CNRS IEA RHAPSSODY
ARgumentation tHeory And natural language ProceSSing fOr e-DemocracY
Joint project with Prof. Mihai Surdeanu from the University of Arizona.
The RHAPSSODY project proposes an automated reasoning framework that can extract, understand, and reason with complex arguments. The project is based on combining computational argumentation theory with natural language processing. The goal is to identify the most important arguments listed on debate platforms, estimate the acceptability degrees of these arguments using information mined from the web, and, using the totality of the arguments (those from the particular debate and those mined from the other web-sites), estimate the decision that will be taken. In addition to reasoning methods from computational argumentation theory, this effort also proposes novel deep learning methods for natural language processing (NLP) that mitigate overfitting to lexical artifacts. This is critical for automated argumentation. Our method combines data distillation (i.e., replacing names with types such as replacing “Hungary” with COUNTRY) and model distillation, in which multiple models trained on different versions of the distilled data compete with each other for the best accuracy.