핵심 개념
Conversational recommender systems benefit from PEARL's persona and knowledge augmentation, enhancing user preferences and recommendations.
초록
Abstract:
Conversational recommender systems are gaining interest.
Existing datasets lack specific user preferences and explanations.
PEARL dataset addresses these limitations with persona and knowledge augmentation.
Introduction:
CRS aims for personalized recommendations through interactive conversations.
Emphasis on high-quality dataset construction.
Existing datasets collected via crowdsourcing have limitations.
PEARL Construction:
User-Review and Item-Review Databases constructed.
Persona-augmented User Simulator enhances preference specificity.
Knowledge-augmented Recommender Simulator provides detailed explanations.
Experiments:
Human evaluation shows PEARL preferred over crowdsourced datasets.
PEARL outperforms in recommendation and response generation tasks.
PEARL is cost and time-efficient compared to traditional datasets.
Case Study:
BART-PEARL consistently provides detailed explanations in responses.
BART-PEARL outperforms BART-ReDial in human evaluation.
Limitations and Ethical Considerations:
Impact of language model choice on dialogue creation.
Considerations for dialogue safety and biases.
References:
Various studies on conversational recommendation systems.
통계
PEARL contains over 57k dialogues.
PEARL covers diverse user preferences and detailed item explanations.
PEARL dataset is preferred by human raters compared to crowdsourced datasets.
인용구
"Despite being fully machine-generated, human raters judge PEARL as better in quality compared to both ReDial and INSPIRED."
"BART-PEARL consistently generates responses with explanations that elucidate what the recommended item is and why the recommender suggests the item."