Neural Clustering on Tree Structured Data: A case study on Argument Mining

This thesis explores the application of Graph Neural Networks based models to investigate the effectiveness in capturing and utilizing tree substructures, named fragments, to improve sentence classification in argument mining. The analysis focuses on two sub-tasks: Argumentative sentence detection, which involves identifying sentences that contain arguments, and Argument component detection, which consists of identifying and classifying the different components within an argumentative text. The dissertation presents a comprehensive study of three architecture variations for both sub-tasks: a classification baseline, a metric learning approach and a prototype network approach, evaluated on two separate datasets. Experimental results reveal that the proposed models achieve satisfactory performance in terms of F1 score: a mean score of 88.35 for USElecDeb60To16 and 63.19 for Persuasive Essays Corpus. However, an analysis of the models’ embeddings sparsity highlights that the performance of few models might not be entirely satisfactory and they require further refinement.