Dual Causes Generation Assisted Model for Multimodal Aspect-Based Sentiment Classification.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Computer Vision

Background:

• Multimodal aspect-based sentiment classification (MABSC) analyzes sentiment in user-generated content.
• Current MABSC methods struggle with limited context in social media, impacting accuracy.
• Identifying the reasons behind expressed sentiments is crucial for better analysis.

Purpose of the Study:

• To propose a novel multimodal dual cause analysis (MDCA) method for MABSC.
• To enhance sentiment prediction accuracy by incorporating reasoning cause (RC) and direct cause (DC).
• To develop and evaluate a multitask learning framework for MDCA.

Main Methods:

• Constructed MABSC datasets incorporating RC and DC using large language models (LLMs) and visual-language models.
• Devised a multitask learning framework to train a small generative model.
• The model generates RC and DC to assist in sentiment prediction.

Main Results:

• The MDCA model achieved state-of-the-art performance on MABSC benchmark datasets.
• The fine-tuned small model demonstrated superior adaptability compared to large models like ChatGPT and BLIP-2.
• Incorporating cause data significantly improved sentiment prediction accuracy.

Conclusions:

• MDCA offers a promising approach to enhance MABSC by providing causal explanations.
• Small, fine-tuned models can be highly effective and adaptable for MABSC tasks.
• This research contributes to more accurate and interpretable sentiment analysis in multimodal contexts.