Leveraging Historical Interactions for Factual Review Generation with Large Language Models

Mingzhe Liu, Yi Qiao, Leilei Sun, Tongyu Zhu, Yi Xu, Hao Sheng

Keywords: Review Generation, Large Language Models, Historical Interaction Representation Learning

Abstract: User reviews play a crucial role in influencing purchase decisions and enhancing recommendation systems. Automatically generating high-quality reviews helps complement existing feedback by providing additional perspectives and uncovering overlooked details for consumers. LLMs have become ideal tools for this task due to their strong text generation capabilities. However, existing LLM-based methods often fail to effectively incorporate user- and item-specific interactions, limiting their ability to generate factually consistent and contextually relevant reviews. To address these challenges, we propose DyGRevLLM, an innovative framework that integrates dynamic graph representation learning with LLM-based text generation. DyGRevLLM encodes and updates user and item embeddings through a pretraining procedure designed to predict future review embeddings, aligning historical interaction data with LLM input formats. By dynamically aggregating user-item interaction information and incorporating temporal behaviors, the framework generates reviews that are factually accurate. Extensive experiments on real-world datasets demonstrate that DyGRevLLM improves the factual consistency and relevance of generated reviews while maintaining coherence. Furthermore, our proposed evaluation metrics validate the effectiveness of the framework in overcoming existing limitations, offering a better solution for dynamic personalized review generation.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 9098