MSR-RAM: Macro-Structure Representation for Robust Argument Mining

ACL ARR 2024 August Submission156 Authors

Abstract: Argument Mining (AM) involves detecting Argument Relations (ARs) between Argumentative Discourse Units (ADUs) to uncover the argument structure. AR detection techniques often rely on micro-structural features derived from the internal structure of ADUs. However, argument structure is guided by a macro-structure representing the functional interdependence among ADUs of the argument. This macro-structure comprises segments, each segment containing ADUs serving specific functions to maintain coherence within that segment ($\textbf{local coherence}$) and cross-segment coherence ($\textbf{global coherence}$). This paper proposes an approach capturing such macro-structure encoding both local and global coherence for detecting AR. Experimental results on heterogeneous datasets showcase a notable performance enhancement, outperforming state-of-the-art models for both in-dataset and cross-dataset evaluation scenarios. The cross-dataset evaluation result underscores that the macro-structure boosts AR prediction skill transferable to new dataset.

Paper Type: Long

Research Area: Sentiment Analysis, Stylistic Analysis, and Argument Mining

Research Area Keywords: Macro-structure for argument mining, Coherence for argument mining, Robust argument mining

Contribution Types: NLP engineering experiment

Languages Studied: English

Submission Number: 156