Abstract:
Due to technological revolution over the years,
bullying which was confined to physical boundaries has now
moved online. Denigration or insult is one form of
cyberbullying. According to Sri Lanka Computer Emergency
Readiness Team, social media cyberbullying incidents are
escalating. Insulting words are dynamic, and same word can
have several meanings according to the context. Simply because
a comment contains such a word, it cannot be classified as
bullying. Hence, when labeling comments, simple keyword
spotting techniques are inadequate. Other languages have
addressed this issue using lexical databases such as WordNet
which provides synonyms and homonyms of words. Since there
is no proper lexical database developed for Sinhala language,
detecting a word as bullying is a challenge. Therefore, we used
rules to overcome this issue. Twitter comments with profane
words were collected, outliers were removed, and remaining
tweets were pre-processed. To determine insult in the text, five
rules were used for feature extraction. Afterward, we applied
Support Vector Machine (SVM), K-nearest neighbor (KNN)
and Naïve Bayes algorithms. The results show that SVM with an
RBF kernel performs better with an F1-score of 91%. Novelty
of this research is the focus on Sinhala language cyberbully
detection which has not been addressed before.
