Can a proper name be a singleton?

Yes. Any eligible referring mention can form a singleton if no other mention of that entity occurs within the relevant scope.

Does a singleton mean the entity is mentioned nowhere else?

No. It concerns the particular text or annotation unit, not every document in existence.

Must every coreference dataset include singletons?

No. Annotation and evaluation conventions differ, so the dataset documentation must be checked.

AI Glossary / Natural language and speech / Intermediate

Singleton

Definition

In coreference analysis, a singleton is an entity represented by only one mention within the text or annotation scope being considered. Its mention forms a one-member cluster rather than a chain of two or more references. This use differs from other meanings of singleton in programming and general mathematics.

What it means

A document may name a person once and never refer to that person again. Another person may appear through a name, a pronoun, and a description. Coreference analysis groups references to the same entity; the first person’s group is a singleton, while the second person’s group is not.
The scope matters. A name mentioned once in one paragraph may occur several times in the complete report. Singleton status describes the chosen text and annotation, not a permanent property of the person or object.
It also does not mean the mention is unimportant. An entity mentioned only once can still be essential to the document’s meaning or to an information-extraction task.

How it works

Identify referring expressions under the chosen annotation guidelines, then group the expressions that refer to the same entity. Count the mentions in each group.
A group with exactly one mention is a singleton. A group with two or more mentions contains an explicit coreference connection within the scope.
Datasets differ in whether they annotate or evaluate singleton mentions. This affects the training signal and how system output should be scored. Omitting singletons from a benchmark does not imply that the corresponding references are absent from the actual text.
Mention detection and linking should therefore be distinguished. A system may recognize a valid one-off mention even when there is no second mention to link it with.

Three practical examples

A coreference annotation exercise records a supplier mentioned only once as a singleton entity cluster.
A dataset team documents whether its evaluation includes one-member clusters or only repeated references.
An information-extraction pipeline retains a one-off location mention even though it does not participate in a longer reference chain.

Think of it this way

Think of sorting references into folders, with one folder per entity. A singleton folder contains one reference card. The analogy is limited because real annotation involves interpreting boundaries and identity; the number of cards depends on which text and mention rules the analyst has chosen.

Worked example

Consider an invented passage: “Rina called Omar. She then emailed the architect. Omar replied.” Assume that “she” refers to Rina and that the architect is a third person.
Under a simple person-mention annotation, Rina’s cluster contains “Rina” and “She.” Omar’s cluster contains the two occurrences of “Omar.” The architect’s cluster contains only “the architect.”
There are five mention occurrences grouped into three entity clusters. Two clusters have two mentions each, and one has a single mention. The architect is the singleton in this scope.
Now append “The architect approved the drawing.” The architect’s cluster gains a second mention and is no longer a singleton. The original sentence has not changed; the text scope has expanded.
If a benchmark excludes singleton clusters, the first version’s architect mention may not appear in its coreference labels. That does not justify concluding that the phrase is nonreferential. It reflects the benchmark’s annotation policy rather than the absence of a person in the passage.

Illustrative case study

In a fictional project, Larch Records builds a coreference dataset from short project updates. Its first annotation export includes only entities mentioned repeatedly, while a downstream directory tool expects every person mention.
The team documents the difference and adds a separate mention-detection review for one-off references. It avoids interpreting an unlinked name as automatically invalid or irrelevant.
The imagined case illustrates a mismatch between task definitions. It does not claim that every benchmark should include singletons or that adding them will necessarily improve every coreference system.

Common misconceptions and mistakes

A singleton is not simply a word that appears once. Several different expressions can refer to the same entity, while identical words can refer to different entities.
It is not the same as a nonreferential expression. A genuine reference can occur only once, while dummy “it” may not be an entity mention at all under the annotation scheme.
Singleton status is also not a confidence score. A one-member cluster can be correct, or it can result from missed links.
Finally, excluding singletons from an evaluation does not establish that detecting them has no value for other applications.

Business use

Understanding singletons helps teams connect coreference outputs with search, document indexing, and entity extraction. Many practical systems need one-off references even when a coreference benchmark focuses on repeated mentions.
Before combining datasets or comparing scores, check whether singleton mentions are included, how mention boundaries are defined, and what text scope is used. Otherwise apparent model differences may partly reflect annotation differences.
Error review should distinguish true singletons from clusters fragmented by missed links. A system that leaves every mention separate would create many one-member clusters, but that behavior would not demonstrate successful coreference resolution.

Check your understanding

“The manager entered. She sat down.” Assume “she” refers to the manager. Is “the manager” a singleton merely because that exact phrase appears once?

Answer: No. The name or phrase need not repeat exactly. The two referring expressions belong to one two-member entity cluster.

Related terms

Coreference resolution
Mention detection
Anaphora
Entity linking
entity mention
coreference cluster
referring expression
annotation scheme
discourse entity

Learn next

Learn how coreference groups mentions rather than matching repeated strings. Then annotate a short paragraph with both repeated and one-off entities. Compare your result with the guidelines of a real dataset and note whether its evaluation includes singletons before interpreting any model score.

Sources

Jurafsky and Martin, Speech and Language Processing, singleton definition and coreference tasks: https://web.stanford.edu/~jurafsky/slp3/24.pdf. Original research, “SPLICE: A Singleton-Enhanced PipeLIne for Coreference REsolution” (2024), author abstract: https://aclanthology.org/2024.lrec-main.1321/. The abstract discusses singleton detection and the absence of singleton spans in OntoNotes. Independent review checked the author chapter’s singleton definition and dataset differences, plus the SPLICE author abstract.

Previous: Single version of the truth Next: Singular extensions