Encoder-Decoder Network with Cross-Match Mechanism for Answer Selection

Abstract

Answer selection (AS) is an important subtask of question answering (QA) that aims to choose the most suitable answer from a list of candidate answers. Existing AS models usually explored the single-scale sentence matching, whereas a sentence might contain semantic information at different scales, e.g. Word-level, Phrase-level, or the whole sentence. In addition, these models typically use fixed-size feature vectors to represent questions and answers, which may cause information loss when questions or answers are too long. To address these issues, we propose an Encoder-Decoder Network with Cross-Match Mechanism (EDCMN) where questions and answers that represented by feature vectors with fixed-size and dynamic-size are applied for multiple-perspective matching. In this model, Encoder layer is based on the “Siamese” network and Decoder layer is based on the “matching-aggregation” network. We evaluate our model on two tasks: Answer Selection and Textual Entailment. Experimental results show the effectiveness of our model, which achieves the state-of-the-art performance on WikiQA dataset.