Abstract
Automatic identification of hateful and abusive content is vital in combating the spread of harmful online content and its damaging effects. Most existing works evaluate models by examining the generalization error on train-test splits on hate speech datasets. These datasets often differ in their definitions and labeling criteria, leading to poor model performance when predicting across new domains and datasets. In this work, we propose a new Multi-task Learning (MTL) pipeline that utilizes MTL to train simultaneously across multiple hate speech datasets to construct a more encompassing classification model. We simulate evaluation on new previously unseen datasets by adopting a leave-one-out scheme in which we omit a target dataset from training and jointly train on the other datasets. Our results consistently outperform a large sample of existing work. We show strong results when examining generalization error in train-test splits and substantial improvements when predicting on previously unseen datasets. Furthermore, we assemble a novel dataset, dubbed PubFigs, focusing on the problematic speech of American Public Political Figures. We automatically detect problematic speech in the $305,235$ tweets in PubFigs, and we uncover insights into the posting behaviors of public figures.
Abstract (translated)
URL
https://arxiv.org/abs/2208.10598
