Study On Text Emotion Analysis Based On Supervised Learning

In recent years, social networks such as weibo, social networking, BBS, wiki, and the shopping platform have gathered a large number of users. These users are not only the information browser and recipient, but also the provider and transmitter. There are objective reports on people, things and affairs, and subjective expressions of them.How to analyze and deal with subjective emotion expression from different social networks has become an urgent problem. The issue of text emotion classification is paid more and more attention and becomes the research hotspot of the present.One of the most representative and successful methods is supervised learning method for text emotion analysis. However, they have many limitations when applied to deal with the problem of emotional ambiguity, combine emotion and implicit emotion.With the development of big data and deep learning technologies, deep features have become the focus of attention in the research field of natural language processing.Compared with the low-level features, which are made manually, the deep features can describe the emotional information contained in the text more entirely and precisely.This Ph.D. thesis dissertation describes a series of novel approaches to text emotion classification based on supervised learning technique. The aim of this thesis is to resolve the problem of the emotional ambiguity, emotion combination and implicit emotion, and improve the performance of emotion analysis system, providing them with additional linguistic information and deep features beyond the surface level of words, and promote the practical application process of emotion analysis system.The main research work of this thesis is as follows:Exploring multi-levels of emotion resources are a large part of this dissertation. For the words and phrases level emotion lexicon construction, we describe two methods:source-target translation dictionary based method and bootstrapping methods. The former uses the annotated English emotional dictionary to translate the English emotional vocabulary into the target Chinese emotional vocabulary. The latter starts from a set of small seed words, and use the Bootstrapping method to extend a larger emotional lexicon. We also establish the sentence level emotion corpus, and use manual annotation method and sentence alignment method from the bilingual corpus.The sentence level emotion corpus is based on manual annotation and the aligned English-Chinese parallel corpus, and then maps the English emotional sentences into the target Chinese emotional sentences. Additionally, we evaluate the quality of the above emotion lexicon and corpus.The thesis proposes a new phrase-based supervised model tackling the problem of emotional ambiguity and emotion combination. This model is new task decomposition for emotion analysis system, which contains three main subtasks:phrase segmentation, phrase level emotion classification and sentence level emotion classification. The first subtask detects and segment the phrases based on dependency grammar. The second subtask completes the phrase level emotion classification task,by applying two kinds of different CRFs-based model: extensions of CRFs and semi-Markov CRFs. The third subtask estimates the average emotion label of sentence via emotion shift rules. We then evaluate our models on two extrinsic tasks:sentiment analysis and emotion label tagging, and achieve better accuracy.The use of active learning method to improve emotion classification quality is also studied. Pool based active learning algorithm extended with combination of uncertainty and diversity techniques with successful results. Additionally, we evaluate the impact and quality of classification on the semi-Markov conditional random field's emotion analysis system. The method makes full use of the characteristics of the probabilistic graph model and the natural languages, and obtains better annotation performance under the condition of insufficient training corpus.Furthermore, the thesis proposes an emotion analysis model tackling semantic information generation through composition semantic features. Composition semantic model is constructed based on distributed semantic model and dependency grammar.A detailed study of the approach is conducted. This includes its initial development from a large amount of unlabeled data and main applications. We introduce two kinds of combination of semantic features in emotion analysis: supervised learning model based on the combination of composition semantic features and neural network application.