基于图神经网络的多源异构知识增强对话模型

Multi-source Heterogeneous Knowledge Enhanced Dialogue Model Based on Graph Neural Network

为解决当前开放域对话系统中端到端模型在生成响应时存在的相关性低、多样性不足的问题,提出了一种多源异构知识增强对话生成模型(multi-source knowledge-enhanced dialogue generation framework,MSGF)。该模型通过整合多个不同的知识源,提高了与对话背景信息相关的知识覆盖率,并采用全局知识选择模块解决不同知识源之间的主题冲突问题,来避免对话主题含义混淆。此外,该模型还引入了融合预测模块,通过获取不同的知识源中的信息来生成响应。实验结果表明,与同类其他模型相比,MSGF模型在性能上具有明显优势,具有更全面的知识覆盖,生成的响应主题相关性更高。可见,所提出的MSGF模型能够很好地理解对话内容,并显著提升对话系统的性能。

In order to address the problems of low relevance and lack of diversity of end-to-end models in generating responses in current open-domain dialogue systems,a multi-source knowledge-enhanced dialogue generation framework(MSGF) was used to investigate various aspects of dialogue generation.Firstly,multiple different knowledge sources were integrated to improve the knowledge coverage related to dialogue background information.Secondly,the adoption of the global knowledge selection module can avoid topic conflicts between different knowledge sources,thereby eliminating the problem of confusion in the meaning of dialogue topics.In addition,a fusion prediction module was also introduced to generate responses by obtaining information from different knowledge sources.The results show that the MSGF model outperforms other similar models with more comprehensive knowledge coverage and higher topic relevance of generated responses.It is concluded that the proposed MSGF model is capable of understanding the content of dialogue conversations and significantly improving the performance of dialogue systems.