CollRec: Pre-Trained Language Models and Knowledge Graphs Collaborate to Enhance Conversational Recommendation System

Abstract

Existing conversational recommender systems (CRS) use insufficient generality in incorporating external information using knowledge graphs. The recommendation module and generation module are loosely connected during model training and shallowly integrated during inference. A simple switching or copying mechanism is used to merge recommended items into generated responses. These problems significantly degrade the recommendation performance. To alleviate this problem, we propose a novel unified framework for collaboratively enhancing conversational recommendations using pre-trained language models and knowledge graphs (CollRec). We use a fine-tuned pre-trained language model to efficiently extract knowledge graphs from conversational text descriptions, perform entity-based recommendations based on the generated graph nodes and edges, and fine-tune a large-scale pre-trained language model to generate fluent and diverse responses. Experimental results on the WebNLG 2020 Challenge dataset, ReDial dataset, and Reddit-Movie dataset show that our CollRec model significantly outperforms the state-of-the-art methods.