Find in Library

With the popularization of social media, Word-of-mouth analysis, as an important market research method, can help companies understand users’ attitudes and opinions towards their brands. However, existing social media word-of-mouth analysis methods face challenges such as insufficient feature fusion, low accuracy of sentiment analysis, low precision of topic identification, as well as data sparsity and high annotation costs, which hinder the comprehensive and accurate analysis of word-of-mouth. To address these challenges, this paper proposes a social media word-of-mouth analysis model combining the Adaptive Vision Transformer (AViT) model and Enhanced BERT (EBERT) model. Firstly, we improve the traditional Transformer model by introducing dynamic attention mechanism and multi-scale processing to enhance its adaptability to different types of images, effectively adjusting computational resources, and enhancing attention to various parts of the images. Secondly, enhancements are made to the BERT model, including learnable positional encodings, sparse Transformer mechanism, and enhanced Masked Language Model, enabling better capture of semantic information and contextual relationships. Furthermore, by introducing the idea of multi-task learning, sentiment classification, topic recognition, and other tasks are combined to achieve comprehensive analysis of word-of-mouth. Through experiments on large-scale social media datasets, the proposed model not only achieves high accuracy in sentiment analysis and topic recognition tasks but also demonstrates good generalization ability, providing an effective social media marketing strategy and decision support tool for enterprises and marketing practitioners.