The AI behind cross-lingual search at Walmart

As a capability of the Walmart Translation Platform, Walmart’s cross-lingual search system uses AI trained on Walmart-specific data to interpret multilingual queries and return relevant results in real time.

When you shop online, you expect the search bar to just work — type a few words, get the right results.

But if the site only understands one language, and it’s not yours, the system can’t interpret your intent. That’s a common barrier to online shopping.

Research shows that 40% of shoppers avoid websites that aren’t in their preferred language, a challenge that is amplified in multilingual markets such as Canada, where 22% speak French as a first language, and the U.S., where 13% speak Spanish at home. For Walmart, solving this is essential for delivering an inclusive, customer-first experience.

The complexities of translation in online search

While it may seem like a translation tool could solve the problem, retail search introduces complexities that make multilingual search particularly difficult for generic machine translation (MT) systems:

• Ambiguous queries: Retail queries are often short and lack context, which makes it difficult for MT systems to determine customer intent.
• Polysemy: Some words have multiple meanings and short queries rarely provide enough context for a model to infer the right one. For example, “pêche” in French can mean “peach” or “fishing.” 
• Non-translatable entities: Brands, product names and model numbers must remain unchanged. A yogurt brand like Liberté should never become “freedom yogurt.” 
• Code-switching: A query like “cake de fresa” mixes languages and requires adaptive translation logic. MT systems trained on monolingual data often misinterpret these hybrid structures.
• Speed: Online shoppers expect results in the blink of an eye. Generic translation tools are not optimized for real-time inference at the scale Walmart operates.

Without a system engineered specifically for retail, search results are often inaccurate, irrelevant or slow, leading to customer frustration and missed opportunities for retailers.

How Walmart’s cross-lingual search works 

To address the complexities of multilingual search, Walmart developed a system designed to interpret queries across languages, understand intent, and retrieve the right products instantly. Rather than treating translation as a standalone task, it integrates catalog knowledge, language detection, and AI models to understand each query in context and return relevant results in under 50 milliseconds.

By training models on Walmart-specific data, including product categories, brand names, and customer behavior, the system adapts to how people shop and how products are organized, enabling far more accurate and localized interpretations than generic translation tools.

• Translation Memory: Many popular items are searched for repeatedly. An LLM-powered cache precomputes translations for high-frequency queries, covering up to 95% of search traffic in some markets. This enables near-instant lookups and highly consistent results.
• Smarter language detection and handling variations: The system identifies the language and dialect of each query, detecting regional phrasing, spelling variations, and mixed-language queries. This ensures localized accuracy. 
• Handling entities and resolving ambiguity: The system identifies brand names, model numbers, product lines, and other fixed terms, and determines which parts of a query should be preserved. It also interprets short or ambiguous queries using available context, so the model has a clearer understanding of what the customer most likely means before translation begins.
• Fine-tuned models: For new or rare queries, a neural machine translation (NMT) model fine-tuned for Walmart’s data generates the final translation. A contextual rule engine reinforces the pre-translation decisions by ensuring that brand names and product identifiers remain intact during translation and are not overwritten by the model.

Together, these components deliver speed, accuracy and scalability, with measurable boosts in search relevance, precision, and conversion rates.

Cross-lingual search is one component of a broader transformation. As we expand into new languages and markets, the system will continue to evolve, becoming the linguistic backbone for search, content, associate tools, and customer experiences across the company. By combining cutting-edge AI with human expertise, we’re building a multilingual foundation that scales globally and removes friction from our customers’ shopping journeys.