Flame prediction based on harmful expression judgement using distributed representation

Abstract

In recent years, flaming on social media has been a problem. To avoid flaming, it is useful for the system to automatically check the sentences whether they include the expressions that are likely to trigger flaming or not before posting messages. In this research, we target two harmful expressions. There are insulting expressions and the expressions that are likely to cause a quarrel. Firstly, we constructed a harmful expression dictionary. Because a large cost requires to collect the expressions manually, we constructed the dictionary semi-automatically by using word distributed representations. The proposed method used distributed representations of the harmful expressions and general expressions as features, and constructed a classifier of harmful/general based on those features. An evaluation experiment found that the proposed method could extract harmful expressions with accuracy of approx. 70%. On the other hand, it was found that the proposed method could also extract unknown expressions, however, it wrongly extracted nonharmful expressions at a rate of approx. 40%.