How should semantic search be built for a low-resource language?

Start with the information need, not the embedding model

"Semantic search" can mean two distinct jobs:

• Understanding a short query written in Mooré, including spelling variation, morphology, word sense, and local terminology.
• Retrieving useful documents, which may be in Mooré, French, both languages, or a multilingual local collection.

These jobs should be evaluated separately. A system may map a query to a good French formulation but still retrieve poor documents. Conversely, a French web engine may retrieve excellent results for a broad query while losing a crucial Mooré meaning in the pivot. Keeping each intermediate decision makes these failures inspectable.

Begin by choosing the search scope. If the product's value is access to official French documents, a French pivot is sensible. If users need Mooré content, local services, or community knowledge, build and search a Mooré or bilingual corpus as well. An external web search engine is not a stable evaluation collection because its results, ranking, personalization, and availability change. Use a frozen, permissioned test collection for engineering decisions, then separately assess the live search experience.

Define success in user terms before choosing models. Examples include finding a local agricultural advisory, understanding an administrative requirement, locating a verified health-service page, or discovering a relevant Mooré document. Do not substitute translated English benchmark questions for these needs. They can help debug a model, but they cannot establish that the system understands local spelling, terminology, or relevance.

Preserve language identity and the original query

Use the name Mooré in the product unless the affected language community asks for another form. Model providers may use a different label, such as "Mossi," and that label should remain a technical mapping rather than a replacement for the community's name.

Store the original query, a normalized search form, and the normalization operations separately. Normalization can apply Unicode normalization, case folding when appropriate, and known accent or apostrophe alternatives. It should never overwrite the original text. The system needs to be able to explain that a match was caused by a spelling-variant rule rather than pretending the user wrote another word.

For interfaces, documents, and APIs, use a validated BCP 47 language tag from the IANA registry and record the script or regional variety only when it distinguishes real content. The W3C's language-tag guidance explains that BCP 47 language tags identify the language of text and that a short tag is preferred unless extra subtags add useful distinction. W3C language-tag guidance Keep this web metadata separate from an embedding model's own language identifier. For example, SONAR documents the model label mos_Latn for Mossi. SONAR supported-language list

Model spelling variation as data

Build a small normalization and expansion layer from the lexicon's variante, plural, pronunciation, synonym, domain, and sense fields. For every observed form, link to a stable lexical entry ID and a specific sense ID. Give exact matches a high score, then allow conservative fuzzy matches for likely typing or orthographic variation. A fuzzy match should be a candidate, not an automatic correction when it creates ambiguity.

Do not flatten the lexicon into a single French word per Mooré headword. A sense record should retain its French definition, English gloss if helpful, domain, synonyms, examples when licensed, and confidence in the mapping. This lets a query such as a term with agricultural and administrative senses produce two candidates instead of silently choosing the more common French word.

Use the lexicon as the first retrieval system

The existing lexicon is valuable because it supplies structured distinctions that generic multilingual models may miss. The first production baseline should not need model training.

1. Accept the Mooré query and retain the raw text.

2. Tokenize and normalize using the documented variant rules. Look up exact forms, then conservative fuzzy forms and morphological alternatives.

3. Retrieve a small set of sense candidates. Score exactness, edit distance, domain hints, neighboring query terms, grammatical compatibility, synonym links, and prior user selections where consent permits.

4. If two senses remain plausible, either present a short clarification in Mooré or issue two labeled French query hypotheses. Do not turn uncertainty into an invented single translation.

5. Construct a French query from the selected sense definitions and domain terms. Preserve named entities, local place names, and user-supplied tokens that should not be translated. Weight the head concept more strongly than an optional expansion term.

6. Retrieve from both a lexical French index and any bilingual or Mooré corpus. Return the original document title, URL, source language, and a transparent indication that a displayed summary was translated when that is the case.

This baseline offers a real diagnostic advantage. When a result is wrong, reviewers can see whether the error came from spelling normalization, sense selection, French query construction, retrieval, or display translation. An end-to-end model can hide all five decisions behind a plausible French sentence.

Add multilingual embeddings as a tested candidate

Multilingual embeddings map texts from different languages into a shared vector space. In theory, a Mooré query and a related French passage land near one another, allowing cross-lingual retrieval without translating every query. In practice, model language coverage, text length, spelling, topic, and the amount of Mooré seen during training determine whether that is useful.

Do not infer support from a model's marketing label such as "multilingual." LaBSE, for example, reported multilingual sentence-embedding results across 109 or more languages, but that publication does not establish quality for every other language or for short, misspelled Mooré queries. LaBSE paper A more directly relevant experiment candidate is SONAR. Its current upstream repository lists mos_Latn for Mossi among its supported text-encoder and decoder languages and labels it low-resource. That is evidence that the model exposes a technical path for the language, not evidence that it meets this application's relevance bar. SONAR supported-language list

Test at least three representation strategies against the same held-out queries:

LASER3 is another candidate family. Its upstream repository describes focus-language encoders and lists the available language-specific encoders, including the NLLB-related set. LASER3 documentation Evaluate it rather than assuming it will outperform a multilingual general-purpose encoder. Model repositories and licenses can change. Pin the exact model revision, preprocessing rules, and language label for each evaluation.

How to test a candidate encoder before using it

Create a compact Mooré to French meaning-retrieval set from cases that are not used to tune the system. For each Mooré query, native speakers should identify the intended lexicon sense or senses, and bilingual reviewers should create one or more acceptable French interpretations. Include hard contrasts in which two Mooré senses map to different French queries. Then ask the model to retrieve the correct French definition or passage from a candidate pool containing close wrong senses.

Report Recall at 1 and Recall at 5 for sense retrieval, plus failures broken down by spelling variant, query length, domain, and ambiguity. A high average can conceal a system that works on common French-like terms but fails on local names or agricultural vocabulary. Inspect nearest-neighbor lists, not only a single metric. If the encoder does not beat the lexicon baseline for the groups users care about, keep it out of the serving path or restrict it to a low-weight expansion signal.

Combine retrieval methods when evaluation supports it

Hybrid search means that lexical and semantic signals both nominate documents, then a ranker combines the candidates. It is not merely putting a text field and an embedding field in the same database. Current Elastic documentation illustrates the pattern by indexing original text for lexical search while also creating embeddings for semantic search. Elastic hybrid-search documentation

For this use case, use four candidate channels:

• Mooré lexical channel: variants, headwords, synonyms, domain labels, and user-selected clarifications against the lexicon and any Mooré corpus.
• French pivot lexical channel: one or more weighted French queries against an indexed French collection.
• Cross-lingual dense channel: a Mooré query embedding against embeddings of French, Mooré, or bilingual documents, using a verified model and its correct language setting.
• Reranking channel: a lightweight, rule-aware ranker that uses source quality, language match, lexical evidence, dense score, domain, and the selected sense. Keep the lexical evidence visible in the result record.

Fuse the channels with a simple reciprocal-rank or calibrated weighted-score method first. Tune weights only on a development split, then lock them before scoring the final holdout. A more complex neural reranker may help later, but it can easily learn artifacts from a small data set. Start with methods that an evaluator can challenge and improve.

The French pivot should create a limited candidate set, not overwrite the Mooré meaning. Keep every French hypothesis with its source sense IDs and confidence. If a query is ambiguous, expose a respectful disambiguation choice before using a high-impact query expansion. If the system cannot identify a sense, return a transparent low-confidence search, suggest an alternate spelling, or ask a short follow-up. Do not claim a translation is correct merely because a web search returned plausible pages.

Build a small, lawful corpus before trying to train a large model

Sparse data is a collection problem before it is a fine-tuning problem. Assemble a curated corpus of documents people actually need, such as public-service information, education material, agricultural advisories, locally produced news, and community-approved Mooré material. For each document, keep the source URL or identifier, publisher, date, language, title, body, geographic applicability, domain, license or permission, and ingestion revision.

Prioritize data provenance and community agreement. Do not scrape a website simply because it is visible to a crawler. Respect copyrights, terms, access limits, and the expectations of speakers and publishers. Do not upload sensitive query logs, personal messages, health records, or private local material to an external embedding or translation provider without an appropriate lawful basis, a data-processing review, and user disclosure.

The lexicon can bootstrap supervised data, but it is not a parallel corpus. Use its senses to make carefully reviewed query-to-definition pairs and hard negative pairs, such as two related French definitions from different domains. Have native speakers author or validate natural short queries rather than treating a dictionary headword as a complete query. Split by lexical sense and concept family so a spelling variant or near-duplicate definition cannot appear in both training and test sets.

Fine-tune an embedding model only after the baseline exposes a repeatable failure class and a distinct training set exists. A reasonable first adaptation objective is contrastive learning from reviewed Mooré query to correct French definition or relevant document pairs, with semantically close wrong senses as negatives. Keep the final native-speaker relevance set untouched. If fine-tuning helps only on synthetic dictionary pairs and not on real information needs, it has not improved search.

Evaluate relevance with speakers, contexts, and disagreement

Translation quality is not retrieval quality. A French query can be a reasonable paraphrase yet retrieve an irrelevant page. Evaluate at three layers:

Use at least two independent speakers for a sample of queries and record disagreement rather than forcing a single label. A third reviewer or a facilitated adjudication can resolve labels for the release set, while the disagreement rate becomes a useful signal that the query is genuinely ambiguous or that the relevance guide is underspecified. Keep language variety, domain, and query form in the annotation record, but minimize personal data.

Create a relevance guide with examples of "directly answers the question," "partly useful," "topically related but not useful," "wrong sense," "untrustworthy," and "unsafe for this context." Culturally relevant evaluation should include locally meaningful entities, public institutions, farming seasons and practices, idioms where contributors agree they are appropriate, and the types of information people seek in Mooré. It should not reduce the language to a French vocabulary test.

For high-stakes topics, use stricter gates. A retrieval system can surface health or legal information, but it should label source and date, avoid presenting generated summaries as professional advice, and direct users to qualified local services where appropriate. Evaluate whether the French pivot accidentally promotes outdated, foreign, or inapplicable guidance over locally relevant information.

The NLLB research program is a useful methodological reminder rather than a substitute benchmark: it combined native-speaker exploratory work, human-translated evaluation, human evaluation, and safety assessment across many translation directions. No Language Left Behind paper That type of human-centered evaluation is more relevant here than a leaderboard score alone.

A worked design for the Mooré to French use case

Example

Hypothetical setup: a user enters a short Mooré query seeking advice about a maize-storage pest. The query contains an accepted local spelling variant, and the lexicon returns two candidate senses: one for a household insect and one for a crop-storage pest. The document collection includes French agricultural advisories, a small licensed set of Mooré leaflets, and source metadata for each item.

The system preserves the raw query, links the recognized variant to its lexicon entry, and uses surrounding terms plus the agricultural domain to give the crop-storage sense a higher score. It produces a French candidate query from the reviewed sense definition and related agricultural terms, while the cross-lingual encoder independently retrieves passages near the original Mooré query. The lexical, French, and dense channels nominate results. A rule-aware reranker favors an official local advisory whose domain and location metadata fit, then presents the original French title and URL with an optional clearly marked Mooré summary.

The evaluation record asks native speakers whether the selected sense matches the user's need and whether the top results would be useful and locally applicable. A failure in which the household-insect sense wins is logged as query understanding, not hidden as a retrieval miss. The vector score helps rank candidates, while the lexicon, local corpus, and speaker reviews let the team check whether the results fit the query.

A practical build sequence

1. Document the data. Version the lexicon, assign stable entry and sense IDs, record its rights and provenance, and write normalization rules with speaker review.

2. Build the lexical baseline. Implement exact and fuzzy variant matching, sense candidates, French query construction, and a simple French index. Log candidate senses and confidence for consented test traffic.

3. Create the evaluation set first. Collect real, consented short queries and relevance labels from speakers. Separate development and final holdout sets by concept family and spelling variant.

4. Add a small bilingual corpus. Ingest permissioned Mooré and French documents with metadata, source quality, language, date, and license.

5. Run representation experiments. Compare the baseline with one or two documented multilingual encoders, such as SONAR or LASER3, plus a translation-pivot variant. Pin exact revisions and preprocessing.

6. Fuse only proven signals. Introduce hybrid candidate fusion and a transparent reranker if the holdout shows complementary improvement. Retain a lexical-only fallback for encoder failure or low confidence.

7. Pilot carefully. Release to a small group of speakers, collect opt-in correction feedback, review false positives and omitted results, and revise the lexicon and evaluation guide before expanding access.

Common mistakes and better alternatives