Yes, under specific conditions. Multilingual AI can improve product search on an international marketplace when it reliably connects a query typed in a shopper’s own language to listings written in another language, or to a catalog organised around a different one. Amazon Science and eBay have published evidence for this. The gains they report are tied to particular systems, language pairs, baselines and metrics, and they should be read that way.
Five ways AI can bridge the language gap in product search
Multilingual search is not one technique. The approaches below work at different points in the pipeline, and the evidence for each is different in kind.
As an Amazon Associate I earn from qualifying purchases.
| Approach | What it changes | Reported evidence | Main limitation |
|---|---|---|---|
| Query translation into the catalog language | Rewrites the shopper’s query into the catalog’s language before it reaches the existing search engine | Amazon Science, 2020: offline nDCG@8 improved by 11% (Spanish-to-English) and 3% (French-to-English); online product-type search defects reduced by 10% and 22% respectively | Can ignore how the search engine responds; sensitive to spelling errors; weaker for languages outside training; may mishandle names |
| Shared multilingual representations | Maps queries and product descriptions into one representation space, so matching can work across languages without an explicit translation step | Amazon Science, 2019: F1 gains of 5% to 24% over monolingual models in its evaluated task | Figures are F1 comparisons in the evaluated task, not measured live search results |
| Graph-based multilingual retrieval | Uses multilingual transformer models and graph neural networks to model interactions between queries and items | Amazon Science, 2021: method described; comparative performance not stated | No head-to-head result in the cited description |
| Retrieval-augmented product-title translation | Supplies similar bilingual product examples to a large language model when it translates listing titles | Amazon Science, 2024: chrF gains of up to 15.3% for language pairs where the model has limited proficiency | chrF measures title translation quality, not search relevance |
| Catalog-language setting and synonyms | Sets the catalog language at upload; one-way or two-way synonyms handle mixed-language queries | Google Cloud documentation: setting and synonym controls described; performance not stated | Documented as a workaround for mixed-language queries against the default catalog language |
Query translation: the 2020 Amazon Science results
Amazon Science’s 2020 paper studies a global store whose catalog is in a primary language, while some shoppers search in a secondary language. The system first identifies the query’s language, then applies a neural machine translation model fine-tuned on a human-curated parallel query corpus. It is designed to copy entities such as model numbers rather than translate them. A traffic re-ranker then chooses among candidate transformations, favouring those likely to help the existing search engine, so the output is judged by how that engine responds and not only by how the translation reads.
The paper compares this system with a state-of-the-art statistical machine-translation system for product search. Its reported results cover two language pairs:
#1 Best Overall
| Language pair | Offline nDCG@8 | Online product-type search defects |
|---|---|---|
| Spanish-to-English | Improved by 11% | Reduced by 10% |
| French-to-English | Improved by 3% | Reduced by 22% |
The two measures did not move together. The offline score alone would rank Spanish-to-English as the larger gain, while the online defect measure ranks French-to-English higher. A team that tracked only offline ranking metrics would reach a different conclusion about which language pair benefited most.
Shared representations and graph-based retrieval
Shared query and product representations
Amazon Science’s 2019 account of multilingual shopping describes mapping queries about the same product from different languages into a shared representation space, and doing the same for product descriptions. A query in one language can then match a listing in another without a separate translation step. The reported gains are F1 comparisons in the evaluated task:
Rank #2
| Multilingual model | Compared with | Reported F1 gain |
|---|---|---|
| French-and-German model | French monolingual model | 11% |
| French-and-German model | German monolingual model | 5% |
| Five-language model | French monolingual model | 24% |
| Five-language model | German monolingual model | 19% |
Against the same two monolingual baselines, the five-language model shows larger gains than the French-and-German model (24% and 19%, against 11% and 5%). The source does not show how an F1 gain on its evaluated task would carry over to ranking in a live store.
Recommended Free Tools
Graph-based multilingual retrieval
A 2021 Amazon Science publication describes graph-based multilingual product retrieval. It uses multilingual transformer language models and graph neural networks to model interactions between queries and items. No comparative performance figure is reported in the description of this work, so it is a design option to test rather than a demonstrated gain.
Rank #3
- Provides quick, reliable answers to your questions about words
- Economically priced to fit your budget
- Makes a great gift for new high school or college graduates
Translating catalog titles, not only queries
A 2024 Amazon Science paper works on the catalog side. It proposes retrieval-augmented generation for product-title translation: similar bilingual product information is retrieved and supplied to a large language model as examples. For language pairs where the model has limited proficiency, the paper reports chrF score gains of up to 15.3%. chrF measures title translation quality. It is not a measure of search relevance, so the figure cannot be read as a 15.3% improvement in what shoppers find.
Mixed-language queries and catalog settings
Not every case needs a translation model. Google Cloud’s documentation for AI Commerce Search says the catalog language is set when the catalog is uploaded. For queries that mix languages against the default catalog language, the documented workaround is a one-way or two-way synonym control. A practical sequence is:
- Decide the catalog language before import, because it is set at upload.
- Pull a sample of real queries from your search logs and flag those that mix two languages, such as a brand name in one language combined with a generic term in another.
- For a mixed term that should map in one direction only, add a one-way synonym from the query term to the catalog term.
- For terms that should match each other in either direction, add a two-way synonym.
- Re-test the flagged queries after each change, and check the current Google Cloud documentation for the exact synonym configuration, since product behaviour changes over time.
Where multilingual search fails
eBay’s engineering article describes translating search queries into the language of each market and matching them against inventory. It highlights several failure modes that a reader should expect in any deployment:
Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchWindows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstall- Noisy, user-generated text: typos and non-dictionary terms that a translation model may not recognise.
- Ambiguity when a query lacks category context, so the same word can point to different kinds of product.
- Entities such as brand names and model numbers, which must be carried across rather than translated. The 2020 Amazon Science system is built to copy them.
- Spelling and grammar errors, and languages that were absent from training data. The 2020 paper lists both as weaknesses of direct translation, alongside its tendency to ignore the search engine’s response.
eBay’s engineering article puts the priority plainly. As Tatyana Badeka writes: “Providing an accurate, grammatically correct translation of a query is never enough; what we always keep in mind is user intent and relevance of the results.”
Best Value
- Designed for student use anywhere
- Hands-on learning resource any time you need to reference a word
- Makes a great gift for new high school or college graduates
How to test a multilingual search change
The 2020 paper argues that standard machine translation metrics are unsuitable for this task. Its statement reads: “standard machine translation evaluation metrics such as BLEU are unsuitable for this application.” Its proposed offline measure combines how accurately the transformed query reflects shopping intent with how well the existing search system responds to it. A 2022 Amazon Science publication likewise frames query-translation evaluation around downstream search ranking and proposes a ranking-based evaluation framework. In practice, that points to the following test design:
- Build a query set for each target language from real traffic, including misspellings, brand names and model numbers, and tag the queries that mix languages.
- Use the current production search as the baseline. If you are testing translation, add a conventional machine-translation baseline as well; the 2020 study compared against a statistical machine-translation system.
- Measure the relevance of the products returned with a ranked metric such as nDCG, with judgements made by people who read the target language.
- Count task-specific failures separately, such as results that are the wrong type of product, and report them next to the ranking score.
- Include an online test, not only an offline one.
- Report the language pairs, catalog, baseline, query mix and evaluation setting with every result.
What the evidence does not establish
- The studies come from the companies that run the marketplaces or sell the services, and they report on their own systems. The sources cited here do not include independent replication on other marketplaces.
- Results cover specific language pairs: Spanish-to-English and French-to-English for query translation, and French, German and five-language models for shopping text.
- The studies date from 2019 to 2024. Models have changed since, so a current system may behave differently from the ones described.
- No single benchmark establishes an improvement rate that applies across international marketplaces. The metrics differ (nDCG, F1 and chrF), so the figures cannot be ranked against each other.
- The sources describe how these systems work, but they do not establish that a given merchant or marketplace can use them.
No regulator or standards body has published guidance on this question that the sources cover, so the evidence here rests on the company studies and documentation listed above.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
Free tools Windows power users keep installed
One-click scans. No signup required.
