Leveraging Semantic Embeddings of User Reviews with Off-the-Shelf LLMs for Recommender Systems

Enhancing Recommender Systems (RS) with plain-text reviews has been a challenging effort despite significant efforts in the past. Recently, Large Language Models (LLMs) have demonstrated exceptional capabilities in understanding natural language semantics, leading to promising applications across various fields. Nonetheless, applying these models to recommendation tasks introduces several challenges, including high computational demands and the potential for generating inaccurate or fabricated content (”hallucinations”). Consequently, instead of directly employing LLMs as generative models for recommendations, our research explores whether embeddings derived from plain-text reviews can enrich traditional recommendation algorithms and analyze the recommendation impact of different LLM embeddings with high effectiveness in NLP tasks. We conduct our experimental analysis using two Amazon Review Datasets, and three pre-trained LLM embedding models.