Knowledge Aware Conversation Generation with Explainable Reasoning over Augmented Graphs

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 1782–1792, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 1782 Knowledge Aware Conversation Generation with Explainable Reasoning over Augmented Graphs Zhibin Liu, Zheng-Yu Niu, Hua Wu, Haifeng Wang Baidu Inc., Beijing, China {liuzhibin05,niuzhengyu,wu hua, wanghaifeng}@baidu.com Abstract Two types of knowledge, triples from knowl- edge graphs and texts from documents, have been studied for knowledge aware open- domain conversation generation, in which graph paths can narrow down vertex candi- dates for knowledge selection decision, and texts can provide rich information for response generation. Fusion of a knowledge graph and texts might yield mutually reinforcing advan- tages, but there is less study on that. To address this challenge, we propose a knowl- edge aware chatting machine with three com- ponents, an augmented knowledge graph with both triples and texts, knowledge selector, and knowledge aware response generator. For knowledge selection on the graph, we formu- late it as a problem of multi-hop graph reason- ing to effectively capture conversation flow, which is more explainable and flexible in com- parison with previous work. To fully leverage long text information that differentiates our graph from others, we improve a state of the art reasoning algorithm with machine reading comprehension technology. We demonstrate the effectiveness of our system on two datasets in comparison with state-of-the-art models1. 1 Introduction One of the key goals of AI is to build a machine that can talk with humans when given an initial topic. To achieve this goal, the machine should be able to understand language with background knowledge, recall knowledge from memory or ex- ternal resource, reason about these concepts to- gether, and finally output appropriate and informa- tive responses. Lots of research efforts have been devoted to chitchat oriented conversation genera- tion (Ritter et al., 2011; Shang et al., 2015). 1Data and codes are available at https://github. com/PaddlePaddle/models/tree/develop/ PaddleNLP/Research/EMNLP2019-AKGCM However, these models tend to produce generic responses or incoherent responses for a given topic, since it is quite challenging to learn seman- tic interactions merely from dialogue data without help of background knowledge. Recently, some previous studies have been conducted to introduce external knowledge, either unstructured knowl- edge texts (Ghazvininejad et al., 2018; Vougiouk- lis et al., 2016) or structured knowledge triples (Liu et al., 2018; Young et al., 2018; Zhou et al., 2018) to help open-domain conversation genera- tion by producing responses conditioned on se- lected knowledge. In the first research line, their knowledge graph can help narrowing down knowledge candidates for conversation generation with the use of prior information, e.g., triple attributes or graph paths. Moreover, these prior information can enhance generalization capability of knowledge selection models. But it suffers from information insuffi- ciency for response generation since there is sim- ply a single word or entity to facilitate genera- tion. In the second line, their knowledge texts, e.g., comments about movies, can provide rich in- formation for generation, but …

Content truncated to 500 words