Title: Explanatory Argument Similarity Measurements with Abstract Meaning Representation and Generated Conclusions
Speaker: Juri Opitz (ICL)
Abstract
In this talk, we will discuss some recent insights into measuring argument similarity. In contrast to most other works, we
are interested in making the similiarity computation more transparent and explainable. In particular, we investigate two
hypotheses: First, that representing arguments with abstract meaning representation (AMR) and measuring distance between
arguments with AMR metrics improves argument similarity assessment. And second, that similar arguments may lead to
similar conclusions. We examine these hypotheses by viewing argument similarity through the lens of conclusions that we
generate using a pre-trained language model. Our experiments strongly confirm hypotheses one and provide weak evidence
for hypothesis two. We also note that the second hypothesis puts the spotlight on a pressing issue: what's a (good)
conclusion?
