计算机工程与应用 ›› 2023, Vol. 59 ›› Issue (3): 158-165.DOI: 10.3778/j.issn.1002-8331.2108-0354

• 模式识别与人工智能 • 上一篇    下一篇

面向多阶段匹配的答案选择模型

陈建贵,张儒清,郭嘉丰,范意兴   

  1. 1.中国科学院 计算技术研究所 网络数据科学与技术重点实验室,北京 100190
    2.中国科学院大学,北京100190
  • 出版日期:2023-02-01 发布日期:2023-02-01

Answer Selection Model for Multi-Stage Matching

CHEN Jiangui, ZHANG Ruqing, GUO Jiafeng, FAN Yixing   

  1. 1.Key Lab of Network Data Science and Technology, Institute of Computing Technology, Chinese Academy of Sciences, Beijing 100190, China
    2.University of Chinese Academy of Sciences, Beijing 100190, China
  • Online:2023-02-01 Published:2023-02-01

摘要: 近年来,信息量成倍增长,获取有效信息的代价越来越高,答案选择技术能够为用户直接提供所需的信息,具有革命性的意义。给定问题和候选答案,答案选择任务要求从候选答案中找出与问题最相关的答案。不失一般性,候选答案根据与问题的匹配程度可以分为三种类型:不相关、相关不合理、相关且合理。然而,已有工作仅考虑问题与答案的相关性,这对于精准问答是远远不够的。为此,提出多阶段匹配模型(MSMM),模拟人的答题过程。具体的,MSMM模型分为两个阶段,第一阶段先将简单易解决的问答对分离出去,第二阶段再综合推理复杂的问答数据。每一阶段都由嵌入层、编码层、对齐层、融合层和池化层组成。此外,为了增强模型的推理能力,还引入语义角色标注信息和单词相似矩阵信息。为了便于评估,基于WikiQA和InsuranceQA数据集构造了两个答案合理性数据集。实验结果表明,对比基准方法,该模型在性能上取得一致的提升。

关键词: 答案选择, 问答系统, 知识增强

Abstract: In recent years, the cost of obtaining effective information is higher and higher. Answer selection(AS) can directly provide users with the information they need, which has revolutionary significance. Given a question and a list of candidate answers, AS requires to find the most relevant answer from the candidate answers. Without loss of generality, candidate answers can be divided into three types according to the degree of matching with the question:irrelevant, relevant and unreasonable, relevant and reasonable. However, the existing work only considers the relevance between questions and answers. Therefore, this paper proposes a multi-stage matching model(MSMM) to simulate the answering process of human. Specifically, the MSMM model is divided into two stages. In the first stage, the easy samples are separated, and in the second stage, the complex samples are synthesized. Each stage consists of embedding layer, encoding layer, alignment layer, fusion layer, and pooling layer. In addition, to enhance the reasoning ability of the model, this paper also introduces semantic role labeling and word similarity matrix. To facilitate the evaluation, this paper builds a benchmark based on WikiQA and InsuranceQA datasets. Empirical studies demonstrate that the proposed model can achieve consistent performance improvement compared with baselines.

Key words: answer selection, question answering, knowledge enhancement