Grunnleggende konsepter
The author presents PEARL, a dataset addressing limitations in existing conversational recommendation datasets by synthesizing persona- and knowledge-augmented dialogues.
Sammendrag
PEARL introduces a novel conversational recommendation dataset synthesized with persona- and knowledge-augmented large language model simulators. The dataset addresses limitations in existing datasets by providing more specific user preferences, expertise in the target domain, and relevant recommendations. Experimental results show that models trained on PEARL outperform those trained on human-annotated datasets.
Key points include:
- PEARL is a large-scale dataset with over 57k dialogues simulating real-world user preferences.
- The dataset includes detailed persona and item knowledge extracted from real-world reviews.
- Simulators are designed to enhance preference specificity and informativeness of collected data.
- Human evaluation shows PEARL is preferred over other crowdsourced datasets.
- Models trained on PEARL demonstrate competitive or better performances in recommendation tasks.
- The dataset is cost-efficient and time-effective compared to traditional dialogue crowdsourcing methods.
Statistikk
Number of dialogues: 57,277
Number of users: 4,680
Number of utterances: 548,061