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

This forum entry discusses the paper "Recommendation as Language Processing (RLP): A Unified Pretrain, Personalized Prompt & Predict Paradigm (P5)" (March 2022), available on arXiv at https://arxiv.org/abs/2203.13366.

Paper Overview

• Authors: Shijie Geng, Shuchang Liu, Zuohui Fu, Yingqiang Ge, Yongfeng Zhang
• Venue: arXiv preprint, March 2022
• Link: arXiv:2203.13366

Motivation

Traditional recommender systems build separate task-specific models for rating prediction, ranking, explanation, and summarization. This fragmentation increases engineering cost and limits knowledge sharing across tasks. P5 proposes Recommendation as Language Processing (RLP): treating recommendation as a natural language problem so that one unified model can serve many recommendation tasks through prompts.

Core Ideas

1. Unified paradigm (Pretrain, Prompt, Predict): A single foundation model is pretrained on user–item interaction data drawn from multiple recommendation scenarios, then applied to different tasks via personalized natural-language prompts. 2. Personalized prompts: User information is encoded into the input prompt, so recommendation targets can be described in language rather than as numeric IDs alone. 3. Multi-task coverage: P5 handles rating prediction, sequential recommendation, explanation generation, review summarization, and direct recommendation within one model. 4. Zero-shot generalization: Because tasks are formulated as language, P5 can generalize to unseen recommendation scenarios without task-specific fine-tuning.

Context Within the Field

The post situates P5 in the evolution of recommender systems:

• Matrix factorization → deep CTR models → sequential Transformers → LLM instruction-following and generative recommendation (Gen-Rec).
• LLMs bring semantic priors and cold-start ability, but production deployment must address inference latency, cost, hallucination risk, and evaluation reliability (offline metrics vs. online satisfaction).

Practical Takeaways

• Architecture: Prompt-driven unified models reduce the number of task-specific pipelines to maintain.
• Data: High-quality instruction data and interaction logs remain critical; beware leakage and distribution shift in synthetic data.
• Evaluation: Cross-validate LLM-based or offline metrics with human evaluation and online A/B tests.
• Deployment: Latency, cost, explainability, and safety are hard constraints in industrial systems, not just benchmark numbers.

Related Entries

• 360Brew: A Decoder-only Foundation Model for Personalized Ranking
• Bridging Language and Items for Retrieval and Recommendation (arXiv:2403.03952)
• Data-efficient Fine-tuning for LLM-based Recommendation (SIGIR 2024)

> Note: For exact quantitative results, consult the original paper PDF; this entry summarizes based on the abstract and public metadata.