High level argumentative errors in Argument Mining: Enhancing argument detection and automatic error prediction

Abstract

The automatic identification of arguments is an essential step for any system that analyzes the argumentative structure within a text document. Despite recent progress, due to the complexity of the task, methods developed to address it still exhibit errors in the identified arguments. In this work, we aim to improve the performance of models in charge of identifying arguments by directly addressing errors at the argumentation level. To achieve this, we propose several strategies that integrate high-level argumentative errors (that is, errors related to inaccuracies in the identified arguments) into the development of such models. We train Transformer-based models and evaluate the proposed strategies using both token-level and argumentation-level metrics across datasets from diverse domains, demonstrating the benefits of incorporating argumentative error information in the development of models. We also propose a method focused on determining whether an argument predicted by a model has been correctly identified, and, if not, the kind of argumentative error it presents, assuming the correct argument is not available. For that, we define a classification task and demonstrate its feasibility with satisfactory results using multiple classification models across several datasets, highlighting its potential to enhance various tasks within Argument Mining and improve real-world applications.