Scholarlink

<p>E-commerce platforms face the challenge of the “cold-start” problem, where new products lack user reviews crucial for buyer decisions. This paper introduces a novel Large Language Model (LLM)-driven approach for generating high-quality augmented reviews for cold-start products by utilizing reviews from semantically similar “warm” products. The proposed method follows a two-stage framework: Initially, an advanced semantic retrieval mechanism is deployed, enhanced through a data-driven attribute weighting strategy. This attribute weighting is derived via a zero-shot aspect-based segmentation technique, guided by a strategically engineered prompting approach to optimize attribute weights. This initial stage ensures the precise extraction of semantically relevant warm product reviews. Subsequently, in the second stage, these retrieved segments are refined, filtering out irrelevant content to maximize alignment with the cold product’s characteristics. Evaluated on two comprehensive datasets from CNET.com (Digital Cameras and MP3 Players), the method demonstrates a substantial average improvement of 65% in ROUGE score metrics compared to traditional baseline approaches (e.g., Query Likelihood) across varying review lengths. Furthermore, it achieves superior performance compared to the state-of-the-art embedding-based method, attaining a ROUGE-2 F1-score of 0.093 on Digital Cameras and 0.113 on MP3 Players (representing improvements of approximately 3%). By enhancing the informativeness and reliability of generated reviews, this work presents a significant step forward in mitigating the cold-start problem, enhancing the online e-commerce experience for both buyers and sellers.</p> Graphical Abstract <p></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Zero to Feedback: Leveraging Large Language Models for Augmented Review Generation in Cold-Start Scenarios

  • 期刊论文

  • 发表日期：

    2026年4月24日

查看全文
  • Meysam Roostaee

摘要

E-commerce platforms face the challenge of the “cold-start” problem, where new products lack user reviews crucial for buyer decisions. This paper introduces a novel Large Language Model (LLM)-driven approach for generating high-quality augmented reviews for cold-start products by utilizing reviews from semantically similar “warm” products. The proposed method follows a two-stage framework: Initially, an advanced semantic retrieval mechanism is deployed, enhanced through a data-driven attribute weighting strategy. This attribute weighting is derived via a zero-shot aspect-based segmentation technique, guided by a strategically engineered prompting approach to optimize attribute weights. This initial stage ensures the precise extraction of semantically relevant warm product reviews. Subsequently, in the second stage, these retrieved segments are refined, filtering out irrelevant content to maximize alignment with the cold product’s characteristics. Evaluated on two comprehensive datasets from CNET.com (Digital Cameras and MP3 Players), the method demonstrates a substantial average improvement of 65% in ROUGE score metrics compared to traditional baseline approaches (e.g., Query Likelihood) across varying review lengths. Furthermore, it achieves superior performance compared to the state-of-the-art embedding-based method, attaining a ROUGE-2 F1-score of 0.093 on Digital Cameras and 0.113 on MP3 Players (representing improvements of approximately 3%). By enhancing the informativeness and reliability of generated reviews, this work presents a significant step forward in mitigating the cold-start problem, enhancing the online e-commerce experience for both buyers and sellers.

Graphical Abstract