Abstract
As data-driven AI models achieve unprecedented feats across previously unthinkable tasks, the diminishing levels of interpretability of their increasingly complex architectures can often be sidelined in place of performance. If we are to comprehend and trust these AI models as they advance, it is clear that symbolic methods, given their unparalleled strengths in knowledge representation and reasoning, can play an important role in explaining AI models. In this paper, I discuss some of the ways in which one branch of such methods, computational argumentation, given its human-like nature, can be used to tackle this problem. I first outline a general paradigm for this area of explainable AI, before detailing a prominent methodology therein which we have pioneered. I then illustrate how this approach has been put into practice with diverse AI models and types of explanations, before looking ahead to challenges, future work and the outlook in this field.
The Black Box Problem: Machine vs. Human Centricity
As AI models, particularly deep neural networks, grow in complexity, they increasingly function as "black boxes." While their predictive performance is high, the internal logic driving those predictions remains opaque. Current Explainable AI (XAI) research often prioritizes machine-centric metrics like faithfulness—the degree to which an explanation accurately reflects the model's internal mathematical state.
However, this focus often neglects the human element. For an explanation to be effective, it must be understandable and trustworthy to a human user. Most people do not process information through gradient weights or feature heatmaps; instead, human cognition is naturally argumentative. We reason by weighing evidence for and against a conclusion. This paper advocates for bridging the gap between mathematical machine logic and human-centric reasoning through computational argumentation.
A Modular Paradigm for XAI
The proposed framework moves from raw AI model data to human-digestible explanations through an intermediate argumentative representation. This process is modular, allowing for different types of AI models to be mapped to various explanation formats.
By translating the model's internal state into a graph of "attacking" and "supporting" arguments, we create a structure that mirrors human dialogue. Once this representation is established, it can be delivered to users in multiple formats, including natural language text, tabular summaries, or interactive conversational agents.
The "Forging" Methodology
The core of this approach is a process called "forging," which systematically extracts argumentative structures from diverse AI architectures. This involves three critical stages:
- Identification (Step 1a): Pinpointing specific components within the AI model (e.g., neurons in a CNN or features in a recommender system) that can serve as potential arguments and identifying the potential relations between them.
- Explanation Mould (Step 1b): Defining the logical constraints that govern these components. This involves "inverting" gradual semantics—taking the continuous mathematical behavior of the model and defining what constitutes a support or attack relation.
- Generation (Step 1c): Constructing the final argumentative graph (typically a Quantitative Bipolar Argumentation Framework, or QBAF) by verifying which potential relations satisfy the defined mould.
Why Argumentation Works
Argumentation is uniquely suited for XAI for several reasons:
- Natural Reasoning: Drawing on cognitive science, the framework aligns with how humans naturally process information and justify decisions.
- Gradual Semantics: Unlike binary logic, "gradual" argumentation allows for varying strengths of support or attack. This maps effectively to the continuous activation values found in neural networks, preserving the nuance of the original model.
- Interactivity: Argumentative frameworks support dialogue. A user can challenge a specific argument (e.g., "I don't like this specific feature"), and the model can update its reasoning in real-time, making the AI a collaborative partner rather than a static oracle.
Empirical Success and Future Outlook
This methodology has been successfully applied to a wide range of architectures, including Multi-Layer Perceptrons (MLPs), Convolutional Neural Networks (CNNs), and tree-based models like Random Forests. In user studies involving movie recommendations, argumentative explanations significantly improved user trust and comprehension compared to standard baselines. Furthermore, in classification tasks, these "forged" representations have shown empirical advantages over popular XAI methods like SHAP in terms of both faithfulness and utility.
Looking forward, several challenges remain. While effective for CNNs, establishing argumentative abstractions for Transformer architectures (the backbone of LLMs) is a key area of ongoing research. Additionally, as legal frameworks like the GDPR introduce the "right to contest" automated decisions, argumentative XAI provides a natural mechanism for building contestable and legally compliant AI systems. The ultimate goal is to move beyond simple transparency toward AI that can engage in meaningful, human-like reasoning.
Summary updated