LALDM: A Multimodal Aspect-Level Text Analysis Method and Its Application in Online Consumer Electronics

Aspect term extraction and aspect level sentiment analysis are key tasks. Although in the multimodal field, performance is enhanced by placing these two tasks in a unified framework, there is still room for improvement in aspect level analysis for short texts. Firstly, existing research has shown that texts usually play a more important role in online reviews than images. Therefore, we use a large language model to automatically label the text, thereby enhancing its contribution of text to aspect level analysis. Secondly, we use a better depth model than most existing studies, DenseNet, to enhance the effectiveness of image analysis. We integrated text analysis and image analysis modules to form a unified framework for aspect term extraction and aspect sentiment analysis to maintain the continuity of the underlying features of these two tasks. The proposed method called Large language model Automatically Labeled and Dansenet for Multimodal (LALDM). The experimental results show that the proposed method improves the performance of existing methods in MABSA tasks. In addition, LALDM has been applied to a cross modal semantic understanding task for online consumer electronics, and experimental results show that it has better performance than the control algorithms.