KOSAC: A Full-Fledged Korean Sentiment Analysis Corpus

KOSAC: A Full-fledged Korean Sentiment Analysis Corpus Hayeon Jang, Munhyong Kim, and Hyopil Shin Department of Linguistics, Seoul National University Gwanak-no Gwanak-gu, Seoul 151-741 {hyan05, likerainsun, hpshin}@snu.ac.kr Abstract This paper aims to introduce the Korean Sentiment Analysis Corpus named KOSAC. KOSAC is a corpus consisting of 332 news articles taken from the Sejong Syntactic Parsed Corpus. These sentences have been manually-tagged for sentimental features. The corpus includes 7,713 sentence subjectivity tags and 17,615 opinionated expression tags based on the annotation scheme called KSML which reflects the characteristics of the Korean language. The results of sentence subjectivity and polarity classification experiements using the corpus show the wide possibilities of application the KSML scheme and the tagged information of the KOSAC comprehensively to other corpus. What is innovative about our work is that it pulls together both the concept of private states and nested-sources into one linguistic annotation scheme. We believe that this corpus could be used by researchers as a gold standard for various NLP tasks related to sentiment analysis. 1 Introduction There has been much research on the automatic identification and extraction of opinions and sentiments in text. Researchers from many subareas of Artificial Intelligence and Natural Language Processing (NLP) have been working on the automatic identification of opinions and related tasks. To date, most such work has focused on opinion, sentiment or subjectivity classification at the document or sentence level. A common sentiment analysis task is to classify documents or sentences by whether they are subjective or objective, and, if the target text is subjective, to classify it as positive or negative (Pang et al., 2002; Wiebe et al., 2005). Along with these lines of research, a need for corpora annotated with rich information about opinions and emotions has emerged. In particular, statistical and machine learning approaches have become the method of choice for constructing a wide variety of practical NLP applications. These methods, however, typically require training and test corpora that have been manually annotated with respect to each language-processing task to be acquired. As such a resource, the Multi- perspective Question Answering (MPQA) Opinion Corpus plays an important role in sentiment analysis. The goal of this paper is to introduce the Korean Sentiment Analysis Corpus, KOSAC 1. We received two years of support (May, 2011- April, 2013) in this corpus construction project from the Korean Research Foundation (KRF). In the first year of the project, we focused on a fine- grained annotation scheme called KSML (Shin et al., 2012) that identifies key components and properties of sentiments based on solid theoretical background. The annotation scheme has been employed in the manual annotation of a 7,713-sentence corpus of 332 news articles from the Sejong syntactic parsed corpus. This manually-tagged corpus includes 17,615 opinionated expression tags. The remainder of this paper is organized as follows. Section 2 gives an overview of KSML focused on differences with the annotation scheme of the MPQA. Section 3 describes observations about KOSAC. Section 4 presents the results of subjectivity and polarity classification experiments …

Content truncated to 500 words