Source-linked AI summary
Argumentative XAI: A Survey
Kristijonas Čyras, Antonio Rago, Emanuele Albini, Pietro Baroni, Francesca Toni
TL;DR
Argumentative XAI lacks a consolidated account of how computational argumentation can support explanations across AI settings. This survey organizes AF-based explanations by the models they explain, the frameworks and delivery forms they use, and their intrinsic or post-hoc deployment, then identifies priorities for future research. It concludes that the area is varied but still needs stronger explanation properties, computational analysis, and broader applications, especially in machine-learning settings.
Problem
The literature on argumentative XAI spans varied models, argumentation frameworks, and explanation forms without a comprehensive survey focused on computational argumentation.
Method
The paper surveys AF-based explanations across intrinsic and post-hoc settings, model categories, argumentation frameworks, delivery forms, and research directions.
Results
AF-based explanations vary substantially in the models they explain, the frameworks they deploy, and the forms they take, including dispute-tree and AF-modification explanations.
Takeaways & Limitations
The survey identifies explanation properties, computational aspects, and broader applications as central directions for developing argumentative XAI.
Takeaways & Limitations
AF-based explanations have been only sparingly deployed in machine-learning classification and probabilistic settings.
Abstract
from arXiv · showhide
Explainable AI (XAI) has been investigated for decades and, together with AI itself, has witnessed unprecedented growth in recent years. Among various approaches to XAI, argumentative models have been advocated in both the AI and social science literature, as their dialectical nature appears to match some basic desirable features of the explanation activity. In this survey we overview XAI approaches built using methods from the field of computational argumentation, leveraging its wide array of reasoning abstractions and explanation delivery methods. We overview the literature focusing on different types of explanation (intrinsic and post-hoc), different models with which argumentation-based explanations are deployed, different forms of delivery, and different argumentation frameworks they use. We also lay out a roadmap for future work.
1 Introduction
Argumentative XAI is surveyed as a computational-argumentation approach to explaining AI systems, motivated by argumentation’s dialectical character and broad applicability. The survey catalogues explanation approaches, delivery forms, and future research directions while excluding informal models without argumentation-framework semantics.
- Motivation: Argumentative explanations are motivated by their use in social-science accounts of explanation and their potential advantages for human-facing XAI.The survey connects this motivation to XAI’s role in bridging AI applications with their developers and users.
- Scope: The survey focuses on overtly argumentative XAI approaches built with computational argumentation, rather than treating every explanation with positive and negative relations as argumentative.Attribution and contrastive methods can appear argumentative, but the survey concentrates on approaches using explicit computational argumentation.
- Contributions: The survey catalogues AF-based explanations by what they explain, which argumentation frameworks they use, and how the resulting explanations are delivered.Its contributions include reviewing the literature, classifying explanation forms, and outlining a roadmap for future work.
- Scope boundary: The survey excludes argumentative explanations based on informal notions or models lacking argumentation-framework aspects, especially semantics.Application domains of argumentative XAI are treated as outside this survey’s scope.
2 Argumentation Frameworks
The surveyed argumentation frameworks represent dialectical reasoning through abstract or structured arguments, relations such as attack and support, and semantics that determine acceptable arguments or extensions. They range from general frameworks to specialized instances whose arguments encode cases, deductions, or rule-based reasoning.
- Abstract arguments: Abstract Argumentation represents arguments as entities connected by attack, with semantics mapping frameworks to extensions satisfying dialectical constraints such as conflict-freeness.An extension is a set of arguments, and conflict-freeness excludes internally attacking pairs.
- Abstract arguments: Bipolar Argumentation adds support to attack, while Gradual Argumentation can use multiple dialectical relations and Abstract Dialectical Frameworks use user-given acceptance conditions.These frameworks broaden the relations or semantic mechanisms available for representing argument interaction.
- Structured arguments: Structured frameworks represent arguments as deductions from assumptions or strict and defeasible rules, using extension-based semantics or dialectical trees.ABA, ASPIC+, and DeLP are the principal examples discussed for explanation.
- Specialized AA instances: Specialized AA instances use cases, abductive deductions, logical deductions, or argument schemes as structured arguments.The survey names these variants AA-CBR, AA-ALP, AA-LD, and AA-AS.
3 Types of Argumentative Explanations
The survey distinguishes intrinsic explanations, built into argumentative models, from complete and approximate post-hoc explanations extracted from non-argumentative models. It organizes representative applications by model category and argumentation framework.
- Categories: AF-based explanations are classified as intrinsic, complete post-hoc, or approximate post-hoc, although the boundaries between these categories are not crisp.The survey also covers recommender systems, classifiers, probabilistic methods, decision-making and knowledge-based systems, planners and schedulers, and logic-programming tools.
- Intrinsic approaches: Intrinsic explanations use argumentation natively, including DeLP-based recommender systems and AA-driven classification and decision-making methods.Examples include explanations for contradictory movie recommendations, legislation outcomes, binary classifications, and decisions evaluated through arguments for and against alternatives.
- Complete post-hoc approaches: Complete post-hoc approaches map non-argumentative models into AFs while preserving their relevant structure, supporting explanations for PageRank, decision problems, knowledge bases, scheduling, planning, and logic programs.Examples include QBA representations of web graphs, ABA mappings of decision models, AA explanations for query answering and scheduling, and exact correspondences for answer sets.
- Approximate post-hoc approaches: Approximate post-hoc approaches use incomplete mappings from models to AFs, including recommender systems and Bayesian networks.TA or BA frameworks approximate rating propagation in recommender systems, while SA frameworks represent variable interplay through Bayesian-network Markov blankets.
4 Forms of AF-Based Explanations
AF-based explanations can expose the substructures, dialogues, extensions, or changes that determine an argument’s acceptability. The survey also emphasizes converting these formal objects into user-presentable natural-language, visual, or conversational explanations.
- Sub-graphs and proofs: A common explanation form is a sub-graph containing the topic argument and relevant dialectical relations, often constrained by a formal property.Dispute trees are prominent examples and can function as proofs of argument acceptance.
- Sub-graphs and proofs: Dispute trees can provide guarantees of existence, correctness, and relevance for explanations of argument acceptability.Correctness means the explanation establishes the topic argument’s acceptance status, while relevance means its arguments contribute toward acceptance.
- Sub-graphs and proofs: Other sub-graph forms include paths, cycles, and branches, while structured arguments and dialectical relations can themselves serve as explanatory structure.These forms are used with BA, QBA, DeLP, and other argumentation frameworks.
- Dialogues and extensions: Dialogue-based explanations use structured, rule-guided exchanges of arguments to establish or explain acceptability.Dialogues may be constructed from dispute trees or from argumentation frameworks more generally.
- Dialogues and extensions: Explanations may also be related admissible extensions or AF modifications whose addition or removal changes a topic argument’s acceptability.The latter form identifies arguments or relations whose alteration would change the topic’s status.
- Delivery: Presenting formal explanations to users commonly requires natural-language generation, visualization, or conversational delivery.Conversational explanations often arise naturally from argument structures, sub-graphs, and dialogue games.
5 A Roadmap for Argumentative XAI
The roadmap identifies gaps in the properties and computation of AF-based explanations, while proposing broader applications and connections to machine learning and logic-based explanations.
- Properties of AF-based explanations: Future work should study properties of AF-based explanations, including fidelity, cognitive tractability, transparency, trust, and user acceptance.The survey notes that properties of AFs are well studied, whereas properties of their explanations are less developed.
- Computational aspects: Efficient AF-based explanations require tractable reasoning, model-to-AF extraction for post-hoc methods, and systematic analysis of explanation extraction complexity.Intrinsic approaches depend on reasoning tasks such as grounded-extension membership, while post-hoc approaches additionally require extracting AFs from models.
- Applications and scope: The survey presents AF-based explanations as having substantial potential for further development across applications and explanation settings.This broadening is framed as an opportunity alongside work on explanation properties and computational foundations.
- Applications and scope: AF-based explanations remain only sparingly deployed in machine-learning classification and probabilistic settings.The survey proposes interplay between machine-learning explanation needs and further research in computational argumentation.
- Applications and scope: Future research could examine whether existing logic-based and counterfactual explanations can be understood as AF-based explanations.The proposed connections include model-agnostic and model-specific logic-based explanations and existing logic-based AFs.
6 Conclusions
The survey synthesizes argumentative XAI built with computational argumentation, covering diverse models, AFs, and explanation forms. It distinguishes intrinsic explanations from post-hoc explanations and proposes a roadmap for future development.
- Survey scope: The survey covers AF-based explanations across the models they explain, the argumentation frameworks they deploy, and the forms those explanations take.It considers both symbolic and statistical or probabilistic models.
- Intrinsic and post-hoc explanations: AF-based explanations may be intrinsic components of argumentatively explainable systems or post-hoc explanations for existing systems.The survey describes this as combining symbolic AF representations with symbolic or statistical/probabilistic models.
- Future directions: The roadmap targets future development and use of AF-based explanations across computational argumentation and application-oriented XAI.The stated audience includes symbolic-AI experts and application experts seeking customisable XAI solutions.