Emotion-enhanced Word Representation Model And Its Applications In Sentiment Analysis

With the rapid development of Internet, there have been a lot of texts on the network, automatically analyze and process these texts using Natural Language Processing (NLP) technology is very important. As an important part of NLP, Sentiment Analysis (SA) task is one of the most important parts to get the aim of machine understanding text automatically. In recent years, there are a large number of researchers using various statistical algorithms to tackle sentiment analysis tasks. With the features of non-structural natural language, emotion complexity and the curse of dimensionality, SA tasks have faced severe challenges. Word representation is a dimension reduction method using lower dimension to represent the semantic means embedded in texts. It has become the hot topic in the field of NLPs. Exiting word representation training algorithms aim to semantic information while semantic information and emotional information are various. Which is one main reason of lower efficiency when using these exiting training algorithms directly for sentiment analysis tasks.Word representations for sentiment analysis tasks are still preliminary. There are a few training algorithms for word representation with emotion. The paper proposes a new emotional word representation model. The model called Emotion-enhanced wOrd Representation Model (simply EeDOM). This model combines semantic information and emotional information into the emotion-enhanced word representation. The emotion-enhanced word representation is verified in analyzing sentiment polarity discrimination and building sentimental lexical resource. The contributions of the paper can be summarized as follows:1) The author proposes an emotion-enhanced word representation model. First, according to the emotional features, an emotion-enhanced probabilistic language model is suggested. Secondly, based on the probabilistic language model, an emotion-enhanced neural network is described. An emotion-enhanced word representation can be obtained from this model.2) The author proposes a SVM training algorithm which is effective to word representation. The algorithm give a solution of the conflict between the uncertainty word number of sentence and the exact feature dimensions of SVM. The algorithm based on the n sliding window protocol, which is an improvement of N-gram feature selection. The algorithm can effectively improve the accuracy of the sentiment polar discrimination task.3) The author suggests a sentimental lexical resource construction algorithm which based on the emotion-enhanced word representation. The algorithm based the DBSCAN clustering algorithm, using word representation cosine distances to measure word emotional distances. Compared with other construction algorithms, using the suggested algorithm can get a better sentimental lexical resource for the sentiment classification task.Several experiments based on a real data sets (User comments on movie review website "Douban") are conducted.. The data set contains a total of 643,242 reviews, involving 2,737 movies. Experimental results show that the proposed emotion-enhanced word representation model can significantly improve the accuracy of sentimental polarity discrimination task. The EeDOM model can reach a 78.93 score of The Micro-F1 measure, far effective to the existing feature algorithms. Furthermore, the emotion-enhanced word representation in emotional similarity also shows a good performance. Based on the similarity of emotion-enhanced word representation, the sentimental lexical resource can obtain better results from the sentimental polarity discrimination.