Self-Supervised learning for Conversational Recommendation

Conversational recommender system (CRS) aims to model user preference through interactive conversations. Although there are some works, they still have two drawbacks: (1) they rely on large amounts of training data and suffer from data sparsity problem; and (2) they do not fully leverage different types of knowledge extracted from dialogues. To address these issues in CRS, we explore the intrinsic correlations of different types of knowledge by self-supervised learning, and propose the model SSCR, which stands for Self-Supervised learning for Conversational Recommendation. The main idea is to jointly consider both the semantic and structural knowledge via three self-supervision signals in both recommendation and dialogue modules. First, we carefully design two auxiliary self-supervised objectives: token-level task and sentence-level task, to explore the semantic knowledge. Then, we extract the structural knowledge based on external knowledge graphs from user mentioned entities. Finally, we model the inter-information between the semantic and structural knowledge with the advantages of contrastive learning. As existing similarity functions fail to achieve this goal, we propose a novel similarity function based on negative log-likelihood loss. Comprehensive experimental results on two real-world CRS datasets (including both English and Chinese with about 10,000 dialogues) show the superiority of our proposed method. Concretely, in recommendation, SSCR gets an improvement about 5% similar to 15% compared with state-of-the-art baselines on hit rate, mean reciprocal rank and normalized discounted cumulative gain. In dialogue generation, SSCR outperforms baselines on both automatic evaluations (distinct n-gram, BLEU and perplexity) and human evaluations (fluency and informativeness).