Extracting Arguments, Not Just Classifying Them: Instruction-Tuned LLMs for Generative Component Detection

Argumentative component detection (ACD) is a fundamental task in Argumentation Mining that involves identifying and labeling parts of a text as specific argumentative elements, such as claims or premises. Traditionally, this has been treated as a classification or sequence-labeling problem, often requiring complex pipelines to first segment the text and then categorize the segments. This paper introduces ITFACD, a new approach that shifts the focus from classification to language generation, using instruction-tuned Large Language Models (LLMs) to identify arguments directly from raw text.

A Generative Approach to Argumentation

The authors argue that existing methods for ACD are often limited by their reliance on pre-segmented text or rigid classification frameworks. By reframing ACD as a language generation task, the ITFACD model can process plain text directly. This allows the model to perform the dual task of delimiting argumentative spans and classifying them into components simultaneously, rather than relying on a multi-step pipeline. The same large language models question is explored in An Empirical Study of Harness Design..., which adds a research perspective.

Leveraging Instruction Tuning

The core of the ITFACD approach lies in the use of instruction-tuned LLMs. By utilizing compact, instruction-based prompts, the researchers guide the model to extract and categorize arguments in a single pass. This methodology demonstrates the effectiveness of instruction tuning for complex tasks in Argumentation Mining, moving away from the simplified sequence-labeling techniques that have dominated the field. The same large language models question is explored in LLM-Generated Feature Pools for Time Series..., which adds a research perspective.

Performance and Impact

Experiments conducted on standard benchmarks indicate that the ITFACD approach outperforms existing state-of-the-art systems. By successfully modeling ACD as a generative task, the authors provide a new, more efficient way to handle the complexities of identifying argumentative structures. This work represents one of the first attempts to fully adopt a generative framework for this specific subtask, highlighting the potential for LLMs to improve performance in Argumentation Mining. The same large language models question is explored in FlashVector, which adds a research perspective. as detailed in the full paper on Arxiv