The Thin Line Between Hate and Profanity

摘要

Hate speech can be defined as a language used to demean people within a specific group. Hate speech often contains explicitly profane words, however, the presence of these words does not always mean that the text instance is hateful. In some cases, text instances with profane words are just offensive language and they do not target any specific group, and so cannot be classified as hate speech. In this work, we build on existing studies to find a better demarcation between hate speech and offensive language. Our main contribution is to introduce the use of typed dependency as new features in our feature set. This new feature enables us to consider the relationship between long distance words in a text instance, thereby provides more identifying information than single word-based features. We evaluate our approach using a dataset with the classes: hate, offensive and neither. Comparing our work with existing studies, our feature set is much smaller but we achieve better accuracy and show comparable results in further analysis. Our detailed analysis also showed instances missed by the lexical features that were correctly predicted by the proposed feature set.