[Paper Review] Deep Argumentative Explanations

Emanuele Albini, Piyawat Lertvittayakumjorn+2 more|arXiv (Cornell University)|Dec 10, 2020

Explainable Artificial Intelligence (XAI)20 references4 citations

TL;DR

This paper introduces Deep Argumentative Explanations (DAXs), a novel framework that uses computational argumentation to generate interpretable, hierarchical explanations from neural networks by modeling neurons and their interactions as arguments with dialectical relations (support/attack). DAXs achieve high fidelity to model internals, low computational cost, and demonstrate superior comprehensibility and alignment with human judgment compared to LIME and SHAP, especially in inspiring trust and insight into model reasoning.

ABSTRACT

Despite the recent, widespread focus on eXplainable AI (XAI), explanations computed by XAI methods tend to provide little insight into the functioning of Neural Networks (NNs). We propose a novel framework for obtaining (local) explanations from NNs while providing transparency about their inner workings, and show how to deploy it for various neural architectures and tasks. We refer to our novel explanations collectively as Deep Argumentative eXplanations (DAXs in short), given that they reflect the deep structure of the underlying NNs and that they are defined in terms of notions from computational argumentation, a form of symbolic AI offering useful reasoning abstractions for explanation. We evaluate DAXs empirically showing that they exhibit deep fidelity and low computational cost. We also conduct human experiments indicating that DAXs are comprehensible to humans and align with their judgement, while also being competitive, in terms of user acceptance, with some existing approaches to XAI that also have an argumentative spirit.

Motivation & Objective

• To address the lack of transparency in existing XAI methods, which focus narrowly on input-output influence without revealing internal reasoning structures.
• To bridge symbolic AI (computational argumentation) with deep learning by modeling neural network components as arguments with dialectical relations.
• To develop a framework that generates explanations reflecting the deep, layered reasoning of neural networks, not just flat feature importance.
• To evaluate DAXs empirically for fidelity, efficiency, and human comprehensibility, comparing them to established XAI methods like LIME and SHAP.
• To validate that DAXs align with human judgment and are perceived as more insightful and trustworthy than flat explanations.

Proposed method

• DAXs are generated by extracting a Generalised Argumentation Framework (GAF) from a trained neural network, where neurons or neuron groups are treated as arguments.
• Dialectical relations—support and attack—are defined between arguments based on activation patterns and relevance propagation (e.g., using LRP).
• Dialectical strength of arguments is derived from quantitative measures (e.g., LRP scores), ensuring alignment with model behavior and satisfying key dialectical properties.
• The framework supports multiple explanation formats, including conversational and structured forms, tailored to user cognitive needs via a parameter χ.
• Three DAX instances are implemented across different architectures (e.g., CNNs for text), using distinct GAFs and semantics derived from underlying model outputs.
• Explanations are evaluated via both automated metrics (deep-fidelity, complexity) and human studies assessing comprehensibility, alignment, and trust.

Experimental results

Research questions

• RQ1 Do DAXs reflect the deep internal structure of neural networks, ensuring high fidelity to model reasoning?
• RQ2 Do DAXs maintain low computational complexity while providing rich, hierarchical explanations?
• RQ3 Do supports between intermediate arguments and predicted classes in DAXs align with human judgment?
• RQ4 Do supports between individual words and intermediate arguments in DAXs align with human judgment?
• RQ5 Do DAXs inspire greater trust and provide better insight than LIME and SHAP in user studies?

Key findings

• DAXs exhibit deep-fidelity, meaning their internal argumentative structure accurately reflects the reasoning process of the underlying neural network.
• DAXs demonstrate low computational complexity, making them efficient for deployment across various neural architectures.
• In RQ4, 96.0% ± 0.4% of participants correctly selected the class supported by n-grams, confirming alignment with human judgment on intermediate argument support.
• In RQ5, 85.0% ± 0.7% of participants correctly identified the strongest supporting or attacking word, validating the semantic soundness of DAX support relations.
• DAXs significantly outperformed LIME and SHAP in providing insight into the CNN’s internal functioning (p < 0.004 and p < 0.006, respectively).
• DAXs inspired significantly greater trust than SHAP (p < 0.002), though the trust advantage over LIME was not statistically significant (p = 0.32), likely due to participant expertise levels.