Can a closed class gain members?

Yes. Relative stability does not exclude historical change.

Are all frequent words closed-class words?

No. Frequency and grammatical membership are different properties.

Should these words always be removed for search?

No. That depends on the search task and evaluation results.

AI Glossary / Natural language and speech / Advanced

Closed class

Definition

A closed class is a grammatical word category whose membership is relatively stable and does not readily expand through ordinary creation of new words. English determiners, pronouns, and conjunctions are familiar examples. Closed classes contrast with open classes, such as English nouns, which regularly acquire new members. “Closed” describes a tendency in vocabulary membership, not an absolute ban on language change.

What it means

People can invent a name for a new device or product quite easily. Introducing a new item that behaves like the, and, or an established pronoun is a different kind of change: it must fit an existing grammatical system and become accepted in use. This is the contrast the term captures.
Closed-class words often express relationships, reference, or grammatical distinctions rather than naming a topic. They can be short and frequent, but neither shortness nor frequency defines the category. Frequent business nouns do not become conjunctions through repetition.
For natural language processing, these words provide important structural cues. Recognizing topic words is insufficient if a system loses negation, alternatives, reference, or relationships among participants.

How it works

A grammatical analysis assigns words to categories using their behavior in context. Closed versus open is a broad grouping of categories, not usually the most detailed label assigned to each token. For example, a tagging scheme can distinguish determiners, pronouns, auxiliaries, and conjunctions within the broader closed-class group.
The exact inventory depends on the language and annotation scheme. Universal Dependencies lists closed-class categories in its tag inventory, but its broad organization should not be interpreted as a claim that every language realizes each category identically. Surface forms can also be ambiguous: can in “we can swim” has a different grammatical role from can in “the empty can.”
Membership can change over time. Calling a category relatively closed means that adding members is less routine than creating new content vocabulary. Grammars and annotation lists can be revised.
A processing system should therefore combine category information with context and documented conventions. A fixed list can support an application, but it cannot settle every grammatical analysis or determine which words are safe to discard.

Three practical examples

A parser uses determiners and prepositions as cues to phrase structure and relationships, alongside the surrounding nouns and verbs.
A document assistant preserves the difference between “send the invoice and the receipt” and “send the invoice or the receipt,” rather than treating both conjunctions as disposable text.
A corpus analyst reports grammatical categories separately from word frequency, avoiding the assumption that all common words belong to a closed class.

Think of it this way

Think of open-class vocabulary as the things being discussed and many closed-class words as instructions for connecting them. The connectors may be less visually prominent, but changing a connector can change the message. This analogy is not a complete definition.

Worked example

Consider two original instructions:
“The clerk can release the parcel”
“The clerk can not release the parcel”
For this teaching example, use the following contextual tags. Both occurrences of the are determiners; clerk and parcel are nouns; can is an auxiliary; release is a verb; and not is a negation particle. Separate can and not spellings avoid a hidden segmentation decision.
The first instruction contains six tokens: three closed-class tokens, the, can, and the, and three open-class tokens, clerk, release, and parcel. Its closed-class token proportion is 3/6=50%.
The second contains seven tokens. Adding not gives four closed-class tokens and the same three open-class tokens. Its closed-class token proportion is 4/7, approximately 57.1%. These percentages describe these annotations, not English generally.
Now suppose a deliberately crude preprocessing rule removes every token assigned to the closed-class categories used here. Both sentences reduce to “clerk release parcel.” Their resulting token sequences are identical, even though one expresses the possibility or permission of release and the other negates it in the intended reading.
The example demonstrates information loss, not a recommendation to infer a complete business authorization from either sentence alone. Context still determines whether can concerns permission or ability and which parcel is intended. Keeping the words is necessary for preserving the distinction in this representation, but keeping them does not automatically guarantee correct interpretation.
A narrower, task-specific preprocessing choice might behave differently. The relevant question is whether the transformation preserves the distinctions the application needs, not whether the removed words are frequent or belong to a relatively stable grammatical class.

Illustrative case study

Fictional example: ParcelNotes groups internal messages by topic. A developer reuses its keyword-cleaning step for a new assistant intended to interpret operational instructions. Reviewers notice that affirmative and negative messages can receive identical cleaned representations.
The team separates the requirements of topic grouping from instruction interpretation. It retains the original text for the latter and adds contrasting examples that differ only in negation or conjunction. The scenario illustrates why a preprocessing choice must be evaluated against its use; it does not claim that every stop-word removal method causes the same error.

Common misconceptions and mistakes

Closed class does not mean meaningless. Grammatical words can carry decisive information about a statement. It also does not mean stop word: a stop-word list is a processing choice, whereas closed class is a linguistic classification.
Another mistake is equating a limited inventory with a limited number of occurrences. A small set of word types can account for many tokens in a document. Finally, the same spelling can have different categories in different contexts, so a global spelling-based rule may confuse an auxiliary with a noun or another use.

Business use

Review closed-class handling when building search, summarization, translation, or instruction-following workflows. Test minimal contrasts involving negation, conjunctions, and reference. Preserve original text so analysts can inspect what preprocessing removed. When comparing reports, record both the tagging convention and whether counts concern distinct word types or individual token occurrences.

Check your understanding

How many closed-class tokens occur in the second example?

Answer: Four under the stated tagging scheme.

Why do both cleaned sequences match?

Answer: The rule deletes all their differing closed-class material.

Is can always an auxiliary?

Answer: No. Its use must be interpreted in context.

Related terms

Learn next

Tag a short passage, count types separately from tokens, and compare what happens when different preprocessing rules remove grammatical words. Check which meaning distinctions survive.

Sources

Jurafsky and Martin, Speech and Language Processing, chapter 18, section 18.1, open and closed word classes: https://web.stanford.edu/~jurafsky/slp3/18.pdf Universal Dependencies, universal part-of-speech inventory: https://universaldependencies.org/u/pos/

Previous: Close-ended question Next: Closed-form matting