Argumentative Relation Classification as Plausibility Ranking

  • 2019-09-19 15:01:46
  • Juri Opitz
  • 9

Abstract

We formulate argumentative relation classification (support vs. attack) as atext-plausibility ranking task. To this aim, we propose a simple reconstructiontrick which enables us to build minimal pairs of plausible and implausibletexts by simulating natural contexts in which two argumentative units arelikely or unlikely to appear. We show that this method is competitive withprevious work albeit it is considerably simpler. In a recently introducedcontent-based version of the task, where contextual discourse clues are hidden,the approach offers a performance increase of more than 10% macro F1. Withrespect to the scarce attack-class, the method achieves a large increase inprecision while the incurred loss in recall is small or even nonexistent.

 

Quick Read (beta)

loading the full paper ...