| Summary: | In recent years, multi-modal sentiment analysis has become more and more popular in the field of natural language processing. Multi-modal sentiment analysis mainly concentrates on text, image and audio information. Previous work based on BERT utilizes only text representation to fine-tune BERT, while ignoring the importance of nonverbal information. Most current research methods are fine-tuning models based on BERT that do not optimize BERT’s internal structure. Therefore, in this paper, we propose an optimized BERT model that is composed of three modules: the Hierarchical Multi-head Self Attention module realizes the hierarchical extraction process of the features; the Gate Channel module replaces BERT’s original Feed-Forward layer to realize information filtering; the tensor fusion model based on self-attention mechanism utilized to implement the fusion process of different modal features. In CMU-MOSI, a public mult-imodal sentiment analysis dataset, the accuracy and F1-Score were improved by 0.44% and 0.46% compared with the original BERT model using custom fusion. Compared with traditional models, such as LSTM and Transformer, they are improved to a certain extent.
|
|---|