Detecting fake review intentions in the review context: A multimodal deep learning approach

Abstract

The proliferation of fake reviews on the internet has had significant repercussions for both consumers and businesses. However, existing research predominantly employs a binary classification approach to ascertain review authenticity, often neglecting the rich multimodal context information and nuanced intentions embedded within them. To bridge this gap, our study introduces a novel task, Fake Review Intention Detection in Review Context (FRIDRC), which aims to detect fake review intentions by leveraging both textual and visual information, and constructs a dataset comprising both manually and AI-generated fake reviews. Additionally, we develop a predictive framework encompassing modules for multimodal representation and modality fusion. These modules, while independent, are synergistic and effectively tackle the challenge of discerning fake review intentions. Our framework demonstrates outstanding performance, achieving an average F1 score exceeding 0.97 and a Macro F1 score surpassing 0.96 in this task and outperforming advanced pre-trained models. This research not only presents an effective methodology for accurately identifying and addressing fake review intentions but also underscores the efficacy of leveraging multimodal review context information in fake review detection. The dataset and code implementation are publicly available for further research.