Scholarlink

With the development of e-commerce, fake review identification is crucial for both platform users and merchants. Most datasets of fake reviews are imbalance, and the number of fake reviews is far less than that of real reviews. Due to the excellent performance of Large Language Modeling (LLM) in text generation, this paper proposes a fake review recognition method based on data augmentation by LLM. Firstly, several typical LLMs are selected to generate fake reviews, and the quality of the generated text is evaluated by the computational novelty and diversity. Subsequently, based on the data generated by different LLMs, and the performances of the different classification methods are compared. Finally, all the reviews generated by different LLMs are added to the original dataset, and the classification performances are compared with the original dataset. The experimental results of fake review identification show that the augmented dataset has about 10% improvement in classification accuracy compared to the original dataset only. It can be concluded that data augmentation through LLM is indeed significant for fake review identification.

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

Data Augmentation Using Large Language Model for Fake Review Identification

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Qingxu Li,
  • Jindong Chen,
  • Wen Zhang

摘要

With the development of e-commerce, fake review identification is crucial for both platform users and merchants. Most datasets of fake reviews are imbalance, and the number of fake reviews is far less than that of real reviews. Due to the excellent performance of Large Language Modeling (LLM) in text generation, this paper proposes a fake review recognition method based on data augmentation by LLM. Firstly, several typical LLMs are selected to generate fake reviews, and the quality of the generated text is evaluated by the computational novelty and diversity. Subsequently, based on the data generated by different LLMs, and the performances of the different classification methods are compared. Finally, all the reviews generated by different LLMs are added to the original dataset, and the classification performances are compared with the original dataset. The experimental results of fake review identification show that the augmented dataset has about 10% improvement in classification accuracy compared to the original dataset only. It can be concluded that data augmentation through LLM is indeed significant for fake review identification.