Multi-modal Sentiment Analysis of Mongolian Language based on Pre-trained Models and High-resolution Networks
Yang Yang, Qing-Dao-Er-Ji Ren, Rui-Feng He · 2024
Massive digital data in the information age promotes the development of multi-modal sentiment analysis research. However, there are relatively few related studies on Mongolian sentiment analysis due to the lack of large-scale data sets and multi-modal feature extraction technology. This paper proposes a Mongolian multi-modal sentiment analysis method based on pre-trained models and high-resolution networks. First, for sentiment analysis of Mongolian texts, this paper uses pre-training models and word segmentation technology to build a Mongolian vocabulary, which solves the problem of unregistered words, and extracts text features through graph convolutional neural networks to achieve sentiment classification. Secondly, for Mongolian audio sentiment analysis, a method based on bidirectional gating units is proposed to effectively capture the temporal relationships in audio data. Third, for Mongolian visual emotion analysis, the high-resolution representation network is used to extract detailed information in the video, which effectively improves the classification performance. Finally, a multi-modal emotional feature fusion method is proposed through the multi-head attention mechanism to fuse text, audio and visual features, which achieves complementary fusion of different modal information and further improves sentiment analysis performance. The experimental results show that the Mongolian multi-modal sentiment analysis framework proposed in this paper shows good performance in the experiment, providing new ideas and methods for the research and application in the field of Mongolian sentiment analysis.