Source-linked AI summary
Towards Unified Conversational Recommender Systems via Knowledge-Enhanced Prompt Learning
Xiaolei Wang, Kun Zhou, Ji-Rong Wen, Wayne Xin Zhao
TL;DR
Existing CRS methods struggle to integrate recommendation and conversation because their modules use different architectures or techniques and cannot fully share outputs. UniCRS addresses this gap with knowledge-enhanced prompt learning over a fixed PLM, and experiments on two public datasets show effectiveness on both subtasks, especially with limited training data.
Problem
Existing CRS approaches use different module architectures or techniques, making seamless integration of recommendation and conversation difficult.
Method
UniCRS reformulates both subtasks as knowledge-enhanced prompt learning over a fixed DialoGPT backbone, using fused knowledge, task-specific soft tokens, dialogue context, and generated response templates.
Results
UniCRS outperforms several competitive methods on recommendation and conversation subtasks, especially when training data is limited.
Takeaways & Limitations
A unified prompt-learning formulation can support both CRS subtasks while enabling response templates to inform recommendation prompts.
Takeaways & Limitations
The authors identify more complicated scenarios, including topic-guided and multi-modal CRS, as future application settings.
Abstract
from arXiv · showhide
Conversational recommender systems (CRS) aim to proactively elicit user preference and recommend high-quality items through natural language conversations. Typically, a CRS consists of a recommendation module to predict preferred items for users and a conversation module to generate appropriate responses. To develop an effective CRS, it is essential to seamlessly integrate the two modules. Existing works either design semantic alignment strategies, or share knowledge resources and representations between the two modules. However, these approaches still rely on different architectures or techniques to develop the two modules, making it difficult for effective module integration. To address this problem, we propose a unified CRS model named UniCRS based on knowledge-enhanced prompt learning. Our approach unifies the recommendation and conversation subtasks into the prompt learning paradigm, and utilizes knowledge-enhanced prompts based on a fixed pre-trained language model (PLM) to fulfill both subtasks in a unified approach. In the prompt design, we include fused knowledge representations, task-specific soft tokens, and the dialogue context, which can provide sufficient contextual information to adapt the PLM for the CRS task. Besides, for the recommendation subtask, we also incorporate the generated response template as an important part of the prompt, to enhance the information interaction between the two subtasks. Extensive experiments on two public CRS datasets have demonstrated the effectiveness of our approach.
1 INTRODUCTION
CRSs must coordinate recommendation and conversation, but existing approaches often use separate architectures or techniques that can produce semantically inconsistent outputs. UniCRS unifies both subtasks through knowledge-enhanced prompt learning and shows effectiveness on both tasks across two public datasets.
- CRSs combine item recommendation with natural-language question or response generation.
- Existing methods use separate recommendation and conversation modules, despite the subtasks being highly coupled.
- KGSF can predict “Frozen 2 (2019)” while generating a mismatched response mentioning “Pretty Woman (1990)”.
- The inconsistency reflects different module architectures and the inability of one module to use the other’s outputs.
- UniCRS formulates both subtasks as prompt learning with task-specific prompts, fused knowledge representations, soft tokens, and dialogue context.
- Experiments on two public CRS datasets show effectiveness in both recommendation and conversation tasks.
2 RELATED WORK
Related CRS research pursues conversational recommendation through predefined actions or jointly generated recommendations and responses, while prompt learning reformulates downstream tasks for pretrained language models.
- Conversational recommendation methods using predefined actions optimize interactions through intent slots, item attributes, bandits, or reinforcement learning.
- CRS systems that generate recommendations and human-like responses typically assign these functions to separate modules.
- Existing integration strategies share knowledge resources or use semantic-alignment pretraining tasks and regularization terms.
- Prompt learning reformulates downstream tasks as the pretraining task through carefully designed prompts.
3 PROBLEM STATEMENT
A CRS processes multi-turn dialogue history to select candidate items and generate a response, which may include those recommendations or contain no item when recommendation is unnecessary.
- A CRS conducts item recommendation through multi-turn natural-language conversations and may ask clarification questions.
- The conversation consists of utterances formed from sequences of vocabulary words.
- At turn t, the system uses the dialogue history and item set I to select candidate items I_t.
- The system generates a response R that includes the selected items, although I_t may be empty when recommendation is unnecessary.
4 APPROACH
UniCRS unifies recommendation and conversation through knowledge-enhanced prompts built around a fixed PLM. It fuses word and entity semantics, adapts prompts to each subtask, and uses generated response templates to connect recommendation with response generation.
- 4.1 Overview of the Approach: UniCRS uses DialoGPT as a fixed base PLM and solves recommendation and conversation through prompt-based generation.Prompt tokens adapt the frozen PLM to different CRS subtasks without fine-tuning.
- 4.2 Semantic Fusion for Prompt Learning: The framework encodes dialogue words with RoBERTa and entities linked through DBpedia and represented with an RGCN.Word and entity embedding matrices are constructed separately before semantic fusion.
- 4.2 Semantic Fusion for Prompt Learning: A bilinear cross-interaction mechanism fuses word and entity representations to bridge their semantic gap.The affinity matrix and transformation matrix connect the two representation spaces.
- 4.2 Semantic Fusion for Prompt Learning: The fusion module is pre-trained by prepending fused entity representations to dialogue context and predicting entities appearing in the response.The prompt-augmented sequence is formed as [fused entities; dialogue context; response].
- 4.3 Prompt Design: Each subtask prompt combines dialogue history, soft tokens, and fused knowledge representations, while recommendation additionally receives the generated response template.Recommendation uses entity representations, generation uses word representations, and template slots are later filled with recommended items.
5 EXPERIMENT
Experiments on ReDial and INSPIRED evaluate UniCRS against CRS methods and adapted PLMs for recommendation and conversation. UniCRS consistently outperforms the baselines, with component ablations and limited-data tests supporting the usefulness of its design.
- Experimental Setup: Experiments use the ReDial and INSPIRED English movie-recommendation CRS datasets, evaluating recommendation and conversation separately.Recommendation uses Recall@k, while conversation uses Distinct-n and human evaluation.
- Recommendation Evaluation: UniCRS outperforms all baselines by a large margin on the recommendation task.The authors attribute the result to knowledge-enhanced prompts, semantic-fusion pre-training, and response templates shared from conversation.
- Recommendation Ablation: Removing any recommendation component degrades performance, with the largest drop occurring when semantic-fusion pre-training is removed.The authors state that this pre-training learns correlations between entities and tokens and aligns entity semantics with the base PLM.
- Conversation Evaluation: UniCRS consistently performs better than baselines on automatic conversation metrics and human evaluation.The authors connect the conversation gains to semantic fusion, prompt pre-training, and injecting task-specific knowledge into DialoGPT.
- Conversation Ablation: Removing any conversation component lowers performance, while removing entity information causes the largest decrease.The authors associate the importance of entities with their domain-specific knowledge about recommendation items.
- Data Scarcity: With 20% of the training data, UniCRS remains comparable to the best baseline trained with full data.Across data-scarcity settings, baseline performance drops substantially, whereas UniCRS remains better than all baselines.
6 CONCLUSION
UniCRS unifies recommendation and conversation through knowledge-enhanced prompt learning with a fixed DialoGPT backbone, and reports stronger performance than competitive CRS and PLM methods, especially with limited training data. The authors identify more complex scenarios and prompt pre-training as future directions.
- UniCRS reformulates recommendation and conversation as unified subtasks using knowledge-enhanced prompt learning with a fixed DialoGPT backbone.The model optimizes prompts rather than fine-tuning the underlying PLM.
- Its prompts combine fused knowledge representations, task-specific soft tokens, and dialogue context to provide task and background information.
- Generated response templates are incorporated into recommendation prompts to enhance information interaction between the two subtasks.
- UniCRS outperforms several competitive CRS and PLM methods, especially when only limited training data is available.
- Future work will apply UniCRS to topic-guided and multi-modal CRS and develop more effective prompt pre-training for rapid adaptation.