Evaluating Novel Features for Aggressive Language Detection

摘要

The widespread use and abuse of social media and other platforms to voice opinions online has necessitated the development of tools to regulate this exchange of opinions in light of ethical and legal considerations. In this work, we aim to detect patterns of aggressive language to gain insight into what differentiates it from non-inflammatory language. Of particular interest are features of comments that, taken together, allow this distinction to be made automatically. To that end, we employ feature selection techniques to find optimal feature subsets. We apply the feature selection and model evaluation process to two independent datasets. Depending on the dataset and model type, between 3 and 19 features are enough to outperform the full set of 68 features. Overall, the best F_1 scores per dataset are 89.4%, using 35 features with a Gaussian SVM and 82.7%, using 17 features with a linear SVM.