AMQAN: Adaptive Multi-Attention Question-Answer Networks for Answer Selection

Community Question Answering (CQA) provides platforms for users with various background to obtain information and share knowledge. In the recent years, with the rapid development of such online platforms, an enormous amount of archive data has accumulated which makes it more and more difficult for users to identify desirable answers. Therefore, answer selection becomes a very important subtask in Community Question Answering. A posted question often consists of two parts: a question subject with summarization of users’ intention, and a question body clarifying the subject with more details. Most of the existing answer selection techniques often roughly concatenate these two parts, so that they cause excessive noises besides useful information to questions, inevitably reducing the performance of answer selection approaches. In this paper, we propose AMQAN, an adaptive multi-attention question-answer network with embeddings at different levels, which makes comprehensive use of semantic information in questions and answers, and alleviates the noise issue at the same time. To evaluate our proposed approach, we implement experiments on two datasets, SemEval 2015 and SemEval 2017. Experiment results show that AMQAN outperforms all existing models on two standard CQA datasets.