Abstract
For a long time, different recommendation tasks typically require designingtask-specific architectures and training objectives. As a result, it is hard totransfer the learned knowledge and representations from one task to another,thus restricting the generalization ability of existing recommendationapproaches, e.g., a sequential recommendation model can hardly be applied ortransferred to a review generation method. To deal with such issues,considering that language can describe almost anything and language groundingis a powerful medium to represent various problems or tasks, we present aflexible and unified text-to-text paradigm called "Pretrain, PersonalizedPrompt, and Predict Paradigm" (P5) for recommendation, which unifies variousrecommendation tasks in a shared framework. In P5, all data such as user-iteminteractions, user descriptions, item metadata, and user reviews are convertedto a common format -- natural language sequences. The rich information fromnatural language assists P5 to capture deeper semantics for personalization andrecommendation. Specifically, P5 learns different tasks with the same languagemodeling objective during pretraining. Thus, it serves as the foundation modelfor various downstream recommendation tasks, allows easy integration with othermodalities, and enables instruction-based recommendation based on prompts. P5advances recommender systems from shallow model to deep model to big model, andwill revolutionize the technical form of recommender systems towards universalrecommendation engine. With adaptive personalized prompt for different users,P5 is able to make predictions in a zero-shot or few-shot manner and largelyreduces the necessity for extensive fine-tuning. On several recommendationbenchmarks, we conduct experiments to show the effectiveness of P5. We releasethe source code at \url{https://github.com/jeykigung/P5}.