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

Source: arXiv:2203.13366, March 2022 Authors: Shijie Geng, Shuchang Liu, Zuohui Fu, Yingqiang Ge, Yongfeng Zhang (Rutgers University)

Overview

This paper proposes Recommendation as Language Processing (RLP), recasting recommendation problems as natural language processing tasks. The authors introduce P5 (Pretrain, Personalized Prompt, and Predict), a unified framework where a single pre-trained language model serves as the foundation for a broad family of recommendation tasks.

Key Idea

Traditional recommender systems build separate, task-specific models for rating prediction, sequential recommendation, explanation generation, and so on. P5 instead:

• Pretrains a transformer-based language model as a shared recommendation foundation;
• Personalizes behavior through natural language prompts that encode user context, item information, and task instructions;
• Predicts outputs directly as text—ratings, item lists, explanations, or summaries—using a unified sequence-to-sequence formulation.

By converting user IDs, item metadata, and interaction histories into textual form, all recommendation objectives become conditional text generation problems that one model can handle.

Contributions

• A unified prompt-based paradigm covering multiple recommendation tasks (rating prediction, sequential recommendation, explanation generation, review summarization, direct recommendation) within a single model.
• A family of personalized prompt templates enabling zero-shot and few-shot personalization without task-specific architectures.
• Experiments on public recommendation benchmarks demonstrating competitive accuracy alongside human-readable outputs such as explanations.

Significance

P5 is widely cited as a foundational work for LLM-based generative recommendation (Gen-Rec), helping establish the research direction of bridging recommender systems and large language models. It anticipates later instruction-tuned and prompt-driven recommenders, and connects naturally to related work on language-item retrieval (e.g., BLAIR) and decoder-only foundation models for personalization.

Notes for Practitioners

• Evaluation should consider not only accuracy metrics but also the quality of generated explanations and natural language outputs.
• Offline gains should be validated against online user satisfaction; latency and cost of LLM-based serving remain hard constraints in production.
• Follow-up literature extends P5 toward data-efficient fine-tuning, knowledge-graph integration, and agentic recommendation pipelines.