Aspect Based Multilabel Text Classification for Identifying Dangerous Speech Twitter Text

As part of hate speech, dangerous speech is any expression that can increase the risk of committing violence against other people. So far, hate speech research only explains whether some sentences is categorized as hate speech. It does not explain aspects of the sentences that make them called dangerous speech. Aspects of dangerous speech are social context, historical context, dehumanization, the accusation in the mirror, women and children attack, loyalty to the group, and group threat. This study uses the multi-label text classification method to determine dangerous speeches on Twitter texts based on seven aspects. Then, we assign a weighted score from those aspects to differentiate dangerous and hate speech. Based on the test results show the best performance is the Naive Bayes method with label-based subset accuracy (±36 %), instance-based (average) accuracy (±86%) and classification accuracy (±77%). However, even though Naive Bayes has the best performance in terms of instance based (average) accuracy, the average difference between all methods with Naive Bayes is only ± 0.014, this indicates that other methods also produce quite good performance.