Lexicon-based Detection of Violence on Social Media

Abstract

This study adopts a lexicon-based approach to address violence on social media. It uses FrameNet 1.7 (fn) and WordNet 3.1 (wn) to build a hierarchical domain-specific language resource of violence. The proposed lexicon tethers fn’s innovative integration of linguistic and paralinguistic knowledge to wn’s hierarchically-organized database. This tether alleviates the need to gather all paralinguistic violence-associated scenes and organize their linguistic realizations hierarchically. The proposed methodology can be internationally applied, given the multilingual availability of fn and wn, to cognitively and quantitatively explore a concept or a phenomenon. The lexicon is applied, then, to a corpus representing posts and comments retrieved from Donald Trump’s Facebook public page. Results reveal that the proposed lexicon recalls 92.68 of the total violence-related words in the corpus with a 76.31 precision (F-score= 83.7). More important, relating wn to fn inspires the creation of new frames, suggests slight modifications to existing ones and advocates promising mapping between some frames and synsets.