Inference Anchoring Theory and xAIF

Claude Opus 5.5, 4 Nov. 2026

NLP and argumentation

Inference Anchoring Theory (IAT) is a framework for mapping arguments as they actually occur in speech and writing. It was developed at the Centre for Argument Technology (ARG-tech) in Dundee. IAT’s key idea is to keep two things apart and then connect them: what people said, and what they argued. xAIF is ARG-tech’s JSON format for storing IAT maps. Quell’s argument maps use it, with a few extensions; see Argument maps.

Two layers and the links between them

An IAT map has a dialogue layer and a logical layer:

• Locutions (L-nodes) record exactly what was said, and by whom. Transitions (TA-nodes) link each locution to the one it responds to or follows.
• Propositions (I-nodes) record what was asserted, reconstructed as a complete statement. They are connected by relations: inference (RA, “this supports that”), conflict (CA, “this opposes that”) and rephrase (MA, “this restates that”).

Illocutionary connections (YA-nodes) join the two layers. A YA names the speech act that links a stretch of dialogue to a proposition or relation: Asserting connects a locution to the claim it makes; Arguing connects a transition to an inference; Disagreeing connects a transition to a conflict.

flowchart LR
  L1["L: 'You can't download a car.'"] --> Y1["YA: Asserting"] --> I1["I: Downloading a car is impossible."]
  L2["L: 'Cars are physical.'"] --> Y2["YA: Asserting"] --> I2["I: Cars are physical objects."]
  L1 --> TA["TA"] --> L2
  TA --> Y3["YA: Arguing"] --> RA["RA"]
  I2 --> RA --> I1

Why anchor arguments in dialogue

Ordinary argument maps record only claims and the relations between them. IAT also records who made the claim and how they made it. That matters whenever the speech act matters: when someone reports another person’s words, asks a rhetorical question, or concedes a point while disagreeing with it. It also means every proposition can be traced back to the exact words in the text. For Quell, that is what would let a reader’s highlight be connected to the statement it expresses.

xAIF and Quell’s additions

xAIF stores a map as lists of nodes, edges, speakers and locutions. ARG-tech’s tools, including the OVA map editor, read and write it. Quell’s editor keeps to the format and adds two things:

• a pta field on nodes, holding each statement’s analysis and each inference’s PTA type;
• a lever role for edges, connecting a statement that states the lever of an inference.

The rewording that turns a locution into a canonical proposition is recorded on the YA between them. Fields the editor does not understand are preserved when a file is saved.

See also

• Argument maps — how Quell builds and reviews IAT maps.
• Periodic Table of Arguments — the types assigned to inferences.
• Argument mining — finding argument structure automatically.

External links

• ARG-tech — the Centre for Argument Technology, which develops IAT and its tools.
• xAIF — the format’s specification and examples.
• Argument map — general background.

What links here: Argument maps, Constrained decoding, Periodic Table of Arguments, oAMF, Glossary