A user searching for “bbq near charlotte” probably wants nearby barbecue restaurants, not reviews containing all three words. Query interpretation turns those few words into a more useful search request by identifying domain meanings, user intent, and constraints such as location. Without it, literal matching can return an irrelevant result while missing restaurants whose reviews never say “near.” section 7.2 section 2.5.2
Queries need different handling from documents: they are usually short and often imply more than they state. Useful context can come from phrases in the document corpus and term co-occurrence in previous search sessions. Domain knowledge may turn separate keywords into a phrase, add synonyms, or correct a misspelling. Keep keyword matching as a fallback when an interpretation is uncertain or a term is unknown. section 2.5.4 section 2.5.2
An end-to-end interpretation pipeline has four phases:
- Parse the query to identify known terms, phrases, entities, and intent. section 7.5
- Enrich it with learned context, such as alternative spellings, related terms, and knowledge-graph relationships. section 7.5
- Transform that interpretation into filters, expansions, and ranking instructions for the target search engine. section 7.0
- Run the optimized search and return ranked results; phases may repeat or be skipped when appropriate. section 7.0
From “top kimchi near charlotte” to search constraints
A literal search finds no review containing every keyword. An interpreted search instead uses “top” to favor high-rated reviews, expands “kimchi” with related terms such as “korean” and “banchan” and a Korean category constraint, and treats “near charlotte” as a 50 km geographic filter. The resulting search returns highly rated local results, including relevant restaurants whose reviews do not say “kimchi.” This works only when the learned associations and location interpretation fit the user’s intent. section 7.3 section 7.5
who is the CEO? support iphone screen blacked out iphone verizon silver iphone 8 plus 64GB sale refrigerators pay my bill
Compare these short queries: some suggest a destination, some a product or filter, and others an action. Matching their words alone does not settle what the user wants. section 2.5.5
