Pearl: A Review-driven Persona-Knowledge Grounded Conversational Recommendation Dataset

Conversational recommender system is an emerging area that has garnered anincreasing interest in the community, especially with the advancements in largelanguage models (LLMs) that enable diverse reasoning over conversational input.Despite the progress, the field has many aspects left to explore. The currentlyavailable public datasets for conversational recommendation lack specific userpreferences and explanations for recommendations, hindering high-qualityrecommendations. To address such challenges, we present a novel conversationalrecommendation dataset named PEARL, synthesized with persona- andknowledge-augmented LLM simulators. We obtain detailed persona and knowledgefrom real-world reviews and construct a large-scale dataset with over 57kdialogues. Our experimental results demonstrate that utterances in PEARLinclude more specific user preferences, show expertise in the target domain,and provide recommendations more relevant to the dialogue context than those inprior datasets.