Making Amazon Search Understand Local-Language Queries Typed in Latin Letters

A shopper on Amazon describes the product they want in a local language, but types it in Latin letters rather than in that language's own script. For example, a shopper enters **laal chudidar** in the search bar. "Laal" is Hindi for red, and a chudidar is a fitted trouser usually worn with a kurta, so the shopper expects the first results to be red chudidars. How would you make sure Amazon returns relevant results when shoppers type product details in a local language, so that in this example the top results are red chudidars? ```hint Split what the shopper said Separate the words that name the product from the words that describe it. For each kind of word, ask how search could fail on it today, and what the system would need to know about both the query and the catalog to get it right. ``` ### Clarifying Questions - Which languages come first: Hindi typed in Latin letters only, other Indian languages too, and should queries typed in native scripts follow the same path? - What happens today for queries like this: no results, the wrong products, or the right products ranked too low? Is there data on how often they fail? - Are product listings written mostly in English, or do sellers also provide titles and attributes in local languages? - Should the results page show the shopper how the query was interpreted, or should the interpretation stay invisible? ### What a Strong Answer Covers - A diagnosis of why such queries fail: local words that never appear in English listings, many spellings of the same transliterated word, and descriptive words such as colors that are not linked to product attributes. - A query-understanding plan that normalizes spellings, maps local terms to catalog concepts and separates the product type from attributes like color. - How the interpreted query changes retrieval and ranking, including catalog attribute quality, behavioral signals and a safe fallback when the system is unsure. - Where training and evaluation data comes from, especially for languages with little search history. - Metrics and an experiment scoped to local-language queries, with guardrails that protect every other search. ### Follow-up Questions - "Laal" could also appear in a brand or product name. How would you avoid a rewrite that hides the products the shopper actually meant? - How would you extend the approach to a new language where you have almost no search history? - Local-language queries may be a small share of all searches. How would you show in an A/B test that your change helped the shoppers who type them?