Abstract

Query expansion aims to mitigate the mismatch between the language used in aquery and in a document. Query expansion methods can suffer from introducingnon-relevant information when expanding the query, however. To bridge this gap,inspired by recent advances in applying contextualized models like BERT to thedocument retrieval task, this paper proposes a novel query expansion model thatleverages the strength of the BERT model to better select relevant informationfor expansion. In evaluations on the standard TREC Robust04 and GOV2 testcollections, the proposed BERT-QE model significantly outperforms BERT-Largemodels commonly used for document retrieval.

 

Quick Read (beta)