Research On Emotional Analysis Model Based On Deep Learning

With the rapid development of society,China’s economy has entered a stage of high-quality development.Enterprises are increasingly paying attention to user feedback on the experience of their own product effects.Text can most intuitively reflect user preferences.Therefore,studying the content of the text well is of great significance for the sustainable development of an enterprise.This paper is committed to studying the direction of emotion analysis in the field of natural language processing.By using deep learning technology,it improves the traditional text classification model,aspect level emotion analysis and small sample data analysis model,solves the limitations of existing models,and improves their effectiveness in practical applications.Overall,the main work of this article is as follows:(1)Aiming at the problems of limited recognition ability and training time doubling with input length in traditional text classification models,a coarse-grained emotion analysis model BLAT(Bi-LSTM with Additive Attention and Text CNN)based on text summary extraction was proposed.The BLAT model introduces the additive attention mechanism of Fastformer to replace the self-attention mechanism of Transformer,enabling the model to have excellent training speed for long text training without losing accuracy.Secondly,the model forms two-way features by abstracting the original text data,integrates short-term memory network and convolutional neural network to form a multi-scale feature extraction network,and is verified on the Chinese e-commerce review dataset through experiments.The accuracy rate can reach 92.26%,which is better than the current mainstream models.(2)A new aspect level sentiment analysis model SEPGCN(Synthetic Embedding Position and GCN)based on syntactic dependency tree and graph convolutional network is proposed to address the issues of missing position information and single feature extraction scale in aspect level sentiment analysis models.The model constructs positional features through the absolute distance of the syntactic dependency tree and the relative distance between aspect level phrases and attributes.The attention mechanism is used to enhance semantic correlation,and the positional features are fused with it.Finally,the model is transformed into a graph convolutional network to find associated nodes and perform classification output.Through experimental verification on Res-14 and Laptop-14 in Sem Eval-2014,the final result has a certain degree of superiority compared to the comparative model.(3)In view of the poor performance of the current deep learning network model on small sample data sets and the problem of overfitting,a small sample classification model PBFT(Prompt BERT Line Tune)based on template hint learning is proposed.This chapter uses the Prompt method to process the original text and generate a template similar to cloze,which ultimately establishes strong contextual relevance by predicting its content.The model uses the BERT word vector as the feature extraction method of the model,and establishes an emotion vocabulary in the emotion mapping layer of the model to improve the model recognition efficiency.The results of selecting small sample datasets from the aspect level for extension show that the Prompt template learning method can enhance contextual semantic relevance,and has reference significance for classification tasks.