Argument Mining

Definition

Argument mining automatically identifies and extracts arguments from natural language, in text and in speech. It finds the claims people make, the premises they offer as reasons, and the relations between them. Opinion mining tells what people think about a topic, while argument mining asks why they think it (Lawrence and Reed, 2019).

Core concepts

Arguments

Argument mining first separates argumentative text from the rest of a document. It then marks each argument component with its boundaries and its type. The two most common types are claims, the statements under discussion, and premises, the reasons given for or against them.

CLAIMPREMISE

Relations

Relations, typically support and attack, connect argument components. Linking the components of a document through these relations gives an argument graph, whose nodes are components and whose edges are relations.

ClaimPremisePremisesupportsattacks

Evaluation

An argument takes a stance on a topic, for or against it. Its quality is scored along dimensions such as clarity, relevance, and persuasiveness. Quality judgements are subjective, so annotators often disagree on them.

ArgumentSTANCEForQUALITY3 of 5

Reasoning

Arguments often leave steps unstated, such as the warrant that links premises to a claim. Reasoning tasks recover these steps, identify the argumentation scheme of an argument, and recognize fallacies. An argumentation scheme is the reasoning pattern an argument follows, and a fallacy is a flaw in reasoning.

WarrantimplicitPremiseClaim

What is an argument?

The literature offers several definitions of an argument. A common one comes from Walton, as reported by Lippi and Torroni (2016). In it, an argument is a set of premises, a conclusion, and an inference from the premises to the conclusion. Argument mining typically calls the conclusion a claim. A premise can support a claim or attack it, and claims on opposite sides can attack each other.

An argument model defines the components of an argument, such as the claim and the premise, and the relations between them. The claim and premise model marks only these two components. The Toulmin model calls premises grounds and adds two more components. The warrant is the general rule that links the grounds to the claim, and the text can leave it unstated. The rebuttal states when the claim does not hold. The original model of Toulmin (1958) adds the backing, a support for the warrant, and the qualifier, the degree of certainty of the claim.

Each domain also names these components in its own way. In debate support, evidence takes the place of premises and supports a claim on a given topic. In clinical trial abstracts, evidence is an observed outcome of a study, and a claim is what the authors infer from it. In court decisions, the claim is called the conclusion, and premises are either factual or legal. The examples below show one argument for each model and each domain.

The most common model. A claim is the statement being argued for, and premises are offered as reasons for it. Some annotations also mark premises that attack a claim.

1 The text

Despite the fact that P1 Premiseadvertisements can be falseful and exaggerated, it is also true that C1 Claimit plays an important role economically. P2 PremiseThey introduce new products.

2 Argument graph

Claim C1
Premise P1
Premise P2
  • Premise P1 "advertisements can be falseful and exaggerated" attacks Claim C1 "it plays an important role economically."
  • Premise P2 "They introduce new products." supports Claim C1 "it plays an important role economically."
Example reported by Sun et al. (2024). Original wording preserved.

A richer model. Grounds support a claim through a warrant, the general rule that links them, while a rebuttal states when the claim does not hold.

1 The text

R1 RebuttalCapital punishment is not a solution, as it cannot be ruled out that the judicial process may make mistakes. [...] After all G1 Groundthere are criminals who enjoy 'living' in prison for its decent board [...]. Still, C1 Claimthe state needs the death penalty as a deterrent to horrific crimes.

W1 Warrant (implicit)The death penalty deters criminals from committing horrific crimes.

2 Argument graph

Rebuttal R1
Claim C1
Warrant (implicit) W1
Ground G1
  • Ground G1 "there are criminals who enjoy 'living'" supports Claim C1 "the state needs the death penalty as a deterrent to horrific crimes."
  • Warrant W1 "The death penalty deters criminals from committing horrific crimes." licenses Claim C1 "the state needs the death penalty as a deterrent to horrific crimes."
  • Rebuttal R1 "Capital punishment is not a solution, as it cannot be ruled out that the judicial process may make mistakes." attacks Claim C1 "the state needs the death penalty as a deterrent to horrific crimes."
Example from Gupta et al. (2024), shortened. The warrant is not written in the text, so it is shown reconstructed below the text, with a dashed border.

In debate support, a system finds claims on a given topic and the evidence for them, for instance in Wikipedia articles. Claims on opposite sides attack each other.

1 The text

While those on the far-right think that C1 Claimimmigration threatens national identity [...]. Proponents of immigration maintain that, E1 Evidenceaccording to Article 13 of the Universal Declaration of Human Rights, everyone has the right to leave or enter a country, along with movement within it. [...] Some argue that C2 Claimthe freedom of movement both within and between countries is a basic human right. [...] E2 EvidenceImmigration has been a major source of population growth and cultural change throughout much of the history of Sweden. The economic, social, and political aspects of immigration have caused controversy regarding ethnicity, economic benefits, jobs for non-immigrants, settlement patterns, impact on upward social mobility, crime, and voting behavior.

2 Argument graph

Claim C1
Claim C2
Evidence E2
Evidence E1
  • Evidence E2 "Immigration has been a major source of population growth and cultural change throughout much of the history of Sweden. The economic, social, and political aspects of immigration have caused controversy regarding ethnicity, economic benefits, jobs for non-immigrants, settlement patterns, impact on upward social mobility, crime, and voting behavior." supports Claim C1 "immigration threatens national identity"
  • Evidence E1 "according to Article 13 of the Universal Declaration of Human Rights, everyone has the right to leave or enter a country, along with movement within it." supports Claim C2 "the freedom of movement both within and between countries is a basic human right."
  • Claim C1 "immigration threatens national identity" attacks Claim C2 "the freedom of movement both within and between countries is a basic human right."
  • Claim C2 "the freedom of movement both within and between countries is a basic human right." attacks Claim C1 "immigration threatens national identity"
Example from the IBM debate corpus, shortened from Figure 1 of Lippi and Torroni (2016).

Clinical trial abstracts separate evidence, the observed outcomes of a study, from the claims the authors draw from it. Evidence is further typed, for instance as a comparison between groups or as a significance result.

1 The text

E1 Evidence · SignificanceThe diurnal intraocular pressure reduction was significant in both groups (P < 0.001). E2 Evidence · ComparativeThe mean intraocular pressure reduction from baseline was 32% for the latanoprost plus timolol group and 20% for the dorzolamide plus timolol group. [...] This study clearly showed that C1 Claimthe additive diurnal intraocular pressure-lowering effect of latanoprost is superior to that of dorzolamide in patients treated with timolol.

2 Argument graph

Claim C1
Evidence E1 Significance
Evidence E2 Comparative
  • Evidence E1 "The diurnal intraocular pressure reduction was significant in both groups (P < 0.001)." supports Claim C1 "the additive diurnal intraocular pressure-lowering effect of latanoprost is superior to that of dorzolamide in patients treated with timolol."
  • Evidence E2 "The mean intraocular pressure reduction from baseline was 32% for the latanoprost plus timolol group and 20% for the dorzolamide plus timolol group." supports Claim C1 "the additive diurnal intraocular pressure-lowering effect of latanoprost is superior to that of dorzolamide in patients treated with timolol."
Example 1 of Mayer, Cabrio, and Villata (2018), shortened. The paper marks the evidence and the claim; we added the support links and the evidence types following the definitions of the paper.

Research areas

Extraction tasks separate argumentative text from the rest, segment it into argument components, and group arguments by the aspects they address.

Argumentative text detection
Given a sentence or a passage, the task is to decide whether it contains an argument.
Argument component identification
Given an argumentative text, the task is to mark the boundaries of each component and to label it, for instance as a claim or a premise.
Claim and evidence retrieval
Given a topic and a set of documents, the task is to retrieve the claims on the topic and the evidence for each claim.
Aspect mining
Given an argument, the task is to identify the aspects it addresses, such as the cost or the safety of nuclear energy. Arguments on the same aspect can then be grouped.

Component identification finds claims and premises, while relation identification decides how they connect, for instance through support or attack. These relations define an argument graph over the whole document.

Relation identification
Given two argument components, the task is to decide whether one supports or attacks the other, or whether they are unrelated.
Argument structure prediction
Given a document, the task is to predict its argument graph, where components are nodes and relations are edges.

Evaluation tasks assign properties to arguments, such as their stance on a topic or their quality.

Stance classification
Given a topic and an argument, the task is to decide whether the argument is for or against the topic.
Argument quality assessment
Given an argument, the task is to score its quality along dimensions such as clarity, relevance, and persuasiveness. A pairwise variant selects the more convincing of two arguments.

Arguments often leave some steps unstated. These tasks recover the missing steps and assess the inference.

Implicit argument reconstruction
Given an argument with missing steps, the task is to generate the unstated premise or the warrant, the general rule that links premise and claim. This typically requires commonsense or expert knowledge.
Argumentation scheme classification
Given an argument, the task is to identify its argumentation scheme, the reasoning pattern it follows, such as an argument from consequences.
Fallacy recognition
Given an argument, the task is to decide whether its reasoning is flawed and, if so, which fallacy it commits.

Generation tasks write or condense arguments, while retrieval tasks find arguments in large collections.

Argument summarization
Given many arguments on a topic, the task is to produce a short summary or a small set of key points.
Argument generation
Given a topic or an argument, the task is to generate a new argument, potentially for the opposite stance.
Argument retrieval
Given a query, the task is to retrieve relevant arguments from a large collection.

Large language models (LLMs) are now widely applied to argument mining tasks. Argumentation, in turn, offers a way to check how these models reason.

LLMs for argument mining
Prompting and fine-tuning allow a single model to address many tasks. Fallacy recognition and the modelling of diverse perspectives remain challenging.
Argumentation for LLMs
Argumentative prompts elicit LLM reasoning, and dedicated benchmarks evaluate it. Computational argumentation, a formal way of modelling arguments and their conflicts, can verify it.

Our focus

Argumentative fallacies

We study how machines can spot and name fallacies, which are flaws or errors in reasoning that can make an argument seem convincing. Spotting them takes careful reasoning, so they are a hard test for language models.

Current

Built from the task definition in Section 3 of the paper. A sentence from a political debate comes in one of three input modes: text only, audio only, or text with its aligned audio, and systems may also use the previous sentences of the debate as context. Fallacy detection (AFD) decides whether the sentence contains a fallacy. A fallacious sentence then goes to fallacy classification (AFC), which picks one of six fallacy types.DEBATE SENTENCE, THREE INPUT MODESText onlyAudio onlyText + audioprevious sentences as contextDetection (AFD)fallacy or notif fallaciousClassification (AFC)one of six typesFallacy type

Overview of MM-ArgFallacy2025 on Multimodal Argumentative Fallacy Detection and Classification in Political Debates

Workshop on Argument Mining (ArgMining) · 2025

MM-ArgFallacy2025 benchmarks fallacy detection and classification in U.S. presidential debates from text, audio, or both, and finds that text-based systems still score highest.

Past work

  • Multimodal Fallacy Classification in Political Debates
Read more

Multimodality

Arguments are often spoken, not written. We study whether the sound of a voice helps a computer find and classify arguments better than the words alone.

Current

Built from the task definition in Section 3 of the paper. A sentence from a political debate comes in one of three input modes: text only, audio only, or text with its aligned audio, and systems may also use the previous sentences of the debate as context. Fallacy detection (AFD) decides whether the sentence contains a fallacy. A fallacious sentence then goes to fallacy classification (AFC), which picks one of six fallacy types.DEBATE SENTENCE, THREE INPUT MODESText onlyAudio onlyText + audioprevious sentences as contextDetection (AFD)fallacy or notif fallaciousClassification (AFC)one of six typesFallacy type

Overview of MM-ArgFallacy2025 on Multimodal Argumentative Fallacy Detection and Classification in Political Debates

Workshop on Argument Mining (ArgMining) · 2025

MM-ArgFallacy2025 benchmarks fallacy detection and classification in U.S. presidential debates from text, audio, or both, and finds that text-based systems still score highest.

Past work

  • MAMKit: A Comprehensive Multimodal Argument Mining Toolkit
  • Multimodal Fallacy Classification in Political Debates
  • Multimodal Argument Mining: A Case Study in Political Debates
  • Argument Mining from Speech: Detecting Claims in Political Debates
Read more

Dialogues

We study dialogue systems that exchange arguments with people. Such systems can explain their answers, which matters when users need to trust the information they receive.

Past work

  • A Chatbot for Asylum-Seeking Migrants in Europe
  • A Dataset of Argumentative Dialogues on Scientific Papers
  • A privacy-preserving dialogue system based on argumentation
  • An Argumentative Dialogue System for COVID-19 Vaccine Information
Read more

Interpretability

We build models that show why they make a decision. This helps people check a prediction and learn which parts of a text make an argument.

Current

Top: the select-then-predict pipeline, where a generator selects a highlight and a predictor labels it. Bottom: GenSPP searches a population of generators with a genetic algorithm. Each generator is frozen and judged by a predictor trained anew, so no predictor adapts to an earlier selection and interlocking cannot arise.SELECT-THEN-PREDICTGeneratorselects tokensPredictorreads only themlabelGENETIC SEARCH OVER GENERATORSG₁frozenG₂frozenG₃frozenG₄frozenPredictor trained anewone for each generatorfitnessfittest generators breed the next generation

Interlocking-free Selective Rationalization Through Genetic-based Learning

Annual Meeting of the Association for Computational Linguistics (ACL) · 2025

GenSPP trains the generator and predictor of a self-explaining text classifier separately with genetic search, removing interlocking and producing more accurate highlights.

Past work

  • Combining Transformers with Natural Language Explanations
  • Tree-Constrained Graph Neural Networks For Argument Mining
Read more

Reasoning in LLMs

We use argumentation to test how large language models reason. This matters because these models now answer questions and judge content, yet their reasoning often fails in ways that are hard to see.

Current

After Figure 1 of the paper. Within the reasoning task of claim verification and a single-step reasoning process, a claim paired with evidence needs one atomic reasoning type: deduction, where the evidence states the answer, or abduction, where the verdict rests on the most plausible hypothesis. Either leads to a veracity. Averaged over LLMs and prompts on CLIMATE-FEVER, the error rate is 15.58% on deductive pairs and 48.58% on abductive ones.TASK: CLAIM VERIFICATIONPROCESS: SINGLE STEPClaim + evidenceDeductionevidence states itAbductionplausible hypothesisveracityLLM ERROR RATE, CLIMATE-FEVER15.58%48.58%

Assessing the Reasoning Capabilities of LLMs in the context of Evidence-based Claim Verification

Findings of the Association for Computational Linguistics: ACL (Findings of ACL) · 2025

RECV, a benchmark of real-world claims labeled by reasoning type, shows that LLMs verify claims by deduction well but fail when abduction is needed.

Past work

  • Language is Scary when Over-Analyzed: Unpacking Implied Misogynistic Reasoning with Argumentation Theory-Driven Prompts
  • Investigating logic tensor networks for neural-symbolic argument mining
  • Neural-Symbolic Argumentation Mining: An Argument in Favor of Deep Learning and Reasoning
Read more

Argumentative fallacies

A fallacy is a flaw or error in reasoning. Examples are attacking the speaker instead of the claim (ad hominem) and warning that one small step leads to disaster (slippery slope). Fallacy detection asks whether a passage contains a fallacy, and fallacy classification asks which type it is. In political debates, how something is said can matter as much as what is said. We ask whether models can recognize fallacies from both the words and the voice of a speaker, and what reasoning each type of fallacy requires.

Current

Past work

  • Releases the first corpus pairing political debate text with audio for classifying six fallacy types, and shows that adding audio improves classification for several models.

Multimodality

Multimodality means using more than one kind of input, here the text of what people say and the audio of how they say it. Audio carries paralinguistic cues, such as tone, pitch, and pauses, that a written transcript loses. We ask whether these cues help with argument mining tasks, and for which tasks they help.

Current

Past work

Dialogues

A dialogue system is a computer program that talks with a person, such as a chatbot. An argumentative dialogue is a conversation in which the speakers give reasons for their claims and question the reasons of others. We ask how a dialogue system can reason with arguments, explain its answers, and discuss a topic the way people do.

Past work

  • ACME, a chatbot built on computational argumentation, helps asylum seekers find the highest protection they can apply for, with explanations and privacy by design.

  • ArgSciChat collects 41 dialogues in which scientists question and argue about NLP papers, a hard test for dialogue agents that must explain and defend answers.

  • A dialogue system that keeps personal data in a language module and picks replies by argumentation over sanitized facts, so answers are defended and explainable.

  • Dialogue systems are widely used in AI to support timely and interactive communication with users. We propose a general-purpose dialogue system architecture that leverages computational argumentation to perform reasoning and provide consistent and explainable answers. We …

Interpretability

A model is interpretable when people can understand why it gives a certain output. One common form is an explanation, a piece of evidence that justifies a prediction. An explanation can be a sentence of background knowledge or a highlight, a span of the input text. We ask how models for argument analysis can explain their own decisions without losing accuracy. We also ask what these explanations tell us about the data.

Current

Past work

Reasoning in LLMs

Large language models (LLMs) are AI systems trained on large amounts of text to read and write language. Reasoning is the ability to move from given information to a conclusion through steps that someone else can check. Argumentation studies exactly these steps: the claims people make, the reasons they give, and the links between them. We ask whether argumentation can serve as a tool to assess, and later improve, the reasoning of LLMs.

Current

Past work