CLICK: Contrastive Learning for Injecting Contextual Knowledge to Conversational Recommender System

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 1875–1885 May 2-6, 2023 ©2023 Association for Computational Linguistics CLICK: Contrastive Learning for Injecting Contextual Knowledge to Conversational Recommender System Hyeongjun Yang, Heesoo Won, Youbin Ahn, and Kyong-Ho Lee Department of Computer Science, Yonsei University {edbm95, hswon97, ybahn, khlee98}@yonsei.ac.kr Abstract Conversational recommender systems (CRSs) capture a user preference through a conversa- tion. However, the existing CRSs lack captur- ing comprehensive user preferences. This is because the items mentioned in a conversation are mainly regarded as a user preference. Thus, they have limitations in identifying a user pref- erence from a dialogue context expressed with- out preferred items. Inspired by the charac- teristic of an online recommendation commu- nity where participants identify a context of a recommendation request and then comment with appropriate items, we exploit the Reddit data. Specifically, we propose a Contrastive Learning approach for Injecting Contextual Knowledge (CLICK) from the Reddit data to the CRS task, which facilitates the capture of a context-level user preference from a dia- logue context, regardless of the existence of preferred item-entities. Moreover, we devise a relevance-enhanced contrastive learning loss to consider the fine-grained reflection of mul- tiple recommendable items. We further de- velop a response generation module to gener- ate a persuasive rationale for a recommenda- tion. Extensive experiments on the benchmark CRS dataset show the effectiveness of CLICK, achieving significant improvements over state- of-the-art methods. 1 Introduction Recommender systems help users find potential items of interest in the rapidly expanding pool of candidates. To do so, the recommender sys- tems aim to accurately identify user preferences. Thus, traditional recommender systems (Guo et al., 2020) utilize click or purchase data to obtain user preferences. However, such traditional approaches lack dynamically modeling user preferences since the data has a temporal validity and does not of- fer a diverse clue to user information. Conversa- tional recommender system (CRS) is a research area that overcomes the limitation of the traditional User I am stressed out from work today. So, I’m looking for a mindless movie. Recommender You must be exhausted! Have you ever watched “21 Jump Street”? It is so much fun. Oh wait, I have a little son who wants to watch a movie together. Then, I recommend “Yes Man”. It’s a family movie. Your son also will like it. Recommender Hello, how are you today? Recommender User Figure 1: An example of a user-recommender conver- sation for movie recommendations. The important in- formation for a recommendation is written in red, and the recommended items are written in blue. recommender systems by capturing users’ interests through natural conversation. The conventional CRSs (Chen et al., 2019; Ma et al., 2021; Zhou et al., 2020a) mainly derive a user preference from item-entities, which appear in a conversation and exist in a knowledge graph (KG). However, users often express their preferences without preferred items. For example, in Figure 1, the user depicts his/her requirement (a mindless movie), feeling (stressed), and situation …

Content truncated to 500 words