Joint extraction of argument components and relations

Argumentation mining (AM) aims at automatically extracting argument structures in documents. AM consists of three sub-tasks: argument identification, argument component classification and relation prediction. Some recent end-to-end AM works produce promising results, but their performances on the latter two AM sub-tasks are not satisfactory. To tackle this problem, in this work, we propose a framework that solves these two sub-tasks at the same time. We approach the problem by selecting linguistic features between sentence pairs, and training supervised learning models to label the argument component types and the relations at the same time. Our experiments on the persuasive essay corpus show that our system can achieve competitive labelling accuracy compared with the state-of-the-art AM techniques.