Does it decide who is right?

Not inherently. Extracting argumentative structure is distinct from evaluating truth and reasoning quality.

Is every opinion an argument?

An opinion may express a position without providing a reason for it.

Can a language model perform the task?

It can be used, but its extracted spans, labels and relationships still need task-specific evaluation and source checks.

AI Glossary / Data science and analytics / Advanced

Argument mining

Definition

Argument mining is the automatic identification of claims, reasons and argumentative relationships in natural-language material. It aims to recover how a text supports or challenges a position, rather than merely recognizing its topic or emotional tone. The extracted structure is an interpretation of the argument expressed; it is not, by itself, a judgment that the argument is true or sound.

What it means

A statement such as “We should extend support hours because customers in other time zones cannot reach us” contains a proposal and a reason offered for it. Argument mining tries to distinguish those roles and connect them. Larger discussions may include objections, replies and implied assumptions.
The business value lies in making reasoning easier to inspect across many documents. It can help a reviewer find why people favor a change, where evidence is offered and which concerns remain unanswered. The output should remain connected to the original wording so that interpretation can be checked in context.

How it works

Define an annotation scheme that specifies the argument components and relationships of interest. Identify relevant text spans, label their roles and infer relationships such as support or attack. Some systems tackle these tasks in stages; others model them jointly. The structure depends on the task.
Training and evaluation require examples labeled under consistent guidelines. A sentence may contain multiple components, and a premise may support one claim while challenging another. Context and speaker attribution therefore matter.
Represent the result as linked spans or an argument graph. Preserve uncertainty and distinguish explicitly stated content from reconstructed assumptions. Review component detection and relationship extraction separately, because a system can find the right sentences yet connect them incorrectly.

Three practical examples

A product team analyses customer feedback about a proposed redesign. It groups proposals with their stated reasons and objections, helping reviewers inspect the reasoning behind requests instead of relying only on positive or negative sentiment.
A learning platform offers writing feedback by highlighting a student's claim and possible supporting reasons. The student can correct the interpretation rather than accepting an automated map as the only valid reading.
An editorial team examines review comments on draft articles. It separates evaluations from suggestions and checks which criticisms include explanations, while leaving decisions about factual accuracy to a separate review.

Think of it this way

Imagine placing notes on a board and drawing arrows between them: this note gives a reason for that proposal, while another challenges it. Argument mining attempts to build that board from language. A neat set of arrows can still misrepresent a speaker, and a well-connected argument can still rest on a false premise. The map helps inspection; it does not replace it.

Worked example

Consider a fictional discussion with three statements. S1 says, “The glossary should include short audio explanations.” S2 says, “Some readers prefer listening while travelling.” S3 says, “Producing audio will delay the launch.” Under a simple annotation scheme, S1 is the proposal, S2 supports it and S3 attacks the proposal by raising a cost.
Suppose a small evaluation set contains four annotated support-or-attack links. A system predicts five links, of which three match the reference links, including direction and label. Precision is 3 / 5 = 60%. Recall is 3 / 4 = 75%. F1 is 2 × 0.60 × 0.75 / (0.60 + 0.75), approximately 66.7%.
These invented metrics evaluate agreement with a specified reference annotation. They do not measure whether audio is a good investment, whether the premises are factually correct or whether the reference interpretation is beyond dispute. A wrong arrow direction counts as a different relation under this example's matching rule.

Illustrative case study

In a fictional project, Northbank Learning receives extensive comments on a new course format. A simple sentiment summary shows mixed reactions but gives little guidance about what to change. The team pilots argument mining to connect proposals, reasons and objections.
Reviewers discover that several negative-sounding comments actually support the proposal while criticizing its implementation schedule. They also correct cases where the system attributed a quoted objection to the person quoting it. The interface keeps each extracted component beside its original comment and allows uncertain relationships to remain unresolved.
Northbank uses the reviewed map to structure a planning meeting and identify questions requiring more evidence. It does not count every extracted support link as an independent vote or treat repeated wording as independent evidence. This fictional case illustrates a review workflow.

Common misconceptions and mistakes

Argument mining is not simply sentiment analysis. A calmly worded statement can attack a proposal, while an emotional statement may provide no supporting reason. It also differs from fact-checking: extracting an assertion does not verify it.
Words such as “because” and “therefore” can help, but their presence is not a complete argument detector. Implicit premises, quotations, irony and context can change the interpretation.
Another mistake is forcing every passage into a tidy tree. Real discussions may contain shared premises, multiple targets and unresolved relationships. The representation should preserve uncertainty in the text.

Business use

Use argument mining to assist document review, feedback analysis and evidence organization where understanding reasons matters. Start with a defined document type and reviewed sample labels.
Measure extraction quality and reviewer effort, including errors that could reverse someone's position. Keep source attribution and allow corrections. A system tested on essays may not transfer well to short support messages or meeting transcripts. Decisions should draw on the underlying evidence and context, with the argument map serving as an aid to navigation and analysis.

Check your understanding

Why keep original text beside the graph?

Answer: To check context, attribution and interpretation.

What does a support link represent?

Answer: A reason presented as supporting a target claim under the chosen scheme.

Why can high component accuracy hide problems?

Answer: Correctly detected components may still be linked with the wrong direction or relationship.

Related terms

Learn next

Annotate a short discussion manually, comparing claim boundaries and relationship labels with another reader. Then inspect disagreements before deciding how an automated system should be evaluated.

Sources

Lawrence and Reed, Argument Mining: A Survey, 2019: task structure, representations and challenges. https://aclanthology.org/J19-4006.pdf Hua and colleagues, Argument Mining for Understanding Peer Reviews, 2019: primary research on propositions and their roles in reviews. https://aclanthology.org/N19-1219/ The discussion, metrics and Northbank Learning case are original illustrations and do not report these papers' experimental results.

Previous: Area under the PR curve Next: Argumentation framework