Research On Sentiment Computing For Text Data

With the rapid development of the Internet,the e-commerce platform and social network generate a large amount of text data.The collection,analysis and value mining of text data will generate huge commercial value.Affective computing is one of the main techniques for analyzing text sentiment,including sentiment object recognition,sentiment retrieval,sentiment classification,and emotional reasoning.Mining the opinions or sentimental tendencies expressed by the author from the text data can assist the observer in inferring and making decisions about relevant information.Although there have been a series of valuable research results in the field of affective computing in recent years,there are still many challenges,including:fine-grained sentiment object recognition,hidden sentiment object recognition,the problem of unbalanced data distribution,the problem of insufficient corpus of the neural network parameters training.This dissertation solves the related tasks of affective computing on text data,with the whole contributions mainly including:1.The problem of sentiment object recognition for text data.On the one hand,using the comment data in the car forum,we propose a method through constructing a two-level conditional random field.It aims at identifying the fine-grained sentiment objects in the text data.The recognition problem of sentiment objects is regarded as a word-level sequence mark problem in the method.By marking different tags in different positions in the sequence,we establish the information interaction between different mark sequences with a potential function.Then we construct a model on the two-layer conditional random field,in order to recognize fine-grained sentiment objects.Finally,experimental results demonstrate that the proposed method of the two-layer conditional random field can meet the requirements of the sentiment entity recognition of car reviews on fine-grained.On the other hand,we use the microblog theme data as the experimental object.In order to identify hidden sentiment objects,we propose a method of global conditional random field method.For the case where the sentiment object does not appear in the text,the implicit object is converted into an abstract target,which is added as a global node to the conditional random field model.Furthermore,the global condition random field(GLCRF)is verified by experiments under different data sets,where the recognition performance of our proposed method is better than NB,SVM and LLCRF algorithm.It further demonstrates that it is more effective to add two global variables to identify hidden sentiment objects.2.The problem of sentiment classification for text data.Firstly,in order to solve the problem of data distribution imbalance with serious subject bias,we propose a method combining Affinity Propogation algorithm,Word2vec and conditional random field model.Specifically,the Affinity Propagation is used to reduce the number of majority samples in the training set and reaches a relative balance of the training set;The Word2vec is used to expand the emotional information of the words;The conditional random field model of the complex features is used for training prediction.The experimental results show that the AWCRF method achieved better results than the SVM method and the BP neural network method in Chinese microblog sentiment analysis task.Secondly,.in order to solve the problem of the insufficient corpus of text data annotation,we propose a text sentiment analysis method embedding logic rules into the recurrent neural network.Specifically,the prior consciousness is extracted from the structured knowledge such as knowledge map,social graph or syntactic dependency tree and decomposed into a set of logic rules.Then the logic rule is embedded into the recurrent neural network through the feedback mask matrix.The experimental results of sentiment classification and the named entity recognition task show that the recurrent neural network embedded with logic rules performs better than the recurrent neural network in the case of the insufficient corpus.3.The problem of emotional reasoning for text data.Taking the review data text of mobile phones in the e-commerce platform as the experimental object,an emotional reasoning method based on path ranking algorithm is proposed.First,we use the crawler to grab the comment information in the product page of mobile phones in e-commerce platform.Then we pre-process the comment information,use the conditional random field model to perform the fine-grained sentiment object recognition,and integrate the sentiment object into NELL through relationship extraction.Taking the mobile phone and each component entity as the minimum analysis granularity,we employ the NELL format and the path ranking algorithm to mark and sort the paths between the entities.This is performed to infer the relationships between the sentiment objects or to infer the sentiment objects according to the relationships.In summary,we analyze the massive text data generated by the Internet platform and propose corresponding methods for the problems of fine-grained sentiment target recognition,the hidden sentiment target recognition,unbalanced data sets,insufficient label data sets,and emotional reasoning.The experimental results show that our proposed methods are effective.