Sentiment Analysis On Chinese Short Texts

With the rapid development of mobile Internet, massive short texts with subjective information had emerged. Information of attitudes and public opinion trends included in short texts contain immeasurable commercial and social value. Subjective information expressed in the text, such as feelings, opinions and positions, will be analyzed and summarized in sentiment analysis tasks. At present, all aspects of sentiment analysis on Chinese short texts are in rapid development, but challenges always go with chances.In this paper, two methods of sentiment analysis on Chinese short texts are proposed, one is based on the searching of sentimental keywords, and the other one combining dictionary and rules is based on machine learning.In the first method, sentimental and topic keywords were selected at first. Then the short texts which were the most associated with the short texts for sentiment analysis were searched by the keywords combined with the emotional word vector which had the emotional information. The search result would decide the sentimental polarity of the short texts. Another method improved the feature extraction and weight calculation of the machine learning methods according to the characteristics of sentiment analysis tasks and Chinese short texts. An MCHI feature extraction algorithm and an MTF-IDF feature weighting algorithm were proposed since they were more suitable for sentiment classification task. Finally, the classification task was completed by Support Vector Machine. Emotional word vector information was added to the method based on the searching of emotional keywords. It reflects the effect of the emotional keywords and topic keywords. Experiment results show that the proper and efficient use of the limited emotional words and their emotional information was critical in sentiment analysis on Chinese short texts.The improved machine learning method completed the two-categories and three-categories classification tasks in the single-topic and multiple-topics datasets respectively. Experiment results also show that the method makes a great improvement.