Research And Implementation Of Sentiment Analysis Mechanism Based On Multimodal Information Fusion

Human emotion is very abundant,and the study of sentiment analysis using text data can no longer meet the needs of sentiment analysis in the age of intelligence,and more modalities need to be introduced to supplement the sentiment information.In multimodal sentiment analysis,the acquisition of sentiment information of each modality has become an important research.At the same time,the introduction of multimodality will lead to an increase in the efficiency and resource consumption of sentiment analysis,so the optimization of the time consumption and memory usage of sentiment analysis is also an important research direction.The research is carried out to optimize the efficiency of sentiment analysis and the complete acquisition of information in different modalities of multimodal sentiment analysis,and the main research work is as follows.(1)For the optimization of efficiency of text sentiment analysis,a new sentiment dictionary EmoDic for mixed sentiment classification is proposed.This sentiment dictionary divides the words in the training set text into three major categories:sentiment words,general words and domain words.The sentiment weight of each word is calculated by counting the frequency of the word and the number of samples of that word in different sentiment labels.The sentiment weights of all words under different sentiments constitute the EmoDic.The dictionary can build a sentiment vector,which can be used either directly for sentiment classification or as a feature representation needed for deep learning.Comparative experiments show that the emotion vector constructed by this dictionary can effectively optimize the speed and memory consumption of text emotion classification without degrading the effectiveness of emotion classification.(2)For the problem that it is difficult to obtain the sentiment information of each modality and remove the redundant information,attention mechanism and auxiliary function,which are used to remove the redundant information,are introduced into the multimodal sentiment analysis model.The loss function of each modality is added to the total loss function as an auxiliary function with a certain weight to ensure that the influence of different modalities on the total sentiment classification is checked and balanced.The effect of the attention mechanism and the auxiliary function in the model is verified by ablation experiments.(3)Based on the above proposed algorithms,a sentiment analysis platform system is constructed to provide sentiment analysis functions for different intelligent products.The system uses EmoDic and AtM-DNN as the core algorithms to provide single text mixed sentiment classification,batch text mixed sentiment classification,and multimodal sentiment classification,while users can choose the corresponding models according to the accuracy and time requirements for sentiment classification.All sentiment classification functions will eventually visualize the sentiment classification results.