Structure-Aware Data Augmentation for Argument Mining
Abstract
AbstractThis paper introduces a novel data augmentation approach designed for argument mining. Argument mining is a task that automatically identifies argumentative structures within texts, typically represented as trees or directed acyclic graphs (DAGs). A major challenge facing current argument mining systems is the limited structural diversity of existing annotated datasets, as manual annotation inevitably captures only a small subset of the argumentative patterns encountered in practice. To overcome this challenge, we propose an innovative data augmentation method. Our approach begins by generating pseudo-argumentative structures through the exploration of various combinations of argument labels. These structures then guide a large language model (LLM) to produce new texts with annotations directly incorporated into the text. We evaluated our method on the Argument Annotated Essay Corpus (AAEC) and the AbstRCT dataset. Experimental results showed that the proposed method outperformed existing augmentation baselines on Component-F1 and Relation-F1 in most settings. Further analysis indicated that the variety in the generated pseudo-argumentative structures plays a crucial role in achieving these improvements.