Chinese Text Sentiment Analysis Based on BI-GRU and Self-attention

Humans have been trying to get machines to understand human emotional states, and using natural language processing techniques for sentiment analysis is the mainstream solution. In order to solve the problem of high complexity and low efficiency in Chinese LSTM-based Chinese text sentiment analysis, this paper uses Bi-GRU (bidirectional GRU neural network) and attention mechanism model to analyze Chinese text sentiment. First of all, the model can extract deep features of the text, and combine the context of the sentence to learn the text features more accurately. Secondly, the model introduces Multi-Head Self-Attention, which reduces the dependence on external parameters, assigns weights to word vectors, highlights text features, and pays more attention to the internal dependencies on sentences. Experiments show that the Chinese text sentiment analysis based on the model can get 87.1% accuracy.

I. Introduction

The Internet has penetrated into people's work and life. People can express their opinions and opinions on many interactive platforms. In these vast opinions and opinions and shopping evaluations, you can mine interesting information and understand people's things. As an important task in the field of NLP (Natural Language Processing), the use of computers to complete automatic analysis of texts, text sentiment analysis or opinion mining is also a hot topic of current research. In 2012, Wei Wei et al. summarized the principle of Chinese text sentiment analysis and its related research techniques^[1]. The current text sentiment analysis methods mainly include traditional analysis methods based on sentiment lexicons and sentiment analysis methods based on deep learning. From the beginning, basic machine learning algorithms such as Logistic Regression, Maximum Entropy, and Naive are adopted. Naive Bayes and Support Vector Machines (SVM). In 2009, Liu Youping proposed the construction of an emotional lexicon of Chinese sentiment words of Chinese sentiment analysis^[2]; In 2012, Chen Xiaodong proposed an emotional sentiment analysis based on sentiment lexicon^[3]; in 2012 Liu Lu et al. proposed Chinese microblog emotion classification based on machine learning^[4]; Zhou Jie et al. proposed machine-based learning. Emotional analysis of online news commentary^[5]; In recent years, with the development of deep learning, using neural networks to learn text features has become the mainstream technology in the field of natural language processing. In 2014, Liang Jun et al. used the recurrent neural network sentiment transfer model to enhance the capture of text relevance^[6]; in 2017, Cheng Yu proposed to apply the two-way LSTM model based on attention mechanism to the emotional classification of Chinese commodity reviews^[7].