Qsearch: Answering Quantity Queries from Text

Quantities appear in search queries in numerous forms: companies with annual revenue of at least 50 Mio USD, athletes who ran 200m faster than 19.5 s, electric cars with range above 400 miles, and so on. Processing such queries requires the understanding of numbers present in the query to capture the contextual information about the queried entities. Modern search engines and QA systems can handle queries that involve entities and types, but they often fail on properly interpreting quantities in queries and candidate answers when the specifics of the search condition (less than, above, etc.), the units of interest (seconds, miles, meters, etc.) and the context of the quantity matter (annual or quarterly revenue, etc.). In this paper, we present a search and QA system, called Qsearch, that can effectively answer advanced queries with quantity conditions. Our solution is based on a deep neural network for extracting quantity-centric tuples from text sources, and a novel matching model to retrieve and rank answers from news articles and other web pages. Experiments demonstrate the effectiveness of Qsearch on benchmark queries collected by crowdsourcing.