Topic-dependent relation prediction in argument mining: ternary classification

Abstract

The automatic identification of interactions among arguments expressed in natural language is a complex task that is mandatory for applications that need automatic argumentative reasoning. In this work, we focus on the prediction of relations between arguments, aiming to determine how a piece of text interacts with a specific discussion topic. Specifically, we aim to identify whether the text attacks, supports, or remains neutral toward a given topic. We frame this problem as a ternary classification task, and evaluate several methods for addressing it, including methods based on the combination of binary classifiers, models following a multitask learning approach, and large language models using prompting strategies. All our experiments were performed in a cross-topic scenario using two corpora, one of which includes more than 300 discussion topics, which brings our approach closer to a real-world scenario.