基于话语间时序多模态数据的情绪分析方法

Sentiment Analysis Method Based on Temporal Multimodal Data Between Utterances

长期以来,传统的基于单模态数据情绪分析方法存在分析角度单一、分类准确率低下等问题,时序多模态数据的分析方法为解决这些问题提供了可能.本文基于话语间的时序多模态数据,对现有的多模态情绪分析方法进行了改进,使用双向门控循环网络(Bi-GRU)结合模态内和跨模态的上下文注意力机制进行情绪分析,最后在MOSI和MOSEI数据集上进行验证.实验表明,利用话语间的时序多模态数据,并且充分融合模态内以及跨模态上下文信息的方法,能够从多模态特征和时序特征的角度进行情绪分析,从而有效提高情绪分析任务的分类准确率.

The traditional sentiment analysis methods based on single-modal data have always had problems such as a single analysis angle and low classification accuracy.The analysis method based on temporal multimodal data provides the possibility to solve these problems.On the basis of the temporal multimodal data between utterances,this study improves the existing multimodal sentiment analysis method and uses the bidirectional gated recurrent unit(Bi-GRU)combined with the intra-modal and cross-modal context attention mechanism for sentiment analysis.The sentiment analysis is finally verified on the MOSI and MOSEI datasets.Experiments show that the method of using temporal multimodal data between utterances and fully integrating intra-modal and cross-modal context information can be applied to sentiment analysis from the perspective of multimodal and temporal features.By doing this,the classification accuracy of sentiment analysis can be effectively improved.