An Argument-Ontology for a Response-Centered Approach to Argumentation Mining

Mark Aakhus, Smaranda Muresan, Nina WacholderCMNA @ IJCAI 2016 · OpenReview

Abstract

Traditional Argumentation Mining (AM) often relies on static linguistic markers and rigid logical structures, which frequently lead to false positives and the omission of implicit reasoning. This paper, originally presented at the 16th Workshop on Computational Models of Natural Argument (CMNA 2016), proposes a shift toward a "response-centered" approach. By treating argumentation as a dynamic, social process of managing disagreement, the authors introduce a two-tier ontology that focuses on the interactional relationship between a "Target" statement and a "CallOut." This framework aims to capture real-world reasoning more accurately by prioritizing how participants respond to one another over simple keyword detection.

Introduction

Argumentation Mining (AM) is the computational task of automatically identifying and extracting argumentative structures from natural language text. Historically, this field has focused on identifying overt linguistic markers (e.g., "because," "therefore") or pre-defined logical syllogisms. However, these methods struggle with the nuances of human discourse.

The authors argue that real-world reasoning is often implicit and social. To address this, they propose a response-centered ontology that views argument not as a static set of premises and conclusions, but as a process of "disagreement management." This approach allows for the identification of arguments even when formal argumentative language is absent.

The Limitations of Traditional AM

The paper identifies two primary flaws in current AM methodologies:

  1. False Positives: Many words typically associated with argumentation, such as "but" or "well," frequently appear in non-argumentative contexts, such as simple clarifications or explanations.
  2. False Negatives: In natural conversation, participants often resolve differences of opinion using "traces" of reasoning. Because this reasoning is implicit, it is often missed by systems looking for formal logical structures.

The challenge lies in distinguishing between language that merely looks argumentative and language that actually functions as an argument within a specific social context.

A Two-Tiered Argument Ontology

To bridge the gap between high-level theory and practical NLP implementation, the authors propose a dual-layered ontology.

1. The Upper Ontology

The Upper Ontology defines the fundamental structural elements of an argumentative interaction:

  • Target: A specific part of a prior action or statement that is being addressed.
  • CallOut: A subsequent action that "selects" the Target and renders it problematic. The CallOut implies that there are viable alternatives to what the Target expressed or implied.
  • Argumentative Relation: The functional link connecting the CallOut to its Target.

2. The Lower Ontology

The Lower Ontology provides finer-grained dimensions to categorize the nature of the interaction:

  • States of Affairs (SoA): This focuses on propositional content—the "what" of the argument, similar to traditional premises and conclusions.
  • State of Felicity (SoF): This focuses on the "how" of communication. It addresses the appropriateness or legitimacy of an action (e.g., questioning if a speaker has the authority to make a specific request).
  • Modality: This categorizes the specific nature of the disagreement, such as reframing, challenging, or dismissing a claim.

Key Insights: Argument as Disagreement Management

The core of this framework is the idea that an argument is "called into being" when a participant makes a prior statement problematic. This "CallOut" acts as the catalyst for the argumentative process.

Furthermore, the authors emphasize the "Disagreement Space"—the network of overarching presumptions that govern a discussion. By focusing on the pragmatics of interaction (SoF) rather than just the truth of statements (SoA), the ontology can track how participants argue about the quality, clarity, or legitimacy of each other's contributions.

Conclusion and Future Work

As a conceptual and theoretical contribution, this paper provides a formal design for an extensible argument-ontology. While it does not provide quantitative benchmarks, it offers a principled roadmap for building more sophisticated NLP systems capable of discovering unconventional argumentative structures.

Future work remains in refining specific argumentative relations—such as the precise distinction between a "rebuttal" and a "dismissal"—and developing machine learning architectures capable of detecting these nuanced "CallOuts" in messy, unstructured multiparty discourse.

Summary updated

BibTeX
@inproceedings{aakhus2016argument,
  author = {Mark Aakhus and Smaranda Muresan and Nina Wacholder},
  title = {{An Argument-Ontology for a Response-Centered Approach to Argumentation Mining}},
  booktitle = {CMNA @ IJCAI},
  year = {2016},
  url = {https://openreview.net/forum?id=YvRvSrzJwFY},
}
Figure