GEMINIO: Finding duplicates in a question haystack

Abstract

Effective reuse of existing crowdsourced intelligence present in Community Question Answering (CQA) forums requires efficient approaches for the problem of Duplicate Question Detection (DQD). Approaches which use standalone encoded representations for each of the questions in the question pair fail to use the cross question interactions between the two questions which impacts their performance adversely. In this paper, we propose two new schemes for DQD task. Our first approach leverages semantic relations and our second approach utilizes fine grained word level interactions across the two question sentences. We achieve test accuracy of 75.7% and 77.8% with our first and second approaches respectively on a publicly available DQD data set, demonstrating that cross question analysis information can help aid DQD task performance.