Towards Topic-Guided Conversational Recommender System

• Existing CRS datasets largely assume clear, immediate requests and underrepresent proactive transitions from non-recommendation conversation to recommendation.Source: §1 · p. 1
• TG-ReDial contains 10,000 two-party movie-domain dialogues between seekers and recommenders.Source: §1 · p. 1
• Topic threads guide each conversation from an initial non-recommendation topic toward recommendation through evolving topics and chit-chat.Source: §1 · p. 2
• Real user data supports recommended movies, topic threads, recommendation reasons, and user identities, while annotators revise, polish, or rewrite conversation data.Source: §1 · p. 2, Figure 1 · p. 2
• The paper releases TG-ReDial and proposes a Transformer-based solution that leverages multiple data signals across the three subtasks.Source: §1 · p. 2
• Topic-guided conversational recommendation comprises item recommendation, topic prediction, and response generation.Source: §1 · p. 2

• Conversation systems include task-oriented systems for specific goals and chit-chat systems for general-purpose dialogue.Source: §2.1 · p. 3
• Existing CRS datasets include synthetic template-based resources and human-annotated datasets such as ReDial, GoReDial, and DuRecDial.Source: §2.3 · p. 3
• DuRecDial uses goal sequences spanning multiple task types, whereas TG-ReDial uses topic threads to characterize evolving content flow.Source: §2.3 · p. 3
• TG-ReDial is designed to be easier to integrate with open-domain dialogue because its topic threads describe content evolution rather than task switching.Source: §2.3 · p. 3
• Unlike DuRecDial, TG-ReDial derives user-related data from real users rather than relying on annotators to generate it.Source: §2.3 · p. 3

Dataset setting

• The two-party setting begins with chit-chat, proactively moves toward a target topic, and then recommends a movie using the seeker’s interests.Source: §3 · p. 3

Collection process

• Three movies sharing the “family” tag illustrate how film records, ConceptNet topic threads, and candidate sentences are combined during collection.Source: Figure 2 · p. 4

Real-world data

• Conversations are associated with real Douban users, enabling watching records, user profiles, and review-derived recommendation reasons to support recommendation.Source: §3 · p. 4

Movie and tag preparation

• Movies are grouped into coherent watching subsequences and annotated with meaningful tags drawn from original metadata, reviews, and manual selection.Source: §3.1 · p. 4

Creating topic threads

• Topic threads connect an initial greeting to a recommended movie’s target tag by traversing ConceptNet and using shortest paths identified with depth-first search.Source: §3.2 · pp. 4–5

Human annotation

• Annotators generate utterances from retrieved topic-relevant candidates and revise them for dialogue consistency, while recommendation reasons come from retrieved review sentences.Source: §3.3 · p. 5

Dataset statistics

• TG-ReDial contains 129,392 utterances from 1,482 users; dialogues average 7.9 topics and 19.0 words per utterance.Source: §3.4 · pp. 5–6

Privacy

• Privacy protection restricts sampling to users with many watching records and anonymizes or modifies derived user data before excluding identifiable users.Source: §3.4 · p. 6

Task formulation

• The task uses user profile, interaction history, prior utterances, and topics to predict the next topic, recommend an item, and generate a response.Source: §4.1 · p. 6

Recommendation module

• The recommendation module combines BERT representations of historical utterances with SASRec representations of the user interaction sequence.These signals form the user representation used for recommendation.Source: §4.2 · p. 6
• The recommendation module ranks items using their learned embeddings and the user representation, selecting the item with the largest probability.Source: §4.2 · pp. 6–7

Dialog module

• The dialog module generates responses for topic guidance or item recommendation using specific models, including GPT-2 for response generation.Source: §4.3 · p. 7
• The topic prediction model uses conversation-BERT, topic-BERT, and profile-BERT to encode historical utterances, topic history, and the user profile.The model ranks candidate topics and selects the one with the largest probability.Source: §4.3 · p. 7

Experimental Setup

• TG-ReDial is split into training, validation, and test sets with an 8:1:1 ratio, and evaluation covers item recommendation, topic prediction, and response generation.The model generates replies or recommendations turn by turn from each conversation's first utterance.Source: §5 · p. 7

Item Recommendation

• Item recommendation uses NDCG@k and MRR@k with k = 10 and 50 to rank all possible items.Source: §5.1 · p. 7
• The proposed model significantly outperforms all item-recommendation baselines by combining historical utterances with interaction sequences.The comparison includes CRS, interaction-history, and text-based baselines; BERT outperforms TextCNN among the text-based models.Source: §5.1 · p. 8

Topic Prediction

• Topic prediction uses Hit@k with k = 1, 3, and 5, comparing heuristic, recurrent, BERT-based, and target-topic ablation baselines.The task predicts the next topic leading toward the recommendation target.Source: §5.2 · p. 8
• The proposed topic-prediction model outperforms all baselines, while removing the target topic significantly decreases performance.The model jointly encodes historical utterances, topics, and user profiles with different BERT models.Source: §5.2 · p. 8

Response Generation

• For response generation, the proposed model outperforms all baselines in most cases, especially for perplexity, Distinct, and human evaluation.Transformer, GPT-2, and the proposed model have similar BLEU scores, while predicted topics or items enhance generated text quality in the proposed approach.Source: §5.3 · p. 9

• TG-ReDial is a high-quality conversational-recommender dataset constructed through human annotation based on real-world user data.Source: §6 · p. 9
• The paper presents topic-guided conversational recommendation and a solution evaluated on item recommendation, topic prediction, and response generation.Source: §6 · p. 9
• The proposed approach demonstrates effectiveness on all three sub-tasks in extensive experiments.Source: §6 · p. 9
• TG-ReDial's potential has not been fully explored, including possible applications to personalized chit-chat, target-guided conversation, and sequential recommendation.The authors identify broader task exploration and more effective approaches as future work.Source: §6 · p. 9