Vocabularies of hate speech terms with descriptions

The dataset includes three vocabularies (in Greek, English, and French) of terms expressing hate speech, i.e. terms that are considered derogatory and offensive towards a group and their members, based on their characteristics or beliefs. Such characteristics include: gender, sexual orientation, ethnicity, race/color, political affiliation, religion, socioeconomic status, age, disability, and physical appearance. The vocabularies are not meant to include common slurs, which may in certain context be directed against certain groups, but are not by themselves derogatory or offensive towards such a group. The vocabularies include the following information which accompanies each term: Source: where the term was sourced from, specifically a link to the respective Wikidata entry. Category: the group characteristics based on which the term is offensive or derogatory Description: Free text description of the possible meanings of the term. The description mentions why and under which circumstances the term is used in an offensive way. The vocabularies have been constructed in a semi-automatic way. The following steps have been followed for their construction: a) As a first step, we collect all Wiktionary terms that are tagged with labels denoting derogatorive, offensive and vulgar entries. b) The initial list of terms os filtered to keep only politically-charged toxic terms. For Greek, the filtering is performed manually by a human agent. For Greek and English, an LLM is instructed to assess which terms are politically charged and humans have validated a sample of the automatic outputs. c) Categories are specified by a human or by an LLM, with a sample evaluation by humans. d) The Wiktionary definitions corresponding to the filtered terms along with appropriate instructions are provided as input to an LLM, so as to produce a continuous text that describes the meaning(s) of the term, with an emphasis on why and under which circumstances the term is used in an offensive way. The LLM of choice was Claude Sonnet 3.7.