Authorship attribution of Korean texts by using phrase patterns

Abstract

When we adopt machine learning for text classification, the process is exactly the same as the usual machine learning process; however, a textual classification requires its own unique pre-process in order to derive feature datasets. Therefore, proposals for new feature datasets are an important research topic in stylometrics. In this paper, we propose to use phrase patterns to establish a feature dataset for authorship attribution of Korean texts. The main idea of phrase patterns is to synthesize the tag and BUNSETSU(Jin, 2013). Although there are large similarities between Korean and Japanese, for example, the order and structure of sentences as well as the role of JYOSHI, there is no research on phrase patterns of Korean texts to date. We focused on the fact that GOSETSU('Eojeol' in Korean) plays the same role as BUNSETSU in Japanese. The corpus comprised 160 editorials from 4 authors and 200 essays from 10 authors. As a result of a simulation study with 5 feature datasets and 5 classifiers, the use of phrase patterns showed a good performance with respect to the authorship attribution in Korean texts. Moreover, there was no statistically significant mean difference between the use of phrase and the non-phrase patterns as input datasets for four of the five classifiers. Out of the five classifiers, RF yielded the highest accuracy, which is consistent with previous research studies on Japanese texts. A noticeable result is that LMT showed better performance than that of SVM. © 2017 International Information Institute.