Utilizing SBERT For Finding Similar Questions in Community Question Answering

The BERT model was fine-tuned to give state-of-the-art results in sentence-pair regressions. However, this model requires that both questions are fed into the network, which leads to a massive computational overhead. Instead of computing on n pairs of sentences, SBERT was proposed to learn sentence representation by computing on only one query question. This model was proven to work effectively on semantic textual similarity (STS). In this paper, we explore SBERT model for question retrieval in Community Question Answering. Results show that SBERT decreases slightly in performance compared to BERT4ECOMMERCE. However, This model reduces the effort for finding the most similar question from 795 seconds with BERT to about 0.828 seconds with SBERT, while maintaining the accuracy from BERT.