End-to-End Argument Mining through Autoregressive Argumentative Structure Prediction

TL;DR

The paper tackles end-to-end Argument Mining (AM) by introducing Autoregressive Argumentative Structure Prediction (AASP), which jointly solves ACI, ACC, ARI, and ARC via a constrained sequence of actions predicted autoregressively by a conditional language model. AASP builds three action families—Span-Identifying, Boundary-Pairing, and Type-Labeling—to construct, bound, and label argumentative structures, and uses larger FFNs to manage long spans and distantly related ACs. Empirical results across three diverse AM benchmarks show SoTA performance on two datasets and robust gains in relational tasks, along with detailed ablations and error analyses that highlight the importance of model size and relational modeling. The approach demonstrates strong potential for end-to-end AM, with practical implications for downstream reasoning, debate analysis, and educational assessment, by explicitly modelling intra-structure dependencies through a structured autoregressive framework.

Abstract

Argument Mining (AM) helps in automating the extraction of complex argumentative structures such as Argument Components (ACs) like Premise, Claim etc. and Argumentative Relations (ARs) like Support, Attack etc. in an argumentative text. Due to the inherent complexity of reasoning involved with this task, modelling dependencies between ACs and ARs is challenging. Most of the recent approaches formulate this task through a generative paradigm by flattening the argumentative structures. In contrast to that, this study jointly formulates the key tasks of AM in an end-to-end fashion using Autoregressive Argumentative Structure Prediction (AASP) framework. The proposed AASP framework is based on the autoregressive structure prediction framework that has given good performance for several NLP tasks. AASP framework models the argumentative structures as constrained pre-defined sets of actions with the help of a conditional pre-trained language model. These actions build the argumentative structures step-by-step in an autoregressive manner to capture the flow of argumentative reasoning in an efficient way. Extensive experiments conducted on three standard AM benchmarks demonstrate that AASP achieves state-of-theart (SoTA) results across all AM tasks in two benchmarks and delivers strong results in one benchmark.

Figures (7)

• Figure 1: Illustration of the decoding process in the AASP framework for end-to-end argument mining. The process involves three main actions step-by-step: (1) Action $a_i$ constructs the target sequence for span-identifying actions; (2) Action $b_i$ finalizes the AC span boundaries by matching the <m> and </m> tokens; and (3) Action $z_i$ connects related ACs by linking their $</m>$ tokens and classifies both the AC types and the types of their ARs.
• Figure 2: Model Architecture of AASP
• Figure 3: Comparison of ACI Micro-F1 scores (%) between AASP and ST across paragraphs of varying lengths (in terms of the number of ACs present in that paragraph) for the AAE, AAE-FG, and CDCP datasets.
• Figure 4: Comparison of ACC Micro-F1 scores (%) for AASP and ST across different AC categories in the AAE, AAE-FG, and CDCP datasets.
• Figure 5: Performance analysis of ARI task Micro-F1 scores (%) for AASP and ST across varying distances (in terms of number of intermediate ACs) in the AAE, AAE-FG, and CDCP datasets.