Interactive knowledge-enhanced attention network for answer selection

Answer selection which aims to select the most appropriate answers from a set of candidate answers plays a crucial role in various applications such as question answering (QA) and information retrieval. Recently, remarkable progress has been achieved on matching sequence pairs by deep neural networks. However, most of them focus on learning semantic representations for the contexts of QA pairs while the background information and facts beyond the context are neglected. In this paper, we propose an interactive knowledge-enhanced attention network for answer selection (IKAAS), which interactively learns the sentence representations of query–answer pairs by simultaneously considering the external knowledge from knowledge graphs and textual information of QA pairs. In this way, we can exploit the semantic compositionality of the input sequences and capture more comprehensive knowledge-enriched intra-document features within the question and answer. Specifically, we first propose a context-aware attentive mechanism to learn the knowledge representations guided by the corresponding context. The relations between the question and answer are then captured by computing the question–answer alignment matrix. We further employ self-attention to capture the global features of the input sequences, which are then used to calculate the relevance score of the question and answer. Experimental results on four real-life datasets demonstrate that IKAAS outperforms the compared methods. In addition, a series of analyses shows the robust superiority and the extensive applicability of the proposed method.