Fine-Grained Sentiment Analysis through Aesthetic Caption Fusion and Semantic Filtering
With the rapid development of social networks,people are increasingly expressing their emotions through tweets on these platforms. This has led to the emergence of Multimodal Aspect-Based Sentiment Classification (MASC), which aims to classify the polarity of the sentiment for specific aspects. Although numerous approaches have been proposed and significant progress has been achieved, there remain gaps in the extraction of deep emotional information and reasoning from an aesthetic perspective. In this work, we propose AIFFM: Auxiliary Information Fusion and Filtering Model, a novel MASC algorithm, which integrates features from three dimensions of the image associated with the tweet: image description, the observer's emotional impression, and aesthetic evaluation. Additionally, we leverage large language models (LLMs) to achieve a lossless emotional integration and a filter module is then used to further extract practical sentiment-aware information. Experimental results on two standard MASC dataset demonstrate that our method achieves comparable performance on the MASC task.
