer doesn't. Onvocabulary-gap, IBM Granite drops to 0.303 and E5 to 0.562, while BM25+Stemmer holds
at 0.764 — the most robust system on the queries that are supposed to be the dense
models' home turf. (I reported this collapse to the Granite team.)
To be clear and honest: E5 wins overall (0.785 vs 0.754). A strong multilingual
embedder does edge out lexical+morphology on average. But it does so at the cost of a GPU,
a heavier pipeline, and a hard collapse on vocabulary-gap — where the cheap, transparent,
CPU-only stemmer is the safest bet.
A negative result I'm keeping in the open
I also tried the obvious "fix" for vocabulary-gap: synonym query expansion. It made
things worse across the board — overall nDCG@10 fell from 0.754 to 0.539 (−0.215).
Expansion added far more noise than signal. Negative results like this rarely get
published; this one is in the benchmark in full, with the sub-analysis of why.
And the end-to-end RAG honesty check
Plugging retrieval into a full RAG pipeline (Qwen2.5-7B, n=300): the stemmer raises
retrieval hit@3 from 0.737 to 0.803, but the end-to-end answer accuracy gain is not
statistically significant (McNemar p=0.63). The bottleneck downstream is the
Kazakh-language generator, not the retriever. The stemmer's value is proven at the
retrieval level — and I say so rather than overclaiming the whole pipeline.
Takeaways
- For low-resource, morphologically rich languages, don't skip the cheap lexical
baseline — and definitely don't skip morphological normalization. It can beat
multilingual embeddings outright. - Measure per query type. Aggregate scores hide that dense models can ace easy queries
and collapse on hard ones. - Publish negative results. They save everyone else the experiment.
Links
- Dataset (this page): https://huggingface.co/datasets/Tim2190/kaz-rag-search-benchmark
- Code & full results (GitHub): https://github.com/Tim2190/Kaz-RAG-search-benchmark
- Preprint with DOI (Zenodo): https://doi.org/10.5281/zenodo.20605663
- Kazakh Stemmer: https://qaz-api.vercel.app/
Disclosure: I'm the developer of the Kazakh stemmer evaluated here (System 2). All systems
were run on identical queries, corpus, and metrics; everything is public and independently
reproducible; and results unfavorable to my system — its statistical insignificance on
vocabulary-gap vs BM25, and E5's higher overall score — are reported in full.