TREA: Tree-Structure Reasoning Schema for Conversational Recommendation

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 2970–2982 July 9-14, 2023 ©2023 Association for Computational Linguistics TREA: Tree-structure Reasoning Schema for Conversational Recommendation Wendi Li1,2, Wei Wei1,2,�, Xiaoye Qu1, Xianling Mao3, Ye Yuan4, Wenfeng Xie4, Dangyang Chen4 1Cognitive Computing and Intelligent Information Processing (CCIIP) Laboratory, Huazhong University of Science and Technology 2Joint Laboratory of HUST and Pingan Property & Casualty Research (HPL) 3Department of Computer Science and Technology, Beijing Institute of Technology 4Ping An Property & Casualty Insurance company of China 1{wendili,weiw,xiaoye}@hust.edu.cn 3maoxl@bit.edu.cn 4{yuanye503,xiewenfeng801,chendangyang273}@pingan.com.cn Abstract Conversational recommender systems (CRS) aim to timely trace the dynamic interests of users through dialogues and generate relevant responses for item recommendations. Recently, various external knowledge bases (especially knowledge graphs) are incorporated into CRS to enhance the understanding of conversation contexts. However, recent reasoning-based models heavily rely on simplified structures such as linear structures or fixed-hierarchical structures for causality reasoning, hence they cannot fully figure out sophisticated relation- ships among utterances with external knowl- edge. To address this, we propose a novel Tree- structure Reasoning schEmA named TREA. TREA constructs a multi-hierarchical scalable tree as the reasoning structure to clarify the causal relationships between mentioned en- tities, and fully utilizes historical conversa- tions to generate more reasonable and suitable responses for recommended results. Exten- sive experiments on two public CRS datasets have demonstrated the effectiveness of our approach. Our code is available at https: //github.com/WindyLee0822/TREA 1 Introduction Conversation Recommender System (CRS) has become increasingly popular as its superiority in timely discovering user dynamic preferences in practice. As opposed to traditional passive-mode recommendation systems, it highlights the impor- tance of proactively clarifying and tracing user interests through live conversation interactions, which notably enhance the success rate of item recommendations. Since sole contextual utterances are insufficient for comprehensively understanding user prefer- ences, there are many efforts devoted to incorporat- � Corresponding Author ing various external knowledge (Chen et al., 2019; Zhou et al., 2020a, 2022; Wang et al., 2022; Yang et al., 2022), which typically enrich the contextual information with mentioned entities recognized over utterances. However, these methods fail to model the complex causal relations among men- tioned entities, owing to the diversity of user inter- est expression and the frequent shift of conversation topic as shown in Figure 1. Actually, it is non-trivial to explicitly model the complex causal relationships of conversations. Al- though there are several reasoning-based methods proposed for CRS, their simplified structures make the objective unattainable. Some researches (Zhou et al., 2021) track the mentioned entities as linear sequential fragments analogous to (1) in Figure 1. However, the linear structure is only suitable for adjacent relation modeling, which may not always work well since the actual causality between men- tioned entities exists multi-hop jumps ("comedy"- "La La Land" in Figure 1). Other studies (Ma et al., 2021) propose other forms of specially-designed structures for reasoning akin to (2) in Figure 1, but they generally have fixed hierarchies, which often degenerate into a simple …

Content truncated to 500 words