[Paper Review] Towards Topic-Guided Conversational Recommender System

Kun Zhou, Yuanhang Zhou+3 more|arXiv (Cornell University)|Oct 8, 2020

Topic Modeling36 references18 citations

TL;DR

This paper introduces TG-ReDial, a new conversational recommender system dataset featuring natural topic-guided transitions from chit-chat to recommendations, created via semi-automatic human annotation on real movie review data. The authors propose a unified model that jointly predicts topics, recommends items, and generates responses using BERT, SASRec, and GPT-2, achieving state-of-the-art performance across all three sub-tasks on the TG-ReDial benchmark.

ABSTRACT

Conversational recommender systems (CRS) aim to recommend high-quality items to users through interactive conversations. To develop an effective CRS, the support of high-quality datasets is essential. Existing CRS datasets mainly focus on immediate requests from users, while lack proactive guidance to the recommendation scenario. In this paper, we contribute a new CRS dataset named extbf{TG-ReDial} ( extbf{Re}commendation through extbf{T}opic- extbf{G}uided extbf{Dial}og). Our dataset has two major features. First, it incorporates topic threads to enforce natural semantic transitions towards the recommendation scenario. Second, it is created in a semi-automatic way, hence human annotation is more reasonable and controllable. Based on TG-ReDial, we present the task of topic-guided conversational recommendation, and propose an effective approach to this task. Extensive experiments have demonstrated the effectiveness of our approach on three sub-tasks, namely topic prediction, item recommendation and response generation. TG-ReDial is available at https://github.com/RUCAIBox/TG-ReDial.

Motivation & Objective

• To address the lack of natural, topic-guided transitions from non-recommendation to recommendation scenarios in existing conversational recommender system (CRS) datasets.
• To develop a high-quality, realistic CRS dataset that supports proactive, contextually grounded recommendation flows through topic evolution.
• To propose a unified framework for topic-guided conversational recommendation that integrates topic prediction, item recommendation, and response generation.
• To enable more effective modeling of user preferences by jointly leveraging dialog history, user profiles, and evolving topic threads.

Proposed method

• The TG-ReDial dataset is constructed by aligning user identities from a real movie review website with synthetic dialogues, ensuring realistic user profiles and recommendation reasons.
• Topic threads are explicitly designed to guide conversations from non-recommendation topics toward final item recommendations, enabling natural semantic transitions.
• The proposed model uses BERT to encode historical utterances, topics, and user profiles, with separate BERT encoders for each modality to capture diverse user preference signals.
• The item recommendation module combines SASRec for sequential behavior modeling and BERT for contextual understanding of dialog history.
• The response generation module is based on fine-tuned GPT-2, conditioned on predicted topics and items to produce contextually relevant and persuasive responses.
• The model is jointly trained on three sub-tasks—topic prediction, item recommendation, and response generation—using multi-task learning to improve overall performance.

Experimental results

Research questions

• RQ1 Can topic-guided transitions from chit-chat to recommendation improve the naturalness and effectiveness of conversational recommendation?
• RQ2 How well can a unified model jointly predict topics, recommend items, and generate responses in a conversational setting?
• RQ3 To what extent does incorporating user profiles, dialog history, and evolving topics improve recommendation quality compared to isolated signals?
• RQ4 How does the proposed model compare to strong baselines in terms of topic prediction, item recommendation, and response generation?

Key findings

• The proposed model outperforms all baselines in topic prediction, achieving the highest F1 score by effectively integrating historical utterances, topics, and user profiles through separate BERT encoders.
• In item recommendation, the model achieves the best performance by combining sequential preference modeling (SASRec) with contextual dialog understanding (BERT).
• For response generation, the model achieves the lowest perplexity (7.223) and highest human evaluation scores (1.520 for relevance, 1.640 for informativeness), significantly outperforming GPT-2 and other baselines.
• Ablation studies confirm that the target topic is a critical input for performance, as removing it from BERT inputs leads to a significant drop in all metrics.
• The semi-automatic data creation process results in more realistic and controllable human-annotated dialogues compared to fully human- or fully automatic-based methods.