Abstract
Conference Title: 2018 International Conference on Machine Learning and Cybernetics (ICMLC) Conference Start Date: 2018, July 15 Conference End Date: 2018, July 18 Conference Location: Chengdu, China In this paper we present a proposal to address the problem of the pricey and unreliable human annotation, which is important for detection of hate speech from the web contents. In particular, we propose to use the text that are produced from the suspended accounts in the aftermath of a hateful event as subtle and reliable source for hate speech prediction. The proposal was motivated after implementing emotion analysis on three sources of data sets: suspended, active and neutral ones, i.e. the first two sources of data sets contain hateful tweets from suspended accounts and active accounts, respectively, whereas the third source of data sets contain neutral tweets only. The emotion analysis indicated that the tweets from suspended accounts show more disgust, negative, fear and sadness emotions than the ones from active accounts, although tweets from both types of accounts might be annotated as hateful ones by human annotators. We train two Random Forest classifiers based on the semantic meaning of tweets respectively from suspended and active accounts, and evaluate the prediction accuracy of the two classifiers on unseen data. The results show that the classifier trained on the tweets from suspended accounts outperformed the one trained on the tweets from active accounts by 16% of overall F-score.