Is mention detection always a separate component?

No. Some systems use an explicit stage; others learn it jointly with linking or coreference.

Does high recall guarantee good resolution?

No. Retaining the right span is necessary for many pipelines, but the system may still choose the wrong entity.

Should every singleton be included?

That depends on the application and annotation scheme. Record the policy before comparing scores.

AI Glossary / Natural language and speech / Intermediate

Mention detection

Definition

Mention detection is the task of identifying the text spans that refer to entities, or that count as mentions under a specified annotation scheme. In coreference systems, it supplies or jointly learns the expressions that may subsequently be grouped as referring to the same entity.

What it means

Before a reading assistant can connect “the supplier” with “it,” it must notice both expressions. Finding only company names is insufficient: descriptions and pronouns may carry essential information. Detection concerns where a mention occurs; resolution concerns which entity it refers to.

The scope is task dependent. A customer database linker may seek named businesses, while a document coreference project may include people, objects, and pronouns. Some datasets record entities mentioned only once; others omit those singletons. These policies change what counts as a correct detection.

An apparent noun phrase is not automatically a usable reference. In “It is raining,” the word “It” does not name an object that the system should merge with an earlier device or shipment.

How it works

A pipeline may propose candidates using noun phrases, names, and pronouns, then score or filter them. Another approach considers text spans up to a chosen width and learns which spans are plausible mentions. Keeping every span is expensive and produces many irrelevant candidates, so pruning reduces the search space.

Joint models can learn mention scores together with coreference decisions. Lee and colleagues' 2017 system illustrates this approach: span representations support both candidate selection and antecedent scoring. A dummy antecedent can represent either a nonmention or a mention with no earlier matching reference, so that decision alone does not cleanly distinguish those cases.

Evaluation must state the boundary rule. An exact-span match requires the same start and end positions; accepting only the same head word measures something different.

Three practical examples

Support: detect “the replacement router” and “it” before checking whether a complaint describes the same device.
Document search: identify “Harbor Tools” as a candidate business reference before looking up its database identifier.
Meeting analysis: identify “our revised schedule” so a later statement about “the plan” can be considered for linking, rather than assumed to match.

Think of it this way

Think of an editor placing removable tabs beside expressions that might need a reference label. Missing a tab means the expression may never reach the linking stage. Adding an unnecessary tab creates extra review work. The analogy is limited: a learned detector scores uncertain boundaries, and joint systems can revise detection through their linking objective.

Worked example

Use this deliberately simple annotation policy: mark referring names, noun phrases, and pronouns, including singletons. The text is “Maya checked the invoice. She approved it.” Its four gold mentions are “Maya,” “the invoice,” “She,” and “it.”

Suppose a detector returns “Maya,” “invoice,” “She,” “it,” and “checked.” With exact boundaries, three predictions match. Precision is 3/5 = 60%; recall is 3/4 = 75%. F1 is 2 × 0.60 × 0.75 / (0.60 + 0.75), approximately 66.7%.

“Invoice” fails the stated full-span rule even though it identifies the head word. Counting it as correct would require a different evaluation. These detection figures do not measure whether “She” was linked to Maya or “it” to the invoice; linking must be tested separately.

Illustrative case study

Fictional case: Bay Desk builds a complaint summarizer that identifies product names accurately but frequently misses descriptions such as “the second unit.” Its first review focuses only on incorrect pronoun links. Inspecting candidate spans reveals that the relevant antecedents were never retained.

The team adds annotated device descriptions, reviews pruning settings, and evaluates both mention recall and final coreference clusters. It also checks that broader candidate coverage does not merge distinct units. This is an illustrative debugging sequence, not a claim of measured performance improvement.

Common misconceptions and mistakes

Do not equate mention detection with named entity recognition. Names are only one possible mention type. Do not treat a first mention as an error merely because it lacks an antecedent. Conversely, not every occurrence of “it” refers to an entity.

Avoid evaluating a production detector only on premarked gold mentions: that removes the boundary-finding challenge. Also avoid transferring annotation rules silently between datasets. A singleton omitted from one benchmark may be essential for a search application.

Business use

For a business workflow, define the entities that matter and retain offsets into the original text. Review boundary errors, missed descriptions, and false pronoun detections separately. Test representative document formats, including tables and messages where sentence segmentation may be unreliable.

Measure the downstream consequence of omissions: a missed device description may remove evidence from a summary even when the final prose sounds fluent. Keep the original passage accessible so a reviewer can inspect the reference rather than trusting an isolated label.

Check your understanding

A detector finds both “the order” and “it” but links “it” to an unrelated delivery. Is that necessarily a detection error?

Answer: No. If both boundaries are correct under the chosen scheme, the failure is in resolution. Detection and linking errors can coexist, but they should not be conflated.

Related terms

Learn next

annotate a short message with explicit span boundaries, then assign entity identifiers in a second pass. Compare the two kinds of disagreement. Follow with mention-ranking models to see how a system chooses among earlier references and the option of no antecedent.

Antecedent

Sources

Sources: Jurafsky and Martin, Speech and Language Processing, Chapter 24, mention detection and annotation scope: https://web.stanford.edu/~jurafsky/slp3/24.pdf . Lee et al. (2017), End-to-end Neural Coreference Resolution, span proposal and joint scoring: https://aclanthology.org/D17-1018.pdf . The worked detection scores and fictional case are original illustrations.

Previous: Mendelian randomization Next: Mention-pair