UsbVisdaNet: User Behavior Visual Distillation and Attention Network for Multimodal Sentiment Classification.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Computer Vision

Background:

• Biased user reviews negatively impact company evaluations and can spread misinformation.
• Identifying users with psychological biases is crucial for accurate sentiment analysis.
• Existing methods may not fully address the nuances of subjective bias in multimodal reviews.

Purpose of the Study:

• To propose a novel method for sentiment classification of multimodal data.
• To identify biased users by analyzing their psychological behaviors.
• To improve the accuracy of sentiment classification by mitigating subjective biases.

Main Methods:

• Developed UsbVisdaNet (User Behavior Visual Distillation and Attention Network) for multimodal sentiment classification.
• Integrated user behavior, text, and image features at multiple hierarchical levels.
• Leveraged psychological behavior analysis to detect positive and negative biased users.

Main Results:

• UsbVisdaNet demonstrated superior sentiment classification performance on the Yelp multimodal dataset.
• The method effectively identified biased users, improving overall sentiment analysis accuracy.
• Ablation and comparison experiments validated the effectiveness of the proposed approach.

Conclusions:

• UsbVisdaNet successfully integrates multimodal features for enhanced sentiment analysis.
• Analyzing user behavior is key to detecting and mitigating review bias.
• This research offers a significant advancement in understanding and classifying polarized opinions.