Research On Text Sentiment Analysis Algorithm Based On Deep Learning

With the vigorous development of 5G and the mobile Internet,people have left a lot of text information on the Internet.It has very important value for the processing of these texts.The analysis of the sentiment polarity information contained in the text is the main research content of this thesis.We understand some difficulties and key points in the current text sentiment analysis tasks by investigating and researching some existing text sentiment analysis methods.Therefore,this thesis proposes two models for sentiment analysis of texts with different granularities,which are:1.A text sentiment analysis model based on the interactive multi-head attention aspect level of Transformer-LSTM structure.For existing methods,the text and target words are often modeled and extracted separately.Ignoring the problem that there is rich correlation and interaction information between the two,an interactive multi-head attention mechanism is proposed to enhance the interaction between aspect words and context,and simultaneously calculate the attention weights of the two aspects,including the context for the aspect.The attention weight of the word and the attention weight of the aspect word depend on the context.In addition,different from the way of using recurrent neural networks combined with attention mechanisms in many models,we combine the encoder structure of Transformer with LSTM to obtain a more capable feature extraction module.In the label mapping stage,the label smoothing coefficient is introduced,which makes the model have stronger generalization ability.Finally,various experiments are carried out to analyze the whole and local modules of the model.2.A sentence-level Chinese text sentiment analysis model based on graph convolutional network of syntactic dependency and part-of-speech tagging.At present,the mainstream text sentiment analysis work focuses on English corpora,and when these models are applied to Chinese text,they often fail to achieve the effect on English datasets.English and Chinese have big differences in the grammar of their languages themselves.In response to this problem,this thesis proposes to add Chinese syntactic dependencies and part-of-speech information to the sentiment analysis model to enhance the performance of the model on the Chinese dataset.The part-of-speech information and the word embedding vector are spliced together,and then fully featured extraction is carried out through the Bi-GRU network.At the same time,since the syntactic dependency is a kind of graph-like non-Eulerian data,the traditional neural network cannot process this kind of data.This thesis introduces a graph convolutional network,constructs a graph network from the syntactic dependency,and then combines the text and part-of-speech information in the graph structure,which utilizes graph convolutional networks for further modeling and feature extraction.In order to maximize the performance of the model,the Chinese Wikipedia dataset is also used to pre-train the Word2 Vec model,which enriches the text representation of word embedding vectors.This thesis also tries to use the percentile pooling method,and compares multiple pooling methods in this model.Finally,comparative experiments with different existing models are set up,and the effectiveness of each module of the model is verified by ablation experiments.