Research On Sentiment Analysis For Understanding

With the rapidly development of Internet,the way people use Internet has been changed.People delivery information by Internet rather than a reader.It makes the information with sentiment has the opportunity to change the real world.How to analyze the sentiment in the internet? How to identify the classification of opinion? How to extract the sentiment relevant information from text? All those issues belong to sentiment analysis,which is a hot topic in nature language processing.This technology is also wildly used in product review analysis and public opinion analysis.However,there are still some unsolved problems in sentiment analysis.On the data level,the sentiment analysis focus on a specific topic.The size of training data limits the performance of sentiment analysis.Hence,the transfer learning based cross lingual/domain methods are proposed.However,how to handle negative transfer in transfer learning is a big issue in this task.On the semantic level,the recent methods usually use representation learning to capture features from text.However,the relation between sentiment components is ignoring and limits the understanding of sentiment.On the cognitive level,the sentiment analysis method focus on expression of sentiment rather than the cause of sentiment.In order to solve these problems,this research focuses on the tasks below:On the data level,in order to solve the negative transfer issue,this research proposes a Gaussian distribution based negative transfer detection method.In this method,a Gaussian distribution based class noise estimation method is proposed to detect missclassied samples in transfer learning.Then,in order to detect negative transfer,the class noise estimation result is used to estimate the performance of final classifier after transfer learning based on PAC theorem.In the task of cross lingual/domain sentiment analysis,the proposed method improves the performance of sentiment analysis.The curve of performance also reveals the increasing trend in transfer learning.However,this method is paradox in the application of real word data.This method is based on manifold assumption and Gaussian assumption,which can not establish simultaneously.In this work,this work propose a novel class noise estimation method which is based on sum of Rademacher distribution.In this method,we use sum of Rademacher distribution instead of Gaussian distribution to estimate class noise.Based on the estimation result,two approaches to handle class noise are proposed.Furthermore,this work proves that the optimal hypothesis on the noisy distribution can approximate the optimal hypothesis on the clean distribution using both approaches.Finally,this estimation is also deployed in transfer learning to identify negative transfer.The experimental result reveals that the class noise estimation method beats several strong baseline in application,including synthetic data and real word dataset.The application in transfer learning also improve the performance of negative transfer detection.It means that our estimation result is stable and reasonable.On the data level,this result also shows the effectiveness of transfer learningBased on transfer learning framework,this research designs a uniform architecture for representation learning in sentiment analysis in order to learn the representation of holder,target and opinion word for sentiment analysis.Representation learning is a hot topic in sentiment classification in recent years.However,the existing method did not consider the relations between language,the users of language and the target of comments.They only use the context to model language in their methods.This work makes use of a heterogeneous network to model the shared stances in product reviews and learn representations of users,products they commented and words they used simultaneously.The basic idea is to first construct a heterogeneous network which links users,products,words appeared in product reviews,as well as the polarities of the words.Based on the constructed network,representations of nodes are learned using a network embedding method,which are subsequently incorporated into a convolutional neural network for sentiment analysis.Evaluations on the product reviews show that the proposed approach achieves the state-of-the-art performance.Furthermore,this research proposes a new sentiment analysis framework which combine representation learning with negative transfer.This method improve the performance of sentiment analysis further.It shows the necessity of combination of data level research and semantic level research.On the cognitive level,this work presents a cause extraction method for sentiment analysis.Since there is no open dataset available,the lacking of annotated resources has limited the research in this area.Thus,this thesis first presents an annotated a dataset we built using SINA city news.Then,an algorithm which is based on memory network is proposed for cause extraction.In this method,inspired by convolutional networks,we propose a new mechanism to store relevant context in different memory slots by convolution operation to model context information.The proposed approach can extract both word level sequence features and lexical features.Performance evaluation shows that our method achieves the state-of-the-art performance on a recently released emotion cause dataset,outperforming a number of competitive baselines.The attention weight can also used to extract keyword in cause,which shows the interpretability of this method.On the cognitive level,this work promote the research of sentiment analysis and understanding.Besides,the fist public dataset for cause extraction constructed in this work can also propel the research of sentiment analysis.