A Stacking Approach for Cross-Domain Argument Identification.

Argument identification is the cornerstone of a complete argument mining pipeline. Furthermore, it is the essential key for a wide spectrum of applications such as decision making, assisted writing, and legal counselling. Nevertheless, most existing argument mining approaches are limited to a single, specific domain. The problem of building a robust system whose models are able to generalize over heterogeneous datasets remains fairly unexplored. In this paper, we tackle the argument identification task on two different datasets (Student Essays and Web Discourse), following two approaches: a classical machine learning approach and a DistilBert-based approach. Moreover, this paper sheds light on a new direction for researchers in this domain since we validate the principle of ensemble learning. In other words, we show that combining multiple approaches via a well stacked model improves the system performance. The results are very promising with respect to the recent findings in the literature.