What is argument mining? Argument mining is the branch of natural language processing that automatically identifies the argumentative structure inside ordinary text — which sentences make claims, which give reasons, and whether each reason supports or attacks the claim it is attached to.
Argument mining (also called argumentation mining) turns unstructured prose — a meeting transcript, a policy consultation, an essay, a forum thread — into a structured map of who argued what and on what grounds. The field is usually described as four linked tasks: finding the argumentative spans in the text, labelling each one as a claim or a premise or evidence, working out which span responds to which, and deciding whether each relation is support or attack. It grew out of argumentation theory — Stephen Toulmin's 1958 account of how everyday arguments are actually built, and Douglas Walton's catalogue of recurring reasoning patterns — and became a computational discipline in the 2010s through annotated corpora such as Christian Stab and Iryna Gurevych's persuasive essays and Andreas Peldszus and Manfred Stede's argumentative microtexts. Argumentree uses argument mining to turn a pasted transcript or document into a pro and con argument tree that people can then read, correct, rate and extend.
What Is Argument Mining?
Argument mining is the branch of AI that reads ordinary text and works out its argumentative structure — which sentences are claims, which are the reasons given for them, and whether each reason supports or attacks the point it is attached to. It is the technology behind turning a meeting transcript into a structured pro and con tree.
Last updated: 2026-07-29
In short
Argument mining (or argumentation mining) automatically extracts the reasoning inside unstructured text. Where sentiment analysis asks how does this text feel and summarisation asks what does it say, argument mining asks what is being claimed, on what grounds, and what speaks against it. It is what lets software convert a transcript, a consultation response or an essay into an argument map rather than a summary.
The four jobs argument mining has to do
Researchers usually break the problem into four linked tasks. Each one is harder than the last, and an error early on propagates through everything after it.
1. Find the argumentative parts
Most sentences in a real document are not arguments — they are greetings, logistics, background, or restatement. The first job is deciding which spans of text actually carry a claim or a reason.
2. Label what each part is
A claim is a position someone is taking. A premise is a reason offered for it. Evidence is the data, study or example backing the reason. The same sentence can be a claim in one document and a premise in another — the role depends on what it is doing, not on how it is worded.
3. Work out what responds to what
Arguments are not a flat list. A reason attaches to a specific claim; a rebuttal attaches to a specific reason. Recovering that structure is what turns a pile of extracted sentences into a tree.
4. Decide support or attack
For every link in the tree: does this strengthen the thing it is attached to, or weaken it? This is the step that most often goes wrong, and the reason is explained further down.
Only when all four are done do you have something you can actually navigate: a claim, the reasons for it, the objections to those reasons, and the replies to those objections.
Where argument mining came from
Argument mining is a young computational field standing on a much older theoretical one. The theory came first by about half a century.
Toulmin breaks with formal logic
In The Uses of Argument, Stephen Toulmin argues that real arguments do not look like syllogisms. He proposes claim, grounds, warrant, backing, qualifier and rebuttal as the parts of an everyday argument — the vocabulary the whole field still uses.
Walton catalogues how people actually argue
Douglas Walton documents recurring patterns of everyday reasoning — appeal to expert opinion, argument from analogy, argument from consequences — each with a set of critical questions that expose where it can fail.
Freeman formalises argument structure
James Freeman distinguishes premises that only work together from premises that each stand alone, and separates attacking a conclusion from attacking the inference that leads to it — a distinction that matters enormously once a machine has to make it.
The corpora arrive
Argument mining becomes a measurable computational task when annotated datasets appear. Stab and Gurevych annotate claims, premises and relations across persuasive essays; Peldszus and Stede build a corpus of short argumentative texts specifically to study argument structure. Shared benchmarks make progress comparable.
Statistical and neural models
Systems learn to tag argumentative spans and classify relations from labelled examples. Component detection becomes reasonably reliable. Relation detection — deciding what attaches to what, and whether it supports or attacks — stays stubbornly hard.
Large language models change the economics
General-purpose language models can perform argument mining tasks without being trained on an annotated corpus first. That removes the biggest practical barrier — needing thousands of hand-annotated examples per domain — and makes argument mining usable on ordinary business text.
Why it is harder than it sounds
Extracting claims is the easy half. The difficulty is in the relations — and there are three reasons for that.
Most arguments leave the connection unsaid
"The bridge is unsafe, so we should close it" hides an unstated principle: that unsafe structures should be closed. Toulmin called this the warrant. People almost never say it aloud, which means a machine has to reconstruct a step that is not in the text.
Attacks are rare, so models under-predict them
In most annotated corpora the great majority of relations are support; attack is a small minority. A model trained or prompted on that distribution learns that guessing support is usually safe — which is exactly wrong for the cases that matter most in a decision.
Agreeing with the topic is not the same as supporting the parent
Under an objection, a reason that reinforces the objection is supporting its parent while opposing the overall proposal. Systems that judge stance by how the sentence feels about the topic get this backwards, and the error compounds down the branch.
How Argumentree uses argument mining
Argumentree applies argument mining to a specific job: turning a transcript, document or thread into a pro and con argument tree that a team can then work with. The design follows the field's practice rather than inventing a private method.
Every argument stays tied to its source
Each extracted argument carries the exact wording it came from. If a claim cannot be traced back to something actually present in the source, it does not make it into the tree.
Support and attack are modelled properly
Following Walton and Freeman, an objection can dispute a conclusion, dispute the evidence behind it, or dispute the inferential step connecting them. Those are different moves and are treated as different relations.
The output is a draft, not a verdict
Extraction produces a proposed tree. People review, correct, re-parent and rate it before it becomes the record. The machine does the tedious first pass; the judgement stays with the group.
Structure survives the conversation
Because the result is a tree rather than a summary, later contributors can attach a new objection to the precise reason it addresses, months after the original discussion.
The practical effect is that the reasoning behind a decision stays inspectable — not just what was decided, but which arguments were made, what answered them, and what was never answered at all.
Explore further
Argument mapping
The manual counterpart: laying out claims and reasons visually by hand. Argument mining automates what argument mapping does deliberately.
Argumentation theory
The wider field that argument mining computes over — logic, rhetoric, dialectic and the formal frameworks behind them.
The Toulmin model
The six-part anatomy of an everyday argument, and the source of the claim, grounds and warrant vocabulary used throughout argument mining.
AI meeting intelligence
Argument mining applied to meeting transcripts — capturing why a decision was made, not just what was said.
Frequently asked questions
What is argument mining in simple terms?
Argument mining is software reading a piece of text and working out the argument inside it: what is being claimed, what reasons are given, and which reasons push for or against which claims. The result is a structured map rather than a block of prose.
How is argument mining different from summarisation?
A summary compresses a text and tells you roughly what it said. Argument mining preserves structure: it tells you that this particular objection was raised against that particular reason, and that nobody answered it. Summaries lose exactly the disagreement that matters for a decision.
Is argument mining the same as sentiment analysis?
No. Sentiment analysis measures attitude — positive, negative, neutral. Argument mining recovers reasoning. A sentence can be phrased negatively while supporting the point it is attached to, and sentiment analysis will get that backwards every time.
Who invented argument mining?
There is no single inventor. The theoretical groundwork is Stephen Toulmin's 1958 model of argument structure, extended by Douglas Walton's argumentation schemes and James Freeman's account of argument macrostructure. The computational field took shape in the 2010s, with the annotated corpora from Christian Stab and Iryna Gurevych and from Andreas Peldszus and Manfred Stede among the work that made it measurable.
Why is detecting attacks harder than detecting support?
Two reasons. Attacks are much rarer than support in the datasets these systems learn from, so models are biased toward predicting support. And an attack often has to be judged relative to its immediate parent rather than to the overall topic — a reason can strengthen an objection while opposing the proposal that objection is aimed at.
Does argument mining replace human judgement?
No, and treating it that way is a mistake. Extraction produces a first draft of the argument structure. Deciding whether an argument is any good — whether the evidence holds, whether the unstated assumption is acceptable — is human work. The value is in not having to build the structure by hand before that work can start.
References & further reading
Toulmin, S. E. (1958). The Uses of Argument. Cambridge University Press.
The Claim-Data-Warrant-Backing-Qualifier-Rebuttal model - the theoretical foundation for argument structure.
View source →Walton, D., Reed, C., & Macagno, F. (2008). Argumentation Schemes. Cambridge University Press.
The catalogue of recurring reasoning patterns and the critical questions that test them - the vocabulary for support and attack relations.
Freeman, J. B. (1991). Dialectics and the Macrostructure of Arguments: A Theory of Argument Structure. Foris / De Gruyter.
Support, rebuttal and undercut in a proponent-opponent exchange - the distinction between attacking a conclusion and attacking the inference to it.
View source →Peldszus, A., & Stede, M. (2013). From Argument Diagrams to Argumentation Mining in Texts: A Survey. International Journal of Cognitive Informatics and Natural Intelligence, 7(1), 1-31.
Operationalizing the Freeman model for automated argument mining.
View source →Stab, C., & Gurevych, I. (2014). Annotating Argument Components and Relations in Persuasive Essays. Proceedings of COLING 2014.
Foundational computational argument mining - the annotated corpus and scheme behind automated argument extraction.
View source →Lawrence, J., & Reed, C. (2019). Argument Mining: A Survey. Computational Linguistics, 45(4), 765-818.
The standard survey of the field - task definitions, methods and open problems.
View source →The argument mining series
Turn a transcript into an argument tree
Paste a meeting transcript, consultation response or document and get a structured pro and con tree your team can review, correct, rate and build on.
Start free