| With the development of Internet technology,human society is entering a highly informative and intelligent stage.As a concise information carrier,texts have always been the main medium of exchanging information between us.At present,a large amount of text data has been accumulated on the Internet,which contains people's emotional tendencies in terms of products and services.Studying the emotional tendencies in massive texts can help merchants or service providers to make relevant decisions,which has important commercial value and academic research significance.In recent years,aspect-based sentiment analysis(ABSA)has attracted a lot of attention between scholars.This task aims to classify the sentiment expressed in different aspects of the text,which is closer to the practical application.According to the aspect expression,ABSA can be divided into Aspect Term Sentiment Analysis(ATSA)and Aspect Category Sentiment Analysis(ACSA).The traditional solution of ABSA is based on the sentiment dictionary method and machine learning method based on feature extraction.While the performance of the method based on a sentiment dictionary is heavily depended on the quality of the dictionary,and machine learning methods based on feature extraction also require heavy feature engineering.In order to solve the problems in the traditional methods,the deep learning methods can automatically learn and extract task-related features through well-designed models and break through the performance bottleneck of the traditional methods.Therefore,this thesis mainly studies the problem of aspect-based sentiment analysis based on deep learning related models and applies the model to the practical scenario.The main work and achievements of this thesis are as follows:(1)A hybrid neural network is proposed to solve the problem of insufficient feature extraction in aspect-term sentiment analysis.This model extracts features from the text by mixing the LSTM-attention feature extraction module and the CNN feature extraction module,which can simultaneously use global and local sentiment semantics of the text.Through experiments on SemEval 2014 Datasets,the algorithm achieves 74.7%and 79.9%accuracy on laptop and restaurant Datasets respectively and is superior to the same type of baseline model,which proves the effectiveness of the hybrid neural network.(2)The improved model based on BERT(Bidirectional Encoder Representation from Transformers)and the strategy of long text interception are proposed and applied to the AI Challenger 2018 Chinese dataset to improve the performance of coarse and fine-grained aspect category sentiment analysis.The texts in this dataset are long texts at the paragraph level,and each text contains multiple fine-grained aspects' emotional tendencies,which belongs to the multi-label classification task.Two improved methods are proposed based on BERT:First,an additional attention layer is added to the fine-tuning structure of BERT,and all the features in the output encoding layer of BERT are further extracted and utilized.Secondly,the sentence pair input method in BERT is used to process aspect-category sentiment analysis.Therefor the multi-label classification problem can be transformed into a multi-class classification problem.For the problem of long text redundancy,this thesis also proposes a text truncation method based on TFN in data preprocessing,which is used to select the sentences related to fine-grained aspects in a long text,so as to reduce the redundancy of text and the interference of noise.Finally,the results of the comparative experiment show the importance of TFN and the superiority of the improved BERT model.(3)Using improved BERT model to design a web application for sentiment analysis of online restaurant reviews.The application can obtain reviews from online merchants,and use the BERT improved model on the back-end to perform real-time fine-grained sentiment analysis,and finally process the results and generate reports to display in the front-end interface. |
PDF Full Text Request