Enhancing Dialogue Generation with Conversational Concept Flows

Findings of the Association for Computational Linguistics: EACL 2023, pages 1514–1525 May 2-6, 2023 ©2023 Association for Computational Linguistics Enhancing Dialogue Generation with Conversational Concept Flows Siheng Li1∗, Wangjie Jiang1∗, Pengda Si1∗, Cheng Yang1 Yao Qiu2, Jinchao Zhang2, Jie Zhou2, Yujiu Yang1† 1Shenzhen International Graduate School, Tsinghua University 2Tencent Inc, Beijing, China {lisiheng21, jwj20, spd18}@mails.tsinghua.edu.cn yang.yujiu@sz.tsinghua.edu.cn Abstract Human conversations contain natural and rea- sonable topic shifts, reflected as the concept flows across utterances. Previous researches prove that explicitly modeling concept flows with a large commonsense knowledge graph ef- fectively improves response quality. However, we argue that there exists a gap between the knowledge graph and the conversation. The knowledge graph has limited commonsense knowledge and ignores the characteristics of natural conversations. Thus, many concepts and relations in conversations are not included. To bridge this gap, we propose to enhance dia- logue generation with conversational concept flows. Specifically, we extract abundant con- cepts and relations from natural conversations and build a new conversation-aware knowledge graph. In addition, we design a novel relation- aware graph encoder to capture the concept flows guided by the knowledge graph. Exper- imental results on the large-scale Reddit con- versation dataset indicate that our method per- forms better than strong baselines, and further analysis verifies the effectiveness of each com- ponent. 1 Introduction With the remarkable development of conversation artificial intelligence (Shang et al., 2015; Adiwar- dana et al., 2020; Thoppilan et al., 2022), response generation has been improved in many ways, e.g., human-like persona (Zhang et al., 2018a), empa- thetic expression (Rashkin et al., 2019) and knowl- edge injection (Dinan et al., 2019), etc. However, there still exists a series of challenges (Gao et al., 2019; Xu et al., 2020a; Huang et al., 2020). One of the most noticeable is that humans are good at naturally switching topics during conversations, while machine-generated responses are relatively dull and tend to keep the topic still (Fang et al., ∗ Equal contribution. † Corresponding author (yang.yujiu@sz.tsinghua.edu.cn). Figure 1: Two cases in the Reddit dataset. We use ConceptNet as the external knowledge graph to show concept flows in conversations. Concepts are marked in blue. Relations in the graph and those in the natural conversation are marked with red solid lines and blue dashed lines, respectively. 2018) or throw unexpected topics (Wang et al., 2018; Tang et al., 2019). To overcome this challenge, previous works treat the topic shifts as concept flows (Zhang et al., 2020a; Zhou et al., 2018b, 2021a), which means traversing in the concept1 space along relations in an external commonsense knowledge graph. Ex- perimental results have shown that explicitly mod- elling concept flows effectively improves the rele- vance and engagingness of responses. However, we argue that there is a gap between the external knowl- edge graph and natural conversations. The most 1Concept is the node in knowledge graph. 1514 frequently used ConceptNet 2 (Speer et al., 2017) is limited to mostly (90%) taxonomic (e.g., IsA) or lexical (e.g., Synonym) knowledge, while contains relatively small portion of commonsense …

Content truncated to 500 words