Research Of Sentiment Analysis For Chinese Micro Blog Based On Conditional Random Field

In recent years, more and more people like to prefer the microblog express their attitudes and opinions for a hot commodity or event. Because these messages often have a strong emotional tendencies, it is very significant for government and business. However, microblog sentence is often short and expression is not standardized, and the current Chinese text processing tools are not well suited for microblog, which leads us difficult to extract emotional information from them. Moreover, the existing works ignore the information between Subjective classification and Polarity classification task, which eventually leads to the results of sentiment analysis results are generally not ideal. To this end, we present WDCRF (Word2vec DCRF), according to depth study of Chinese microblog features and key factors influencing sentiment analysis result. The details of our work are as follow:(1) Leveraging Word2vec to extend microblog, which is to find the Top-k similar words of the word in original microblog for extending original microblog. To begin with, given a weibo sentence, after word segmentation, we can obtain word sequence. Secondly, we utilize Word2vec to find the Top-k similar words of each word in the sentence. Finally, the first k similar words of each word are added to the sentence behind the original word sequence, so as to extend microblog. Experiments show that the extended microblog is likely to contain richer emotional information than the original sentence, and ultimately improves the algorithm performance.(2) Establishing the links between subjectivity classification task and polarity classification task. Firstly we utilize DCRF (Dynamic Conditional Random Field) model to establish the links between subjectivity classification task and polarity classification task, so that this two tasks can be conducted at the same time. Moreover, the DCRF model also can combine the dependence between neighboring sentences.Combining the above two parts, we propose to apply the DCRF model in the extended microblog. Experimental results show that the WDCRF method achieves much better final results than the state-of-the-art approaches, and it can better identify the emotional tendencies of Chinese microblog.