This container employs a transformer based deep neural network to predict the argument stance of a given text chunck. In default mode it uses chkla/roberta-argument by Christopher Klamm. Our implementation also supports the refinement of the roberta-argument model with argumentative and non-argumentative statements taken from the summetix (formerly known as ArgumenText) search-engine. Usage of refinement is commenced by

Please be aware that model fine-tuning takes quite a while. It is highly recommended to pre-train the model before the debate takes place. The new fine-tuned model can then be replaced in the folder 00_custom/argument_models/chkla.

Users can also change the threshhold for a chunk being classified as argument in the settings, see ARGUMENTEXT_CUTOFF.

Technical Framework

This framework is written in Python.

Main Contributor

Kolja J. Holzapfel