Research On Target-level Sentiment Analysis Of Texts Based On Deep Learning

With the increasing participation of internet users,the Internet has become a centralized platform for internet users to express their opinions.The Internet has generated a large number of people's opinions,which are valuable research resources.Studying the sentiment of Internet users on public figures,products,and services can help governments to grasp the state of public opinions and further make decisions,and it has very high commercial value.In recent years,deep learning has achieved success in various natural language processing tasks.However,sentiment analysis of texts is a very difficult natural language processing task,which achieves a lot of attention from researchers.Therefore,it is very significant from both theoretical and practical perspective to study how to represent and extract people's opinions from the texts on the Internet with the theories of deep learning.This thesis focuses on target-specific sentiment analysis of texts.Specifically,the thesis focuses on two kinds of tasks: entity-level sentiment analysis of microblogging texts and aspect-level sentiment analysis of review texts.The main contributions and innovations of this dissertation are as follows:1.We propose an entity-level sentiment analysis method for microblogging texts and use it for opinion poll of public figures.For microblogging texts,we first define entity-level sentiment analysis for opinion poll of public figures.After that,we propose a sequence labeling method for entity-level sentiment analysis of microblogging texts.Further,we use recurrent neural network and its improved recurrent units to conduct sequence labeling models.At last,we construct a Chinese dataset with the data from hot topics in microblog for entity-level sentiment analysis.The statistical results of the dataset show that using sequence labeling to jointly extract entity and sentiment has obvious advantages than traditional pipeline methods.The experimental results show that the proposed sequence labeling method with recurrent network achieves a substantial increase of F-score compared with traditional pipeline methods.2.We propose a convolutional recurrent neural network-based method for entity-level sentiment analysis of microblogging texts.In view of the disadvantage of recurrent neural network in extracting the local contextual feature of words,we first propose a convolutional neural network-based sequence labeling method and explore the relationship of the label of a word and its local contextual feature.Further,we propose a convolutional recurrent neural network-based sequence labeling method,taking advantage of convolutional neural network to extract the local contextual feature of words and recurrent neural network to extract the global sequential feature of the text.Experimental results on entity-level sentiment analysis task show that the proposed convolutional recurrent neural network-based method performs better than both convolutional neural network-based method and recurrent neural network-based method,indicating that the label of a word depends on both the local contextual feature of the word and the global sequential feature of the text.3.We propose a hierarchical attention network for aspect-level sentiment classification of review texts.For review texts,we focus on aspect-level sentiment classification,which is the key task of aspect-level sentiment analysis.The model use an aspect attention layer to grasp the aspect information of the text and a sentiment attention layer to grasp the aspect-specific sentiment information of the text.In addition,we use the aspect information extracted by the aspect attention layer to help the sentiment attention layer extract the aspect-specific sentiment feature,aiming to gain better classification results.Further,we improve the attention mechanism of aspect attention layer to extract the aspect words.The experimental results show that the proposed model has obvious advantages over the existing attention network-based methods,indicating that extracting aspect information of a text can help the extraction of the aspect-specific sentiment feature.4.We propose an attention network-based framework and specific models for aspect-level sentiment analysis of review texts.According to the conclusion of 3,we propose an attention network-based framework for aspect-level sentiment analysis of review texts.After that,we propose a category-specific word embedding training method and train aspect-specific word embedding and sentiment-specific word embedding with a large domain-specific unsupervised dataset.Then we propose an attention network-based aspect detection method,which can use the attention weight to locate the aspect information of the text.We further use the aspect information to help an attention network extract aspect-specific sentiment feature for aspect-level sentiment classification.The experimental results show that category-specific word embedding performs better than general word embedding on both aspect detection task and aspect-level sentiment classification task.Further,the proposed aspect detection method achieves better result than existing methods.Finally,the proposed aspect-level sentiment classification methods achieve comparable accuracies with the hierarchical attention network-based method without labeling the aspect words in the texts.In summary,this thesis focuses on target-specific sentiment analysis of texts.We study the key technologies for entity-level sentiment analysis of microblogging texts and aspect-level sentiment analysis of review texts based on deep learning theories.These technologies and research results have important theoretical significance and application value for public opinion situation analysis,opinion poll of public figures,and product market analysis,et.al.