Source-linked AI summary
Towards Knowledge-Based Recommender Dialog System
Qibin Chen, Junyang Lin, Yichang Zhang, Ming Ding, Yukuo Cen, Hongxia Yang, Jie Tang
TL;DR
Recommender dialog systems must generate informative responses and high-quality recommendations while capturing users’ preferences beyond implicit feedback. KBRD integrates both systems through external knowledge and reciprocal information exchange. Experiments show better performance than baselines on recommendation and dialog generation, with analyses identifying mutual benefits between the systems.
Problem
Implicit feedback from searching, clicking, and purchasing captures only part of users’ interests, motivating systems that use explicit dialog preferences for both recommendation and response generation.
Method
KBRD is an end-to-end framework that links recommender and dialog systems through external knowledge and recommendation-aware information exchange.
Results
KBRD achieves better performance than baselines in both recommendation and dialog generation, with dialog and knowledge jointly improving recommendation most strongly.
Takeaways & Limitations
Dialog information helps recommendation especially in cold-start settings, while recommender-side user preferences and knowledge improve dialog consistency and diversity.
Takeaways & Limitations
Human consistency evaluation excluded Transformer baseline utterances because their generations differed significantly from REDIAL and model awareness could affect annotator scoring.
Abstract
from arXiv · showhide
In this paper, we propose a novel end-to-end framework called KBRD, which stands for Knowledge-Based Recommender Dialog System. It integrates the recommender system and the dialog generation system. The dialog system can enhance the performance of the recommendation system by introducing knowledge-grounded information about users' preferences, and the recommender system can improve that of the dialog generation system by providing recommendation-aware vocabulary bias. Experimental results demonstrate that our proposed model has significant advantages over the baselines in both the evaluation of dialog generation and recommendation. A series of analyses show that the two systems can bring mutual benefits to each other, and the introduced knowledge contributes to both their performances.
1 Introduction
Recommender dialog systems combine recommendation and natural-language response generation, using dialog to expose explicit preferences and ideally allowing both systems to benefit from one another. KBRD integrates them end to end with external knowledge, improving recommendation and dialog generation while enabling mutual benefits.
- Dialog provides more explicit and preference-related feedback than conventional implicit user actions such as searching, clicking, and purchasing.
- A recommender dialog system must generate informative responses and provide high-quality recommendations from users’ utterances.
- KBRD integrates the recommender and dialog systems end to end, incorporating external knowledge to strengthen their connection and performance.
- The dialog system supplies item context to recommendation, while the recommender returns recommendation-aware vocabulary bias to promote dialog generation.
- KBRD improves both recommendation and dialog generation, with analyses showing dialog information helps recommendation and vocabulary bias improves generated-dialog quality.
2 Preliminary
The preliminary framework combines a recommender that retrieves items from user information with a dialog generator for multi-turn natural-language interaction. End-to-end training combines recommendation and vocabulary distributions, while KBRD further connects both systems through knowledge propagation.
- In cold-start recommendation, the system initially lacks user knowledge and accumulates information from conversation to build a preference-based user profile.
- The dialog system generates multi-turn responses with a natural-language generation model that encodes dialog history and decodes responses recursively.
- The recommender uses mentioned items from dialog history, while its item probability distribution is combined with the dialog system’s vocabulary distribution.
- A switching mechanism controls whether the decoder generates a vocabulary word or an item from the recommender at each timestep.
- The baseline framework has weak recommender–dialog interaction, whereas KBRD links informative entities to a knowledge graph and sends knowledge-enhanced user representations back as vocabulary bias.
- The combined system can be trained end to end using switcher parameters and a sigmoid function to control decoder selection.
3 Proposed Model
KBRD integrates recommendation and dialog generation through knowledge propagation, allowing dialog context to inform recommendations and recommendation information to guide responses. Its framework links mentioned items and non-item entities through a knowledge graph and adds recommendation-aware vocabulary bias to the dialog system.
- 3 Proposed Model: KBRD integrates the recommender and dialog systems through knowledge propagation so they can provide mutual benefits.The framework is designed as an end-to-end bridge between both systems.
- 3.1 Dialog-Aware Recommendation with Knowledge: Dialog contents include mentioned items and linked non-item entities, enabling recommendations from contextual preferences beyond explicitly mentioned items.Items and informative entities are matched or linked to entities in the external knowledge graph.
- 3.1 Dialog-Aware Recommendation with Knowledge: Relational graph convolutional networks encode structural and relation-specific information from neighboring knowledge-graph nodes into entity representations.Each node aggregates transformed messages from relation-specific neighbors and combines them with its current representation.
- 3.1 Dialog-Aware Recommendation with Knowledge: Self-attention combines knowledge-enhanced representations of interacted entities into a fixed-size user representation while weighting entities by importance.The recommendation output masks non-item entities so that only items are recommended.
- 3.2 Recommendation-Aware Dialog: The dialog system uses a Transformer encoder-decoder and incorporates the recommender’s user representation as a vocabulary bias during decoding.The bias is computed by a feed-forward network from the user representation and added to the decoder’s output layer.
4 Experiments
The experiments evaluate KBRD on recommendation and dialog generation using REDIAL, automatic and human dialog metrics, and Recall@K. KBRD achieves the best reported performance across recommendation metrics and dialog-generation evaluations.
- Experimental Setup: The experiments use REDIAL, containing 10,006 conversations and 182,150 utterances involving 956 users and 51,699 movies.
- Evaluation: Recommendation is evaluated with Recall@1, Recall@10, and Recall@50, while dialog generation uses perplexity, distinct 3-gram, distinct 4-gram, and human consistency scores.
- Recommendation Results: KBRD reaches the best performance on Recall@1, Recall@10, and Recall@50, and both dialog contents and external knowledge improve recommendation performance.
- Recommendation Results: +0.7 Recall@1, +3.4 Recall@10, and +5.1 Recall@50 are obtained when dialog and knowledge are combined, exceeding the separate contributions.
- Dialog Results: KBRD improves diversity over the Transformer baseline while maintaining the best perplexity, indicating more diverse generated content without decreased fluency.
- Dialog Results: KBRD outperforms REDIAL by +0.22 consistency score, a 15% advantage, in human evaluation of dialog-history consistency.
5 Discussion
The discussion analyzes how dialog and recommendation mutually improve one another. Dialog helps recommendation especially with few mentioned items, while recommendation-aware information improves dialog consistency, diversity, and explainability.
- 5.1 Does dialog help recommendation?: Dialogs containing only 0-2 mentioned movies comprise 62.8% of the test set, making high-quality recommendation with sparse dialog information important.
- 5.1 Does dialog help recommendation?: With no mentioned item, dialog incorporation substantially outperforms the baseline and knowledge-only system by using contextual preferences such as directors and actors.
- 5.1 Does dialog help recommendation?: As more items are mentioned, knowledge becomes more significant than dialog, while combining both information sources performs best on average.
- 5.2 Does recommendation help dialog?: Recommendation-aware vocabulary bias gives the dialog system user- and item-relevant words, improving response consistency and diversity.
- 5.2 Does recommendation help dialog?: Top biased words can relate to movie attributes and creators, providing explicit clues that help explain recommendation results.
6 Related Work
Related work covers recommender systems, dialog systems, and their intersection in conversational recommendation. Prior recommender research addresses interaction sparsity and cold start, while conversational systems connect dialog with other intelligent tasks.
- Recommender Systems: Traditional recommender systems rely on collaborative filtering, while recent approaches increasingly use neural networks.
- Recommender Systems: Researchers have proposed methods to address cold start and sparse user-item interactions in recommendation.
- Recommender Systems: Explainability is identified as an important recommender-system evaluation aspect alongside accuracy.
- Dialog Systems: End-to-end neural dialog systems have been studied in open-ended, goal-oriented, visual-dialog, and conversational-recommendation settings.
7 Conclusion
The conclusion presents KBRD as an end-to-end framework linking recommendation and dialog through knowledge propagation. Experiments report better performance in both tasks and mutual benefits between the systems.
- Conclusion: KBRD bridges the recommender and dialog systems through knowledge propagation in an end-to-end framework.
- Conclusion: Experiments show better KBRD performance than baselines in both recommendation and dialog generation.
- Conclusion: Dialog information is especially effective for recommendation in cold-start settings, while recommendation information enhances dialog consistency and diversity.