Sentiment Analysis for Domain-Specific Texts

We develop a sentiment analysis system for domain-specific texts. The sentiment analysis estimates a polarity of a text, for example positive or negative. To develop such a system a dictionary, which consists of words and their polarities, is needed. The dictionary reflects specialists 'knowledge and affects a sentiment analysis precision. Hence, it is important to construct an appropriate dictionary. And making a dictionary needs much human cost generally. As a good dictionary is constructed, the human cost must be decreased. To achieve the goal our proposed approach uses a bootstrap method to decrease human cost and a chi(2) statistic to estimate a polarity of a word correctly. To evaluate a performance of the proposed approach we carried out an evaluation experiment using real stock market news, T&C news. We confirmed that the proposed approach could construct a dictionary but had some problems which were inappropriate words and that the dictionary did not have enough words.