点击次数：

论文类型：期刊论文

发表时间：2020-08-18

发表刊物：NEUROCOMPUTING

收录刊物：SCIE

卷号：402

页面范围：80-88

ISSN号：0925-2312

关键字：Answer selection; Question answering; Fragment-level interaction; Attention

摘要：Answer selection in question answering (QA) denotes a task which selects the most appropriate one from candidate answers for a given question. Previous researches on answer selection usually conduct it by isolated word-level interaction between questions and answers. In these methods, the abundant contextual information is hardly captured, which affects the choice of the correct answer. To overcome this problem, we propose to exploit a Multiple Fragment-level Interactive Network (MFIN) for this task. The MFIN can extend the search space from word-level to fragment-level, which is conducive to obtaining more contextual information. In MFIN, we apply the multiple fragment-level attention mechanism to select key fragment pairs and achieve multiple fragment-level interaction. Meanwhile, we utilize the recurrent representation encoding to integrate multiple interactive information to reduce noise. The experimental results demonstrate that our proposed model is efficient compared to the existing methods on the WikiQA and SemEval-2016 CQA datasets. (C) 2020 Elsevier B.V. All rights reserved.