Research On Public Opinion Analysis Algorithm For Multimodal Social Network

The emotions contained in social network data are essential for obtaining the thoughts and opinions of network users,and their research results have successfully provided effective references for product promotion,policy formulation,and other fields.One of the difficulties of the current social network text public opinion sentiment analysis research is that the text is not used to dig out enough effective emotional features.In addition,images in social network platforms often imply user sentiment information,but image sentiment is often subjective.If sentiment analysis is performed on images independent of the text,this is not very helpful for public opinion sentiment analysis.Therefore,this thesis conducts the following research on social network public opinion sentiment analysis from two levels: single-modality and multi-modality:1)Most of the data in today's social networking platforms are in the form of text,and most of the text sentiment studies focus on using deep learning models to obtain sentiment features of the text,and only consider the contextual information of utterances but rarely consider using syntactic information of text utterances.Accordingly,this thesis establishes a graph convolutional sentiment classification model based on Chinese syntactic dependency information by combining Chinese text grammar structure.The contextual feature representation of the given text is learned and combined with the syntactic dependency information of the given text,which is jointly inputted into the GCN neural network to classify the sentiment polarity of the text.2)To obtain the features of images and text in multimodal data,this thesis introduces a 16-layer VGG network pre-trained on the Image Net dataset for parameter migration,extracts the local feature representation of images by parameter fine-tuning,and then uses a two-layer gated neural network and an attention mechanism at the word level and an attention mechanism at the sentence level to obtain the sentiment feature representation of sentences.3)For feature fusion between different modalities,an attention mechanism is used to assign different weights to image and text features,to measure their contribution to user sentiment analysis,and to obtain a text sentiment feature representation based on image features,which is then fed into the classifier to achieve multimodal sentiment classification.4)An opinion sentiment analysis system for social platforms is designed and implemented.The sentiment analysis algorithm proposed in this thesis is applied to the sentiment analysis module of the system.By receiving user input on the page,the corresponding analysis results are returned,and the analysis results are displayed in a visual way.In summary,this thesis investigates sentiment analysis algorithms in social networks at two levels,unimodal and multimodal,and constructs a text sentiment analysis model and a picture-text multimodal sentiment analysis model in social networks.By comparing the experiments with the existing models on several datasets,the accuracy of the two models exceeded the benchmark models by 4.45% and 5.2%,respectively,which verified the effectiveness of the two models.