Argument mining using graph neural networks

Summary

Automatically extracting arguments (Argumentation Mining) from texts has always been elusive topic. Research has been limited by the availability of large annotated corpora and efforts have focused almost exclusively on highly engineered feature oriented Machine Learning models. Recent advancements in Natural Language Processing and curation of a large, annotated English corpora for AM have sparkled new interest in this area. In this thesis we look into possible alternatives to highly engineered ML approaches. The thesis focuses particularly on the two more difficult tasks of identifying the type of argumentation unit (ADU) and the link between them. Specifically, we will be re-purposing a Graph neural network, used successfully for generating novel molecular structures [3], to the task of Argumentation Mining. The idea is to learn how to model the graph structure of arguments composed of Claims and Premises. Contrary to previous approaches, the model tries to reconcile the two tasks by learning simultaneously how to perform best on both tasks. The vanilla model uses only embeddings while an augmented model uses positions and cue words. The results showcase that the vanilla model is worse for the task of identifying the type of the argumentative unit when compared to the highly engineered ML model and a hard to beat baseline that exploits specific intricacies of the corpus. In contrast, when augmented with text positions the model performs on a par with the ML model and even beats it in specific metrics. On the other hand, for the task of ADU linking, all versions of the proposed model fail to beat the ML model. Finally, we showcase an interesting method to align translated text. The method relies on the use of multilingual, aligned embeddings and minimization of the L2 norm.