venue: WSDM 2020 (Demonstration)

demo link: https://qsearch.mpi-inf.mpg.de/

Traditional search engines do no understand quantities and often fail to return results expected to satisfy certain quantity conditions in the query.

This paper introduces Qsearch, which is originally proposed in “ Qsearch: Answering
quantity queries from text (ISWC 2019)” . The overview framework is shown in Figure 1 which consists of two phases: Extract and Answer.

1. Extract

The 1st step is to recognize entities with AIDA and quantity mentions with Illinois Quantifier. Then, LSTM is used to extract triples like (entity, quantity, context), called Qfacts. For each quantity, the LSTM identifies the referred entity and relevant context tokens. For the example in Figure 2, q1 and q2 are both linked to e1. As a result, we can find all the relevant quantities and contexts for a single entity (block 3 in Figure 1).

2. Answer

First, a query is parsed into (answer type, quantity condition, rest tokens in the query), or (t,q,X). The query parser uses YAGO types and predifined quantity units to recognize t and q.

To match the answer, given the identified answer type and quantity condition, a list of candidate entities and their quantity-context pairs are returned. A list of unit conversion rules are used to filter Qfacts with the unmatched quantity condition. After that, each candidate entity is assgined a score based on the context matching. Finally, the candidate entities are ranked by the matching scores.