A BERT-Based Semantic Matching Ranker for Open-Domain Question Answering

Open-domain question answering (QA) is a hot topic in recent years. Previous work has shown that an effective ranker can improve the overall QA performance by denoising irrelevant context. There are also some recent works leveraged BERT pre-trained model to tackle with open-domain QA tasks, and achieved significant improvements. Nevertheless, these BERT-based models simply concatenates a paragraph with a question, ignoring the semantic similarity of them. In this paper, we propose a simple but effective BERT-based semantic matching ranker to compute the semantic similarity between the paragraph and given question, in which three different representation aggregation functions are explored. To validate the generalized performance of our ranker, we conduct a series of experiments on two public open-domain QA datasets. Experimental results demonstrate that the proposed ranker contributes significant improvements on both the ranking and the final QA performances.