基于动态图神经网络的会话式机器阅读理解研究
Research on Conversational Machine Reading Comprehension Based on Dynamic Graph Neural Network
-
摘要: 近年来, 基于深度学习的机器阅读理解模型研究取得显著进展, 但这些模型在全局语义关系构建以及较长距离推理上仍有显著缺陷, 在对段落文本进行推理时, 大多只把文本信息看作词的序列, 而没有探索词与词之间丰富的语义关系。为了解决上述问题, 该文提出一种新的基于动态图神经网络的会话式机器阅读理解模型。首先, 提取文本中的实体, 使用句法结构与句子之间的语义关系进行建模;然后通过语义融合模块, 将基于序列化结构得到的上下文嵌入表示与基于图结构得到的实体节点嵌入表示进行融合;最终使用图神经网络实现对答案的预测。同时, 该模型可在每轮对话过程中动态地构建问题和会话历史的推理图, 能有效地捕捉对话中的语义结构信息和会话历史流程。实验结果表明, 在两个最近提出的会话挑战 (CoQA 和 QuAC)上表现了出色的性能。Abstract: Machine reading comprehension models based on deep learning have achieved remarkable success recently. However, these models have significant defects in constructing long-distance and global semantic relationships, which affect their performance in reading comprehension tasks. Moreover, when reasoning over passage text, most of them simply regard it as a word sequence without exploring rich semantic relationships between words. In order to solve this problem, this paper proposes a new system effective graph structure named Dynamic Conversational Graph Network (DCGN). Firstly, named entities are extracted from the text, and the semantic relationship between syntactic structure and sentence is used for modeling. Then, the context-embedded representation based on serialization structure and the entity node embedded representation based on graph structure are fused by semantic fusion module. Finally, dynamic graph neural network is used to realize machine reading comprehension. The model dynamically builds inference graphs of questions and session history during each round of conversation, which can effectively capture the semantic structure information and the history flow of the conversation. Experimental results show that the model performs well on two recent session challenges (CoQA and QuAC).