An unsupervised method to extract Chinese comment targets automatically

With the Internet developing, the quantity of comments accumulates rapidly, and sentiment analysis becomes a hot research task. A key preparation for sentiment analysis is to extract comment targets. In this paper an unsupervised method is proposed to judging the degree of their importance. We separate the method into two parts. One part uses explicit phrases' cooccurrence, calculating conditional entropy. We take nouns and adjectives into consideration, and make full use of their relevance. The other part mines the relationship between clusters and terms based on LDA result, using a modified PageRank method. The two algorithms are combined to rescore these candidate targets. In the end, we use two ways to test our LPCE (LDA, PageRank, and Condition Entropy) model's reliability, including word aspect and sentence aspect. Experiment results prove LPCE model is reasonable and effective.