Towards an Artificial Argumentation System

Ramon Ruiz-DolzIJCAI 2020 · DOI

TL;DR

This paper proposes a comprehensive framework for building artificial argumentation systems that can analyze, evaluate, and even generate arguments. The authors aim to bridge the gap between current AI capabilities and the nuanced human art of argumentation, with a focus on applications like privacy management and educational tools.

What Problem Does This Paper Address?

Argumentative discourse, the art of persuasion and debate, has been studied for centuries. While Artificial Intelligence (AI) has made strides in areas like Natural Language Processing (NLP), automating the full spectrum of argumentation remains a significant challenge. Current AI systems struggle to:

  • Identify and analyze arguments: Automatically pinpointing premises, conclusions, and the relationships (support or attack) between them in text or speech is complex.
  • Evaluate arguments: Determining the strength or acceptability of arguments within a given framework is computationally intensive.
  • Generate arguments: Creating persuasive and contextually relevant arguments, especially tailored to specific audiences, is an open area.

The authors highlight that existing datasets for argument mining might be insufficient for advanced NLP models, and the cross-domain robustness of these models is largely unexplored. Furthermore, understanding how to make AI systems persuasive and explainable is crucial for their adoption in decision-making and other applications.

Methods

The paper outlines a research plan to tackle the entire argumentation process, which is generally divided into four tasks:

  • Argument Mining (AM): This involves identifying argumentative propositions and their relations. The authors plan to develop a generic pipeline using state-of-the-art NLP approaches, particularly transformer-based models, and evaluate their performance across different domains.
  • Argumentation Solving: This task focuses on evaluating arguments, typically represented using Argumentation Frameworks (AFs). The research will explore efficient algorithms and suitable semantics for determining argument acceptability within these frameworks.
  • Argument Invention (Generation): This involves creating new arguments. The authors are interested in computational persuasion and tailoring arguments to specific audiences, considering factors like personality traits and social network usage.

The paper also references prior work by the authors on an Argumentation Framework for Online Social Networks (AFOSN) and a method for automatically generating explanations to prevent privacy violations.

Datasets

Not discussed in the paper. The authors mention the potential limitation of existing Argument Mining corpora being "not big enough" but do not specify which datasets are used or if new ones are being created.

Key Insights and Ideas

  • Bridging AI and Argumentation: The core idea is to leverage recent AI advancements (like NLP) to build systems that can fully engage in argumentation, moving beyond simple analysis to persuasion and generation.
  • Holistic Approach: The paper advocates for a complete argumentation system that covers all four key tasks (identification, analysis, evaluation, invention), rather than focusing on isolated components.
  • Explainable AI through Argumentation: Computational argumentation is seen as a powerful tool for creating explainable AI systems, providing users with reasons and justifications for decisions.
  • Privacy Management as a Key Application: The authors have already explored using argumentation systems to assist users with privacy management in social networks, demonstrating a practical application of their research.
  • Audience-Centric Persuasion: A significant insight is the need to tailor argumentative strategies to specific audiences to maximize persuasive impact, acknowledging the complexity of human communication.

Key Results

Not discussed in the paper. The paper outlines a research plan and preliminary work but does not present final quantitative results.

Limitations and Future Work

Limitations:

  • Argument Mining Corpus Size: Existing corpora for argument mining may not be large enough to fully leverage advanced NLP models.
  • Cross-Domain Robustness: The behavior of argument mining models across different domains is not well-explored, and domain robustness is crucial.
  • Audience Variation: The significant variations between different audiences and the difficulties in gathering persuasion-related data pose challenges for computational persuasion.

Future Work:

  • Generic Argument Mining Pipeline: Developing a robust pipeline for extracting argument graphs from text and speech.
  • Efficient Argumentation Solving: Exploring efficient algorithms and suitable semantics for evaluating arguments in various domains.
  • Computational Persuasion: Investigating the effect of user characteristics (e.g., personality traits) on argument persuasiveness and developing refined persuasive strategies.
  • Application Domains: Continuing research in privacy management, journalistic text analysis (detecting fallacies), and educational debate systems.

Summary updated

BibTeX
@inproceedings{ruizdolz2020towards,
  author = {Ramon Ruiz-Dolz},
  title = {{Towards an Artificial Argumentation System}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.24963/ijcai.2020/743},
}
Figure