Abstract: The core of target-guided open-domain conversation is topic sequence planning and response retrieval according to the dialogue history and current topic, in that way the conversation reaches the target topic. The extraction and modeling method of topic features affects the accuracy of topic prediction, thereby affecting the performance of response retrieval. The current common method is to introduce external knowledge to enhance topic semantics, but this method relies on the quality of external knowledge. This paper proposes a topic semantics enhancement target-guided conversation model based on graph convolution and subgraph attention. The model firstly constructs an undirected graph of keywords, and uses a graph convolutional neural network to encode keywords to enhance topic semantics. At the same time, the dialogue history features are encoded, and the dialogue features and topic features are used for prediction and retrieval tasks. After each prediction task, the attention weight of the topic subgraph related to the current topic is calculated to update the sub-graph node representation, which is used to obtain information that is closer to the target topic. Extensive experiments on two real chat datasets show that the model can construct a coherent conversation sequence and achieve the conversation goals with a high achievement rate.

Key words: dialogue system, topic-guided, graph convolutional network, semantic matching

摘要： 目标导向式开放域对话的核心是根据对话历史与当前话题进行话题序列规划与回复检索，使得对话达到目标话题，话题特征的提取与建模方式影响话题预测的准确性，从而影响回复检索性能。目前常见的方法是引入外部知识来增强语义，但这种方法依赖外部知识质量。提出一种基于图卷积与子图注意力的话题语义增强的对话引导模型（KWGE），该模型首先构建关键词无向图，利用图卷积神经网络对关键词编码以增强话题语义，同时编码对话历史特征，使用对话特征与话题特征做预测与检索任务，每次预测任务后计算与当前话题相关话题子图的注意力权重来更新子图节点表示，用于获取更近于目标的话题。通过对两个真实闲聊数据集进行的广泛实验，表明该模型可以构建话题连贯的对话序列，并以较高的目标达成率实现对话目标。

关键词: 对话系统, 话题引导, 图卷积神经网络, 语义匹配