ABDA-NL: A Natural-Language Scenario Explorer for Argument-Based Reasoning

1. Introduction

• ABDA-NL extends ABDA with a natural-language interface so users can inspect, explore, and extend scenarios without engaging in ASPIC- notation.ABDA supports a grounded discussion game over ASPIC- knowledge bases, while ABDA-NL adds natural-language interaction.Source: §1 · p. 1
• Informal argumentation exchanges claims, objections, and defenses, whereas formal argumentation often computes accepted claims as a static property of argument graphs.The contrast motivates an interface that connects formal outcomes with dialog.Source: §1 · p. 1
• ABDA’s reasoning over the argumentation framework remains authoritative, so conclusions are formally grounded, inspectable, and contestable.Figure 1 illustrates the tool on Popov v. Hayashi.Source: §1 · p. 1, Figure 1 · p. 2

2. Background: ASPIC- and the Grounded Discussion Game

• ASPIC- knowledge bases combine exception-free strict rules with defeasible rules that normally hold but may have exceptions; arguments arise by chaining rules.Attacks include rebutting a conclusion or sub-argument and undercutting a defeasible rule; attacks on premises are treated as rebuttals of sub-arguments.Source: §2 · p. 2
• Rule and assumption strength levels govern rebut success under weakest-link democratic ordering, while undercuts always succeed; grounded semantics labels arguments in, out, or undec.A literal is accepted when concluded by an in argument, undecided when concluded by an undec argument and no in argument, and rejected otherwise.Source: §2 · p. 2
• The Grounded Discussion Game is sound and complete for grounded semantics: a proponent has a winning strategy for argument A exactly when A is in the grounded extension.The proponent asserts that an argument has to be the case; the opponent challenges with an attacking argument that can be the case.Source: §2 · p. 3
• ABDA lets users play the game through formal move primitives, ending when no move is possible; the proponent wins if the main argument was conceded.The opponent concedes once every attacker has been retracted and retracts a challenge once one of its attackers has been conceded.Source: §2 · p. 3

3. ABDA-NL: Scenarios, Architecture, and Interface

• A scenario combines an ASPIC- knowledge base, a glossary linking identifiers to plain-language descriptions, and a corpus of reference documents.Its knowledge base declares facts, assumptions, intermediate propositions, key conclusions, and rules; assumptions and defeasible rules have strength blocks and active flags.Source: §3 · p. 3
• Commonsense and legal principles enter scenarios as defeasible assumptions, so their validity can itself be contested.The illustrated rules include rebutting conclusions and an undercut that targets every argument using rule wt1.Source: §3 · p. 3
• The shipped scenarios include Popov v. Hayashi, with 19 facts, 20 assumptions, 38 rules, and 8 key conclusions, plus five smaller scenarios.The smaller scenarios concern everyday life, medicine, environmental management, and sports; the reference corpus provides context for the language model but does not enter reasoning.Source: §3 · pp. 3–4
• The deterministic core rebuilds arguments and attacks after each change, computes grounded labels, and supplies the framework used by natural-language renderings.For built-in scenarios, recomputation takes milliseconds; limits on constructed arguments and attacks guard against pathological knowledge bases stalling the server.Source: §3 · p. 4
• The main view presents conclusion labels, facts and assumptions, rules, and chat, with identifiers connecting natural-language descriptions to formal items.A key conclusion that cannot currently be derived either way is labeled absent.Source: §3 · p. 4

4. Using the Demo

• The explanation view expands a selected conclusion into its argument, premises, sub-arguments, attacks, and proponent replies.Branches indicate whether challenges are retracted, stand, or are conceded; only counter-arguments that matter under grounded semantics are shown.Source: §4 · p. 4
• Figure 2 presents interactive explanation through the discussion game alongside natural-language rule authoring.Source: Figure 2 · p. 5
• Suspending an assumption or defeasible rule previews which conclusion labels would change before the user applies or cancels the what-if scenario.Users can also change preferences between conflicting defeasible rules, with changed labels highlighted and Reset restoring the baseline.Source: §4 · p. 5
• The chat answers questions using the case corpus and current scenario state, citing the documents it used.Users can add a conclusion, fact, assumption, or rule to the question being composed.Source: §4 · p. 5
• Plain-English authoring produces a proposed formalization with review notes, which users may refine, cancel, or apply before labels are recomputed.The proposal is displayed both as a sentence and in ASPIC- syntax.Source: §4 · p. 5
• The conclusion graph aggregates the argumentation framework by conclusion, distinguishing labels and mutual rebut, one-way rebut, and undercut edges.The ASPIC- listing includes glossary comments and suspension markers and can be re-imported.Source: §4 · p. 5

• ABDA computes arguments, attacks, and grounded labels deterministically after each change; the language model handles document-and-state questions and translates plain-English edits into candidate formal statements.Source: §5 · pp. 5–6
• For questions, the system combines lexically ranked excerpts from every document with the engine’s current labels, supporting rules, arguments, and scenario changes.Source: §5 · p. 6
• Before displaying answers, the server checks citations against the corpus and quotations against supplied passages, retrying one failed draft before showing a notice.Source: §5 · p. 6
• Plain-English authoring produces a structured proposal that undergoes deterministic validation and advisory review, then the user can refine, cancel, or apply it; nothing is applied automatically.Source: §5 · p. 6
• Different runs may propose different formalizations, but applied proposals are well-formed, use declared symbols, are shown to the user, and have consequences computed by the engine.Source: §5 · p. 6
• Figure 3 pairs an ABDA-computed label preview during scenario creation with a source reader displaying a quotation verified against the saved document.Source: Figure 3 · p. 7
• A model is admitted only after every question-answering and authoring test passes in every repetition, including checks of quotations, acceptance statuses, and proposal operations.Source: §5 · p. 7

• ABDA-NL is available as a public service for exploring built-in scenarios without installation or sign-in, and its keyboard-operable interface passes automated WCAG A and AA accessibility checks.Source: §6 · p. 7
• Users can build scenarios in a Guided plain-English editor or ASPIC- syntax view, attach reference documents for chat citations, validate and preview drafts without a model call, and import or export scenarios.Source: §6 · p. 7
• Signed-in users can save private scenarios, share revocable read-only links, and suggest scenarios for community listing after author consent and curator review.Source: §6 · p. 8
• The chat supports editable, branchable conversations tied to the scenario state of each question, while evidence cards open cited passages and Derivation links arguments to their ASPIC- counterparts.Source: §6 · p. 8
• An MCP server lets AI coding agents inspect scenarios and grounded outcomes, make version-checked edits that trigger recomputation, and request questions or proposals; agents must review and explicitly apply proposals.Source: §6 · p. 8

• The source code is available on GitHub, and the hosted service is at demo.abda-nl.org.Source: §7 · p. 8
• ABDA-NL was demonstrated at COMMA 2026.Source: §7 · p. 8
• The work acknowledges support from a Cardiff–DPI seed grant and use of cloud-computing resources through CloudBank and ACCESS allocations.Source: §7 · p. 8