The joint system submission of HPI and DEDIS (Julian Risch, Anke Stoll, Marc Ziegele, Ralf Krestel) won at the Shared Task of the Identification of Implicit and Explicit Offensive Language at the GermEval workshop 2019. The workshop was held on the 8th of Ocotber at the Conference of Natural Language Processing (KONVENS) in Erlangen.
The organizers describe the task: "Implicit offensive language is a form of offensive language where the expression of hate, condemnation, inferiority etc. as directed toward an explicitly or implicitly given target has to be inferred from the ascription of (hypothesised) target properties that are insulting, degrading, offending, humiliating etc. Rather than explicitly expressing their aversion, the writers hint at something degrading, i.e. their tweets imply that the target is unworthy etc.".
Our winning system is an ensemble of multiple fine-tuned BERT models that were pre-trained on a large corpus of German-language documents. Our paper can be found here and the code is published on GitHub. An overview paper by the task organizers that summarizes all submissions can be found here. The task results are published here and a brief summary is in the table below.