Parsing Argumentation Structures in Persuasive Essays

Abstract

In this article, we present the first end-to-end approach for parsing argumentation structures in persuasive essays. We model the argumentation structure as a tree including several types of argument components connected with argumentative support and attack relations. We consider the identification of argumentation structures in several consecutive steps. First, we segment a persuasive essay in order to identify relevant argument components. Second, we jointly model the classification of argument components and the identification of argumentative relations using Integer Linear Programming. Third, we recognize the stance of each argument component in order to discriminate between argumentative support and attack relations. By evaluating the joint model using two corpora, we show that our approach not only considerably improves the identification of argument component types and argumentative relations but also significantly outperforms a challenging heuristic baseline. In addition, we introduce a novel corpus including 402 persuasive essays annotated with argumentation structures and show that our new annotation guideline successfully guides annotators to substantial agreement.