Application Of Clustering Algorithm And Convolution Neural Network In Text Emotion Analysis

As the popularity of social networks and online shopping platform, there are more and more user-generated content, which contain some useful information, including user opinions, reviews, emotions and attitude. It has attracted so much attention to automatically identify sentiment information expressed through words and sentences. Sentiment analysis is a general term indicating a series of problems in natural language processing, and its main objective is to analyze the sentiment information implied in the texts. This sentiment can be represented in some discrete categories, such as positive, negative, neutral and multiple emotions, and it can also be represented by continuous sentiment intensity. As the limitation of text length, traditional approaches to sentiment analysis of short texts often suffer from feature sparsity problem when present texts as a feature vector, which may lead to reduce the performance of classifier and increase the complexity of models. Besides, for continuous sentiment prediction, traditional methods tend to build on sentiment lexicons or other external resources. However, these resources are often difficult to keep update to capture the new phrases and informal expressions, which would reduce the performance of these models when dealing with informal texts.In this thesis, we firstly introduce the categorical approach of sentiment analysis technique and propose a new sentiment analysis method for short texts based on clustering algorithms to reduce the sparsity problem of feature representation. The purpose of using clustering algorithms is to cluster short texts according to their sentiment similarity. The clustered texts are then used to train a classifier such that the sparsity of feature vector is reduced. Then, deep learning is introduced, and based on that, we propose a serious of convolutional neural network architectures for continuous sentiment intensity prediction. The convolutional neural networks can automatically learn the representation of texts and do not depend on any external resources to avoid the small coverage problem of sentiment lexicons. Finally, we conduct experiments on English and Chinese datasets and make some analysis about the experiment result, which show the effectiveness of clustering based method for sentiment classification of short texts, and the good performance of deep learning method for sentiment intensity prediction.