Detecting Nasty Comments from BBS Posts

Detecting Nasty Comments from BBS Posts Tatsuya Ishisaka and Kazuhide Yamamoto Nagaoka University of Technology 1603-1, Kamitomioka, Nagaoka, Niigata 940-2188 Japan {ishisaka,yamamoto}@jnlp.org Abstract. We propose a method to detect Japanese nasty comments from posts on bulletin board systems (BBS). Nasty comments can cause many social problem, because they express potentially harmful words and phrases. There are methods to recognize harmful words, but they are insufficient. Therefore, we present a method for detecting such comments on a BBS with many posts using an n-gram model. In addition, we compared our method with a support vector machine (SVM) that is based on nasty words. As a result, we detected nasty comments that are different to those by the SVM. We also observe higher detection accuracy by combining two methods. Keywords: sentence detection, nasty comment, n-gram, SVM, BBS 1 Introduction We focus on Japanese nasty comments that are seen on the Web, particularly, on bulletin board systems (BBS). BBSs are used mainly for information-sharing, consultation, and discussion; how- ever, unfortunately we also see some nasty comments posted on them. Recently, young people, such as primary and secondary students, have been posting such comments. Furthermore, there are many cases where these comments drain the victim emotionally for a long time, and can negatively affect the victim’s social life. In a worst-case scenario, the victim commits suicide. These com- ments are increasing every year all over the world, and have become a social problem, which has been classified as cyber-violence or cyber bullying. This is an effect of the insufficient regulation of the Internet. The nasty comments must be managed automatically. There are companies in Japan that patrols Web pages manually to find nasty comments. However, manual patrolling is very expensive, so the process to detect nasty comments on BBSs should be automated as much as possible. Research on detecting nasty comments is similar to research on classifying harmful contents and detecting spam blogs. The POESIA project (Hidalgo et al., 2002)(Hidalgo et al., 2003), which was funded by the European Commission, created a filter for harmful content. This filter classifies whether Web text is pornographic, and is adapted for the English, Italian, and Spanish languages. In other cases, the NET PROTECT project (Grilheres et al., 2004) developed a text classifier for harmful information. Harmful information in Web texts, including pornography, bomb-making, drugs, and violence, is classified based on machine learning using an support vector machine (SVM) (Lee et al., 2007). In addition, Kolari et al. discussed how SVM models based on local and link-based features can be used to detect spam blogs (Kolari et al., 2006). These studies achieved a high level of accuracy, and several companies are now introducing filtering services against harmful sites; therefore, we can expect to find such harmful sites. How- ever, this filtering is limited only to harmful words that deal with subjects such as pornography and drugs. In contrast, because there can be many patterns for nasty comments, they are difficult to detect. Since nasty comments can express harmful …

Content truncated to 500 words