Learning First-Order Logic Rules for Argumentation Mining

Argumentation Mining (AM) aims to extract argumentative structures from texts by identifying argumentation components (ACs) and their argumentative relations (ARs). While previous works focus on representation learning to encode ACs and AC pairs, they fail to explicitly model the underlying reasoning patterns of AM, resulting in limited interpretability. This paper proposes a novel First-Order Logic reasoning framework for AM (FOL-AM), designed to explicitly capture logical reasoning paths within argumentative texts. By interpreting multiple AM subtasks as a unified relation query task modeled using FOL rules, FOL-AM facilitates multi-hop relational reasoning and enhances interpretability. The framework supports two flexible implementations: a fine-tuned approach to leverage task-specific learning, and a promptbased method utilizing large language models to harness their generalization capabilities. Extensive experiments on two AM benchmarks demonstrate that FOL-AM outperforms strong baselines while significantly improving explainability. Kuribayashi et al. (2019b) enhances AC representations using discourse markers, while Bao et al. (2021a) employs a transition-based network for AC and AC pair modeling. Morio et al. (2022b) leverages biaffine functions to encode relational representations, and Sun et al. (2024) incorporates task interactions into AC and AC pair representations. However, these methods fail to explicitly model the reasoning patterns underlying AM, potentially leading to sub-optimal performance. Moreover, they lack interpretability, as they rely on black-box neural networks. For example, as shown in Figure 1 , the stateof-the-art (SOTA) model PITA misclassifies the relation between AC1 and AC5 as Attack due to their shared contextual cue increase taxes, whereas AC5 actually expresses a negative stance toward the topic. In contrast, humans can deduce the correct relation using logical reasoning, recognizing that "if AC1 promotes AC4, and AC4 entails AC5, then AC1 and AC5 are in a Support relation." In this paper, we propose a novel first-order logic reasoning framework for multi-task AM (FOL-AM). As indicated in Figure 1 , our method aligns with human reasoning by introducing interpretable FOL rules derived from the argumentative text for AM. For example, we may use the following logic rule to identify the relation between AC pair 𝐴𝐶1 and 𝐴𝐶5 as "Support": Support(AC1, AC5) ← Promotion 1 (AC1, AC4) ∧ Entailment(AC4, AC5). This logic reasoning framework has two clear benefits over previous approaches: First, it naturally captures multi-hop relations among ACs owing to the compositional structure of the reasoning chain. Second, it improves interpretability as the reasoning processes are visible. Specifically, we uniformly transform ACTC and ARC subtasks as relation query 𝑞(𝑥 1 , 𝑥 𝐿+1 ) associated with an AC pair (𝑥 1 , 𝑥 𝐿+1 ) within the argumentative text. The query is modeled as the FOL rule to conduct logical deduction: ). The body of the rule (on the left) is formed by attentively selecting proper ACs 𝑥 𝑖 and predicates 𝑟 𝑖 upon given argumentative text. We implement the FOL-AM framework in two settings: a fine-tuned method and a prompt-based method using LLMs. Intuitively, the former method can learn the nuanced patterns specific to the AM task directly from the dataset,