Research on Multimodal Aspect-Based Sentiment Analysis Based on Image Caption and Multimodal Aspect Extraction

Sentiment Analysis is an important task in NLP. In recent years, more and more attention has been paid to Multimodal Sentiment Analysis. For Multimodal Aspect-Based Sentiment Analysis (MABSA), most of the existing methods take the original information of each modality as input and extract each modality’s features and the alignment between them. Based on the existing methods, additional information, including aspect terms and image caption, is extracted. The original and additional information is used as a compound feature so that the model can have a better understanding of each modality and the relationship between them. Experiments show that the method achieves better results on MABSA and its two sub-tasks: Multimodal Aspect-Sentiment Classification (MASC) and Multimodal Aspect-Terms Extraction (MATE).