Does support mean the target has been proved?

No. It means the source is presented as a reason for the target under this interpretation. Proof or persuasiveness is a separate assessment.

Can a claim also support another claim?

Yes. An intermediate conclusion can receive support and then function as a reason for a further conclusion.

Are attack and contradiction identical?

No. An objection can weaken a recommendation or challenge an inference without asserting the exact logical negation of a statement.

AI Glossary / Natural language and speech / Intermediate

Argumentative relations

Definition

Argumentative relations are the connections that show how one argument component bears on another, such as providing support or raising an objection. A relation identifies a source component, a target component, and a type of connection. In argument mining, these links help reconstruct the reasoning structure of a passage rather than merely list its claims and premises.

What it means

A support relation means that a statement is presented as a reason for accepting another statement. An attack relation means that it is presented as a reason against a statement or, in some schemes, against an inference. The relation describes a role in the argument; it does not certify that the reason is true or persuasive.

Annotation schemes vary. Some use a small support/attack inventory, while others distinguish objections to a conclusion from challenges to the connection between a premise and conclusion. The chosen labels and edge direction must be stated so readers can interpret the resulting graph correctly.

How it works

A system first identifies candidate argument components, then decides which pairs or groups are connected and how. Words such as because, therefore, and however provide clues, but relation direction and type require interpretation of the actual statements.

In a graph representation, an edge can run from a reason to the claim it supports. Under that convention, reversing the edge changes the meaning. A chain can contain an intermediate claim that is supported by one statement and itself supports another.

Not every related-looking pair has an argumentative edge. Two sentences can share a topic or appear next to each other without either being offered as a reason for the other. Systems need a no-relation possibility and should preserve uncertainty when context does not determine the intended link.

Three practical examples

1. Business proposals: A cost estimate may support a recommendation to postpone a purchase, while a reliability concern supports replacing the equipment sooner. Mapping the targets keeps competing reasons distinct.

2. Review summaries: “The battery lasts all day, so it suits travel” contains an explicit support connection. Merely labeling both clauses positive would miss the explanatory relationship.

3. Debate analysis: A reply may challenge a study's relevance rather than deny its reported result. A richer relation scheme can preserve that difference instead of treating every objection as a contradiction of the data.

Think of it this way

Think of argumentative relations as labeled arrows between notes on a board. An arrow marked supports says how one note is being used in relation to another. The analogy does not make the arrows logically valid: someone can draw a support arrow between weakly related claims. Extracting the intended arrow and assessing whether it is justified are separate tasks.

Worked example

Consider four invented components. C: “We should replace the meeting-room projector.” P1: “Its lamp fails repeatedly.” P2: “Replacement parts now take three weeks to arrive.” O: “A new unit would exceed this quarter's equipment budget.”

Using arrows from reason to target, a simple annotation has P1 supporting C, P2 supporting C, and O attacking C. There are four nodes and three labeled directed edges. With four distinct nodes, there are 4 times 3 = 12 possible ordered source-target pairs if self-links are excluded, but only three are positive relations in this chosen annotation.

A classifier that predicts no relation for every pair is correct on nine of twelve pairs, or 75 percent, yet recovers none of the argument's three actual links. This illustrates why overall pair accuracy can be misleading when non-relations dominate. Precision and recall for the relevant edge types, together with structural review, provide more information.

The example does not settle whether the projector should be replaced. The reasons may all be accurate yet carry different weights. These invented counts illustrate evaluation, not a procurement recommendation.

Illustrative case study

Fictional case: Cedar Planning maps feedback on a new service process. Its first system treats every negative sentence as an attack on the main proposal, even when the sentence criticizes the existing process and therefore supports changing it.

The team asks reviewers to identify the target of each reason before assigning support or attack. It retains the original spans beside the graph so the direction can be checked. The fictional case demonstrates why sentiment polarity cannot substitute for an argumentative relation; it reports no measured outcome.

Common misconceptions and mistakes

Do not equate support with positive language or attack with hostile language. A politely expressed cost concern can attack a proposal, while a negative description of the current situation can support a change.

Do not infer direct edges solely from topic similarity or adjacency. Also avoid multiplying the amount of support by counting paraphrases as independent evidence. A graph may contain several restatements of one premise, and its structure alone does not measure how much independent justification exists.

Business use

Argumentative relations help teams preserve why a position was advanced or challenged in proposals, customer discussions, and review documents. They can make disagreements easier to inspect by showing whether participants contest the conclusion, a premise, or the relevance of a reason.

An evaluation should check source span, target span, direction, and relation label. A correctly identified support word is insufficient if the link points to the wrong claim. Document whether the system represents individual-premise links, jointly supporting premises, or challenges to inferences, since these choices affect what the graph can express.

Check your understanding

“The current system loses messages, so we should replace it.” Does the negative first clause attack the replacement proposal?

Answer: No. In this passage, the problem with the current system is offered as support for replacement. The target and argumentative role matter more than the clause's negative sentiment.

Related terms

Argumentation mining
Premise
Relation extraction
claim
support
rebuttal
undercutting
stance
argument graph

Learn next

Compare sentiment with support/attack labels on the same text. Then study the difference between challenging a premise, challenging a conclusion, and challenging the inference connecting them, using a clearly defined annotation scheme.

Premise

Sources

Source: Stab and Gurevych, Parsing Argumentation Structures in Persuasive Essays, original paper, read for directed argumentative relations and structure-identification tasks: https://aclanthology.org/J17-3005.pdf . The edge-count exercise and business examples are original. Richer relation distinctions vary across argumentation frameworks and should not be assumed to be labels in every dataset.

Previous: Argumentation mining Next: Argumentative zoning