Learning a Query Parser for Local Web Search

Parsing unstructured local web queries is often tackled using simple syntactic rules that tend to be limited and brittle. Here we present a data-driven approach to learning a query parser for local-search (geographical) queries. The learnt model uses class-level ngram language model-based features; these ngram language models, harvested from structured queries logs, insulate the model from surface-level tokens. The proposed approach is compared with a finite state model. Experiments show significant improvements for parsing geographical web queries using these learnt models.