RecInDial: A Unified Framework for Conversational Recommendation with Pretrained Language Models

Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 489–500 November 20–23, 2022. ©2022 Association for Computational Linguistics 489 RecInDial: A Unified Framework for Conversational Recommendation with Pretrained Language Models Lingzhi Wang1,2∗, Huang Hu4, Lei Sha3, Can Xu4, Kam-Fai Wong1,2, Daxin Jiang4† 1The Chinese University of Hong Kong, Hong Kong, China 2MoE Key Laboratory of High Confidence Software Technologies, China 3University of Oxford, United Kingdom 4Microsoft Corporation, Beijing, China 1,2{lzwang,kfwong}@se.cuhk.edu.hk; 3lei.sha@cs.ox.ac.uk; 4{huahu,caxu,djiang}@microsoft.com Abstract Conversational Recommender System (CRS), which aims to recommend high-quality items to users through interactive conversations, has gained great research interest recently. A CRS is usually composed of a recommendation mod- ule and a generation module. In the previous work, these two modules are loosely connected in the model training and are shallowly inte- grated during inference, where a simple switch- ing or copy mechanism is adopted to incor- porate recommended items into generated re- sponses. Moreover, the current end-to-end neu- ral models trained on small crowd-sourcing datasets (e.g., 10K dialogs in the ReDial dataset) tend to overfit and have poor chit-chat ability. In this work, we propose a novel unified framework that integrates recommendation into the dialog (RecInDial1) generation by introduc- ing a vocabulary pointer. To tackle the low- resource issue in CRS, we finetune the large- scale pretrained language models to generate fluent and diverse responses, and introduce a knowledge-aware bias learned from an entity- oriented knowledge graph to enhance the rec- ommendation performance. Furthermore, we propose to evaluate the CRS models in an end- to-end manner, which can reflect the overall performance of the entire system rather than the performance of individual modules, compared to the separate evaluations of the two modules used in previous work. Experiments on the benchmark dataset ReDial show our RecInDial model significantly surpasses the state-of-the- art methods. More extensive analyses show the effectiveness of our model. 1 Introduction In recent years, there have been fast-growing re- search interests to address Conversational Rec- ommender System (CRS) (Li et al., 2018; Sun ∗Work performed during internship at Microsoft STCA. †Corresponding author: djiang@microsoft.com. 1The code is available at https://github.com/ Lingzhi-WANG/PLM-BasedCRS and Zhang, 2018; Zhou et al., 2020a), due to the booming of intelligent agents in e-commerce plat- forms. It aims to recommend target items to users through interactive conversations. Traditional rec- ommender systems perform personalized recom- mendations based on user’s previous implicit feed- back like clicking or purchasing histories, while CRS can proactively ask clarification questions and extract user preferences from conversation history to conduct precise recommendations. Existing gen- erative methods (Chen et al., 2019; Zhou et al., 2020a; Ma et al., 2020; Liang et al., 2021) are generally composed of two modules, i.e., a rec- ommender module to predict precise items and a dialogue module to generate free-form natural re- sponses containing the recommended items. Such methods usually utilize Copy Mechanism (Gu et al., 2016) or Pointer Network (Gulcehre et al., …

Content truncated to 500 words