Font Size: a A A

Chinese Sentiment Analysis Based On Attention Mechanism

Posted on:2020-04-05Degree:MasterType:Thesis
Country:ChinaCandidate:S D TangFull Text:PDF
GTID:2428330590973537Subject:Applied statistics
Abstract/Summary:PDF Full Text Request
Sentiment analysis,also called opinion mining,refers to the process of classifying the feelings of text,and on the basis of the induction and reasoning of the meaning of the text,for example,the evaluation of a certain movie or the evaluation of a certain product is emotionally positive,negative or neutral,etc.,to analyze the popularity of a movie or product.Studying sentiment analysis is very important for countries,companies and even individuals.Sentiment analysis is a classification problem in supervised learning,and it is also a problem of sequential modeling.However,it is worth noting that most of the domestic and foreign sentiment analysis methods are mature and effective in the application of English texts,while the research on the application of Chinese text analysis is relatively rare.The main reason is that the evaluation resources of Chinese texts are scarce and cluttered.At the same time,compared with English,the complexity of Chinese word segmentation is huge,so there are huge challenges in processing.Therefore,this topic will build an effective model for Chinese sentiment analysis based on the attention mechanism,focusing on the emotional analysis of Chinese text.This paper explores and proposes a neural network model different from convolutional neural network or recurrent neural network-SAE(self-attention network)by testing the Chinese sentiment analysis data set.The paper will introduce two self-focused network architectures called SAE and SAE2.SAE was developed on the work of the SAN,which aims to create the simplest SAN architecture that can be compared to basic circular neural networks and convolutional neural network architectures.SAE2 is an extension of our SAE architecture-using a multi-attention mechanism based on SAE,L2 regularization,replacing simple pooling operations.We used SAE2 in Section 5.2 to get better results from the sentiment analysis dataset.After a detailed comparison of the simple SAE,BiLSTM,and BERT algorithm models on four sentiment analysis datasets,we found that SAE(Self-Attention Network)can achieve higher classification accuracy than the BiLSTM and BERT models on the Chinese dataset.At the same time,it also has good features such as faster training and inference,and fewer parameters in comparison to recurrent neural network.
Keywords/Search Tags:sentiment analysis, Self-Attention Network(SAN), neural network
PDF Full Text Request
Related items