Abstract
The rapid development of social media has driven the need for opinion mining and sentiment analysis based on multimodal samples. As a fine-grained task within multimodal sentiment analysis, aspect-based multimodal sentiment analysis (ABMSA) enables the accurate and efficient determination of sentiment polarity for aspect-level targets. However, traditional ABMSA methods often perform suboptimally on social media samples, as the images in these samples typically contain embedded text that conventional models overlook. Such text influences sentiment judgment. To address this issue, we propose a text-in-image enhanced self-supervised alignment model (TESAM) that accounts for multimodal information more comprehensively. Specifically, we employed Optical Character Recognition technology to extract embedded text from images and, based on the principle that text-in-image is an integral part of the visual modality, fused it with visual features to obtain more comprehensive image representations. Additionally, we incorporate aspect words to guide the model in disregarding irrelevant semantic features, thereby reducing noise interference. Furthermore, to mitigate the semantic gap between modalities, we propose pre-training the feature extraction module with self-supervised alignment. During this pre-training stage, unimodal semantic embeddings from both modalities are aligned by calculating errors using Euclidean distance and cosine similarity. Experimental results demonstrate that TESAM achieved remarkable performances on three ABMSA benchmarks. These results validate the rationale and effectiveness of our proposed improvements.