Abstract:

Multimodal product reviews， which include both text and images， have become the mainstream way for customers to express opinions and share word-of-mouth on e-commerce platforms. Previous studies on multimodal review usefulness primarily focus on feature representation and fusion， often neglecting the reliability of multimodal data. Considering the heterogeneity of these two types of multimodal data and their varying degrees of reliability in influencing the value of online product reviews， an online product review helpfulness recognition method is propased that incorporates multimodal credibility. The method fully accounts for the heterogeneity of textual and visual modalities and captures consistency information between modalities through their interaction. A credible multi-view fusion module is designed to estimate the uncertainty of both unimodal and cross-modal views， improving the overall reliability of the model through a dynamic evidence fusion strategy and a contrastive learning strategy. Empirical validation on multiple product review datasets from Amazon demonstrates that the proposed method effectively enhances the accuracy of online product review helpfulness recognition while increasing the interpretability of the model’s decision-making results.

Key words: online product reviews, helpfulness prediction, multimodal fusion, trusted multi-view learning