Deep Emotion Feature Extraction For Text Sentiment Analysis

As one of the core tasks in natural language processing,text sentiment analysis is to extract the sentiment semantic feature of text by analyzing,learning and summarizing the context information of unstructured text,so as to explore the sentiment tendentiousness expressed in the subjective text.Traditional machine learning algorithms mainly focus on the feature engineering such as the construction of manual dictionary and the design of grammar rules,which can represent the sentiment feature of the raw text.However,the feature engineering is too complex to deal with large-scale text data.In recent years,with the application of deep learning in the field of natural language processing,the feature extraction model based on deep neural network can improve the effect of text sentiment analysis without relying on complex feature engineering.Nevertheless,the deep learning-based model still has a series of defects,such as ignoring the effective use of sentiment knowledge,the unicity of extracted sentiment feature,and the difficulty in capturing the dependence between different sentences in large-scale text.As a result,there is a lot of room to improve the performance of extracting text sentiment feature.To alleviate the aforementioned limitations,in this dissertation,we improve the ability of deep neural network to extract the sentiment feature of text from two aspects:knowledge embedding and model architecture.This dissertation mainly contains the following two parts:(1)Aiming at the problem that deep neural networks cannot make full use of sentiment knowledge and specific deep neural network can only extract relatively single text semantic features,we propose a text sentiment analysis model based on sentiment attention mechanism and deep neural networks.This model leverages the sentiment lexicon as sentiment knowledge attending to context words in the word embedding matrix via an attention mechanism,so that deep neural networks can learn deep text semantic features closely related to sentiment polarity under the guidance of sentiment knowledge.In addition,this model can effectively combine the gated recurrent unit neural network and the convolutional neural network to give full play to their respective advantages,so that the text sequence information extracted by the gated recurrent unit neural network and the text local information extracted by the convolutional neural network are fused together to enhance the sentiment feature representation of the text.(2)In view of the fact that the convolutional neural network is limited by the width of the convolution kernel,it cannot fully model the dependencies between different positions of text,and the computational complexity of the convolution layer is too high to process the document-level text data,we propose a text sentiment analysis model based on bidirectional long short-term memory neural network and multi-head attention mechanism.This model improves the sentiment attention mechanism proposed by the first work so that it can learn the high-recognized sentiment word embedding matrix.At the same time,under the premise of protecting the input and output sequence information in neural network layers,this model maps the text sequence to multi dimensional subspaces where it can further extract the long-distance dependency relationship between different positions in the text data.In addition,we use a hierarchical network structure to model document-level text data,which enables it to mine deep sentiment semantic features in the document under the guidance of sentiment knowledge.