부산대학교 도서관

Due to the exponential growth of the number of texts available online, automatic text mining techniques have been receiving much attention recently. Pattern-based text mining techniques have been specially studied because they allow users to extract points of interest from text automatically. However, the quality of obtained results is highly dependent on the quality of the rules that are used in the mining process. Currently, the rules that are more effective are the ones that are defined manually, but the creation of such rules is resource-consuming as it requires experts to define them. In this paper, we develop a method of automatic rule definition for pattern-based text mining. Our method is investigated in terms of aspect classification and sentiment analysis. We also examine combining pattern-based text mining with machine learning techniques. The evaluations were performed using two text data, namely, those mentioning API from Stack Overflow and those of Amazon review on PC peripherals. Evaluation results showed that meaningful rules were defined automatically by our method. We also show that our methodology can be used together with manually defined rules, resulting in outperforming other methods.