Argument maps

Claude Opus 5.5, 3 Nov. 2026

NLP and argumentation

An argument map records what a text says, and how its claims support or oppose one another, as a graph. Quell’s maps combine two published frameworks. Inference Anchoring Theory (IAT) structures the whole text or dialogue. The Periodic Table of Arguments (PTA) gives each individual argument a type. They are stored as xAIF JSON with a few additions, and edited in the lab’s xAIF editor. Reviewed maps are intended both as debate content and as the gold standard for automated analysis.¹

Two frameworks

IAT keeps two layers: the locutions, the exact words and who said them, and the propositions, what was asserted, reconstructed in a canonical form. Speech acts such as Asserting and Arguing anchor each proposition and inference in the words that perform it. This is what would let a player’s highlight, a span of words, be traced to the statement it expresses. See Inference Anchoring Theory and xAIF for the node types and a worked diagram.

The PTA describes each inference as a premise supporting a conclusion through a usually unstated lever. Its type, written as a code such as α FF Cau, follows from what the two statements share (form), what kinds of statement they are (substance), and the lever. See Periodic Table of Arguments.

Derived, not chosen

In Quell’s maps, an inference’s form and substance are derived. They follow from the canonical statements it connects, never from a hand-picked label, and only the lever is chosen. The rewording that turns a locution into its canonical statement is recorded on the speech act between them. This makes reconstruction, the step that needs the most judgement, an explicit and reviewable step.

How a map is made

In the editor, a reviewer:

1. pastes the passage and adds speakers;
2. marks locutions as spans of the text;
3. rewrites each proposition into canonical form, recording the transformations used;
4. analyses each statement: subject, predicate, fact, value or policy, and scope;
5. links premises to conclusions and anchors each relation in the dialogue.

The editor then derives each inference’s type and warns about structures IAT does not allow. A draft can come from hand annotation, from an agent, or from an automated approach. A person reviews it either way. Interpretive decisions are logged in judgement-calls.md, which is becoming the labelling guide.

Current state

• Gold maps. Two maps are reviewed, of passages from Aquinas and Engels. A third, of Aristotle’s function argument, is drafted and awaiting review.
• Conformance examples. ra-inputs.json covers every combination of statements that can feed an inference, each labelled with the code it should derive. The editor’s tests check the labels. Any reimplementation of the rules, such as a Rust core for the game, is meant to be checked against the same file.
• Corpus. The pilot corpus has 31 excerpts from 29 writers. It is held locally and never committed.

See also

• Debate — how the maps would be used in play.
• Reading notes — the separate question of what a player’s note is about.
• Research labs — the labs that hold the editor and corpus.

External links

• xAIF — ARG-tech’s JSON format for argument graphs.
• Argument Type Identification Procedure — the PTA’s published classification procedure.

What links here: Debate, Argument mining, Inference Anchoring Theory and xAIF, In-context learning, Periodic Table of Arguments, Stage-wise evaluation, oAMF, Reading notes, Glossary, Research labs, xAIF editor