The Book That Was Supposed to End a Career

When Stephen Toulmin published The Uses of Argument in 1958, the reception from professional logicians was hostile. He had written a book about reasoning that declined to be about formal logic, and he was told, in effect, that he had wandered off the map. Toulmin later recounted that a colleague took to calling it his anti-logic book.

It did not end his career. What it did was find a completely different audience. Departments of rhetoric and communication — especially in the United States — picked the book up and built curricula on it, because Toulmin had done something formal logic had never bothered to do. He had described what an argument looks like when a real person makes one.

Sixty-odd years later, if you ask a machine to read a meeting transcript and tell you what was argued, it will use Toulmin's vocabulary to answer. That is the story of this series, and this post is the arc of it: how a philosophy book that logicians rejected became the specification for a machine-learning task.

What Toulmin Actually Changed

Formal logic is concerned with validity: given these premises, does this conclusion follow? It is a superb tool and it describes almost nothing people actually say. Nobody argues in syllogisms.

Toulmin's move was to ask a different question — not is this valid but how is this built. He broke a real argument into six parts, and three of them do the heavy lifting:

• Claim — the position being advanced. What you are actually asking someone to accept.
• Grounds — the facts offered in support. The evidence, the data, the example.
• Warrant — the principle that licenses the move from grounds to claim. Almost never stated aloud.

The warrant is the interesting one, and it is the reason argument mining is hard rather than tedious. The bridge is structurally unsound, so we should close it contains a hidden premise: that unsound structures should be closed. Nobody says that sentence. Everybody assumes it. A machine reading the text has to reconstruct a step that is not in the text.

The other three parts — backing, qualifier, rebuttal — handle the hedging and the exceptions. Together they describe an argument as a structure rather than a chain, and that structural view is what made the whole thing computable later.

Walton and the Question That Breaks the Argument

If Toulmin gave the field its anatomy, Douglas Walton gave it a field guide. Working through the 1990s and 2000s, Walton catalogued the recurring patterns of everyday reasoning — appeal to expert opinion, argument from analogy, argument from consequences, argument from sign — as argumentation schemes.

The catalogue is useful. What makes it powerful is the second half: each scheme comes with critical questions, the specific challenges that expose where that pattern fails. For appeal to expert opinion: is this person actually an expert in this field? Are other experts in agreement? Is the claim consistent with the evidence?

This is the single most useful thing in the whole tradition for anyone sitting in a meeting, and it costs nothing to apply. When someone argues from authority, you do not have to disagree with the conclusion. You ask the critical question that the pattern itself invites. It is not aggressive; it is what the argument is asking for.

Freeman and the Distinction That Made It Computable

James Freeman's 1991 work on argument macrostructure added the piece that mattered most for machines. He drew a hard line between two things people casually call disagreement.

You can attack a conclusion — argue that it is false. Or you can attack the inference — accept every fact offered and deny that the conclusion follows from them. Your study is flawed and your study is fine but it doesn't show what you think it shows are entirely different moves, and treating them as one loses most of the information in a real disagreement.

Freeman also separated premises that only work together from premises that each stand alone — the difference between an argument that collapses if you remove one leg and one that merely gets weaker. Once you have those distinctions, an argument is no longer prose. It is a labelled graph, and a labelled graph is something a computer can be scored on.

Then Nothing Happened for Twenty Years

The theory was in place by the early 1990s. The computational field did not arrive until the 2010s, and the reason is unglamorous: there was nothing to measure against.

A research field becomes a research field when competing approaches can be compared on the same data. That required someone to sit down with thousands of documents and mark, by hand, which spans were claims, which were premises, and which supported or attacked which. Christian Stab and Iryna Gurevych did it for persuasive essays. Andreas Peldszus and Manfred Stede built a corpus of short argumentative texts specifically to study structure.

That work is covered properly in the next post in this series. What matters here is the shape it revealed. Once you can score systems, you can see which parts are hard — and the answer was immediate and has never really changed.

Finding Claims Is Easy. Everything After That Isn't.

Published benchmarks tell a consistent story. On the persuasive-essay corpus, identifying and labelling the argumentative components reaches roughly 73% F1. Deciding how those components relate to each other — what attaches to what, and whether it supports or attacks — sits near 48%. On the CDCP corpus of public-comment data built by Joonsuk Park and Claire Cardie, relation detection drops to around 34%.

That gap is the field. It is not an artefact of weak models, and it has survived every architecture change since. Detection improved steadily. Relations barely moved.

The reasons are structural rather than technical, and they are the subject of the third post in this series: disagreement is rare in the data, it depends on context the sentence does not contain, and the connective step is usually unstated — Toulmin's warrant, still missing, sixty years later.

But Isn't This Just Natural Language Processing?

It uses the same machinery, so the objection is fair. The difference is in what is being recovered.

Sentiment analysis asks how a text feels. Summarisation asks what it says. Named-entity recognition asks who and what appear in it. All three are about content. Argument mining asks about structure — which sentence is doing argumentative work on which other sentence, and in which direction.

That distinction has a practical edge. A sentence can be phrased negatively while supporting the claim it is attached to: timeline pressure adds to the risk strengthens an argument that the project is risky. Any system scoring by tone gets that backwards, every time. The information argument mining recovers is not in the words. It is in the relations between them.

What Large Models Changed, and What They Didn't

The most recent chapter is the one everyone knows. General-purpose language models can perform argument-mining tasks without being trained on an annotated corpus first, which removes the single biggest practical barrier the field ever had: needing thousands of hand-labelled examples for every new domain.

That is a genuine discontinuity, and it is what made argument mining usable on ordinary business text rather than only on curated essay corpora. It is covered in the fourth post.

What it did not do is close the gap. The structural problems are still structural. A model that has read the entire internet still has to decide whether this objection was aimed at that reason or at the conclusion behind it, and the text still does not say.

What Sixty Years of This Suggests About Your Own Arguments

• The warrant is where the disagreement usually lives. Two people who agree on every fact and still disagree are disputing an unstated principle. Name it and the conversation gets shorter.
• Attacking the inference is not the same as attacking the facts. Deciding which one you are doing — out loud — saves the other person from defending the wrong thing.
• Every appeal to authority comes with its own test question. Walton wrote them down. Using them is not hostility; it is the pattern working as intended.
• Structure is what survives. Six months later nobody remembers the wording. What they need is which objection was raised and whether anyone answered it.

The Anti-Logic Book

Toulmin's critics were right about one thing: The Uses of Argument really was not a book about formal logic. It was a book about how people reason when the stakes are real and the premises are contestable and nobody has time for a syllogism.

That turned out to be the harder problem and the more useful one. Every system that reads a transcript today and tells you what was claimed, on what grounds, and what argued against it is working from his description. Not bad for an anti-logic book.

The argument mining series

Five posts on how machines learned to read arguments — where the field came from, why its hard problems are hard, and how we apply it.