P5 - 리뷰

Recommendation as Language Processing (RLP): A Unified Pretrain, Personalized Prompt & Predict Paradigm (P5)

논문

• https://arxiv.org/abs/2203.13366

코드

• https://github.com/jeykigung/P5

참고자료

• (백본 모델인 T5 설명 자료) https://youtu.be/v7diENO2mEA

---

문제의식

• Different recommendation tasks require task-specific architectures and training objectives.
  • Hard to transfer the learned knowledge and representations (from one task to another)

아이디어

• Utilize (Task-agnostic) text-to-text paradigm

  • formulating multiple recommendation tasks as prompt-based natural language tasks
    • (model inputs) user-item information + prompt templates(* for the task)

  (참고) T5 in NLP : Unified framework - convert every language problem into a text-to-text format
  

프롬프트 구성

• With the prompts, directly build input–target pairs from raw data for training / testing
  • Simply substitute the fields in braces with the corresponding information in the raw data
• (Ground-Truth explanations) extract the sentences from reviews that comment on one or more item feature words
  • Sentires toolkit

모델구조

• text-to-text encoder-decoder architecture : pretrained T5 (backbone)