Does the nearest noun control the verb?

Not necessarily. The grammatical subject controls the relevant agreement, even when another noun appears closer.

Can a model pass agreement tests without understanding everything?

Yes. The tests probe a particular linguistic dependency, not every aspect of meaning or reasoning.

Do all languages use the same agreement features?

No. Languages and constructions differ in which features are expressed and where they appear.

AI Glossary / Graphs and symbolic AI / Beginner

Subject-verb agreement

Definition

Subject-verb agreement is the grammatical relationship in which a verb's form matches relevant features of its subject, such as person and number. In English, familiar examples include “the sensor works” and “the sensors work.” In natural-language processing, agreement is useful both for producing grammatical text and for testing whether a model tracks sentence structure rather than simply matching nearby words.

What it means

The subject is the grammatical expression associated with the verb, not necessarily the nearest noun or the person performing an action. In “the parcels are inspected,” parcels is the subject even though someone else performs the inspection.

For many English present-tense verbs, a third-person singular subject takes an ending such as “works,” while a plural subject takes “work.” Other forms behave differently: “I am,” “you are,” and “she is” illustrate person as well as number. The patterns depend on the language and construction, so a single English rule cannot serve as a universal agreement algorithm.

How it works

A grammar-based system can identify the clause, locate the subject and its head, determine relevant grammatical features, and check whether the finite verb has a compatible form. A neural language model may instead learn statistical patterns that support similar choices without an explicitly supplied parse tree.

Intervening nouns make the task informative. In “the label on the boxes is damaged,” boxes is plural but belongs to a modifying phrase; label controls the agreement. A nearest-noun shortcut would choose the wrong form.

Researchers can compare minimally different continuations, such as “is” and “are,” under the same prefix. A preference for the grammatical continuation is evidence on that particular test. It does not by itself establish complete grammatical knowledge, human-like processing, or reliable comprehension of the sentence's meaning. Evaluation should examine multiple constructions and sources of distraction.

Three practical examples

Writing assistance: flag a possible mismatch in a report while showing enough context for the author to assess the proposed correction.
Template generation: produce a suitable verb when an application inserts one item or several items into a notification, rather than always using a fixed singular phrase.
Model evaluation: test singular and plural subjects separated from their verbs by distracting nouns, measuring whether predictions follow the intended subject.

Think of it this way

Think of the subject and verb as two linked settings that must be compatible even when other words appear between them. The analogy is limited because language is not a simple wiring diagram: pronouns, coordination, tense, dialect, and interpretation can affect the appropriate form. A real editor needs context rather than blindly changing an ending.

Worked example

Consider an original four-item English test. The intended standard-English completions are: “The invoice for the parts is ready”; “The invoices for the part are ready”; “The invoice is ready”; and “The invoices are ready.” The first two contain an intervening noun with a different number from the subject head. The last two are controls without that distraction.

Suppose a hypothetical model assigns the following conditional probabilities to the two candidate next words. For item one, P(is)=0.18 and P(are)=0.12. For item two, P(is)=0.16 and P(are)=0.14. For item three, the values are 0.25 and 0.05; for item four, they are 0.08 and 0.24. Other possible next words account for the remaining probability, so each displayed pair need not sum to one.

Selecting the more probable candidate yields is, is, is, are. The model answers three of four items correctly, or 75%. It gets both controls right but only one of the two distracting-noun items right, or 50% on that tiny subgroup.

The arithmetic illustrates a targeted evaluation, not an actual model result. Four sentences are far too few for a dependable performance estimate. The grammaticality comparison also does not show that the model knows whether any invoice is truly ready.

Illustrative case study

Fictional case: a software team localizes notifications and finds that a template produces “three requests is waiting.” Its initial fix adds a plural verb whenever a number exceeds one, but a reviewer notes that this approach will not cover every sentence structure or language.

The team separates message templates by supported grammatical construction and asks language specialists to review the variants. It tests inserted noun phrases as well as counts. This fictional case illustrates agreement-aware generation without claiming a universal grammar engine or a measured improvement in translation quality.

Common misconceptions and mistakes

The closest noun does not always determine agreement. A plural-looking word ending is not sufficient to identify grammatical number, and the form appropriate for one English tense may not vary in the same way in another. Agreement checking should not be confused with deciding whether a statement is factual.

High average accuracy can hide failure on particular constructions. Conversely, a disagreement with a simplistic checker is not automatically a writer's error: intended meaning and language variety can matter. Editing tools should distinguish clear mismatches from cases requiring contextual or stylistic judgment.

Business use

Agreement quality matters in customer communications, generated reports, and multilingual interfaces because small grammatical mistakes can make otherwise useful output harder to read. Evaluate the constructions the product actually generates, including long noun phrases and changes in the number of entities.

When benchmarking an AI model, document the language variety, target construction, candidate-scoring method, and test size. Keep agreement performance separate from factual accuracy and task success. A system can produce grammatical wording while assigning it to the wrong customer, object, or event.

Check your understanding

Which is appropriate in standard English: “The folder containing the reports is missing” or “are missing”?

Answer: “Is missing. The subject head is singular folder; reports belongs to a modifying phrase and does not control the main verb.

Related terms

Syntax
Language model
grammatical number
person
dependency parsing
subject
finite verb
agreement attraction
minimal pair

Learn next

Create singular and plural versions of five sentences, then insert a noun phrase with the opposite number between each subject and verb. Identify the subject head before choosing the verb. Next study dependency parsing and targeted syntactic evaluation in NLP.

Noun phrase

Sources

Purdue OWL, subject-verb agreement guidance: https://owl.purdue.edu/owl/general_writing/grammar/subject_verb_agreement.html Linzen, Dupoux, and Goldberg, Assessing the Ability of LSTMs to Learn Syntax-Sensitive Dependencies, original research abstract and publication record: https://aclanthology.org/Q16-1037/ All example sentences, probabilities, and the fictional case are original. The research reference is historical evidence about a test approach, not a claim about current model performance.

Previous: Subject indexing Next: Subjective logic