Are event names temporal expressions?

An event name such as “the launch” may provide an anchor in context, but event mentions and explicit time expressions are distinct annotation categories.

Can “soon” be normalized to a date?

Not reliably without additional conventions or context. Its vagueness should be preserved.

Is a time of day enough to identify an instant?

Usually not. A date and relevant time-zone information may also be required, depending on the application.

AI Glossary / Natural language and speech / Intermediate

Temporal expressions

Definition

Temporal expressions are words, numbers or phrases that refer to time. They can describe dates, times of day, durations, intervals or recurring schedules, such as “14 October 2026,” “at noon,” “for three hours” and “every month.”

What it means

Some expressions identify time directly; others depend on context. “Tomorrow” needs a reference day, while “three weeks later” needs an earlier time or event. “In March” leaves the year unresolved unless surrounding information supplies it.

A temporal expression is not necessarily an event. A date can be mentioned without saying anything happened on it, and an event can be described without an explicit date. An application needs to connect the right expression to the right event rather than treating every nearby date as interchangeable.

The distinction matters for business records. A request created today may quote a message written last week about a meeting next month. The reader's current date is not automatically the reference point for every relative phrase.

How it works

A temporal-expression processor identifies the relevant span of text and classifies its type. It then attempts to normalize the meaning into a structured value, interval, duration or recurrence representation.

In TimeML's TIMEX3 scheme, the principal categories are DATE, TIME, DURATION and SET. A context-dependent value can be associated with an anchor. The scheme also permits modifiers for expressions that are approximate or bounded.

Normalization needs enough evidence. Locale can change how a numerical date is read, and the relevant calendar or time zone may need to be specified. Recurrence requires a rule rather than one timestamp.

The system should retain both the original phrase and the interpreted value. When context is missing, it can preserve partial information or request clarification instead of guessing. Detecting the correct words and resolving their meaning should be evaluated separately.

Three practical examples

A scheduling assistant interprets “tomorrow afternoon” using the conversation's established date and relevant local context.
An archive search recognizes “during 2019” as a year-level interval rather than a precise instant.
A maintenance planner distinguishes “every three months” from “three months from now,” because one is recurring and the other is relative to an anchor.

Think of it this way

Think of directions on a calendar. Some name a square directly; others say to move a certain distance from a known square or repeat a pattern across many squares.

The analogy is incomplete because natural language can express vague periods, multiple calendars and uncertain boundaries. A phrase such as “around the end of the year” should not be forced into one exact day merely because a calendar has day boxes.

Worked example

Resolving relative dates without inventing precision

Use a fictional message with an explicitly stated reference date of 2026-10-14. It says: “The review is tomorrow. The trial lasts three days. We will repeat the review every two weeks.”

Under the stated reference, tomorrow resolves to 2026-10-15. Three days describes a duration; it does not establish the trial's start date because the message does not say when the trial begins. Every two weeks describes recurrence, not a single future date.

If the message additionally states that reviews recur every 14 calendar days beginning on 2026-10-15, the next date is 2026-10-29, and the following one is 2026-11-12. Those dates follow from the added schedule rule, not from the duration of the trial.

Now consider “03/04/2026.” Without an established date convention, it can mean 3 April or March 4. A reliable output records that ambiguity rather than silently selecting one. All dates here are teaching inputs, not references to the reader's current day.

Illustrative case study

A fictional training provider imports registration emails into a planning system. It initially resolves every “next week” against the date when the import runs. Messages written months earlier are consequently assigned to the wrong period.

The team changes the process to retain message dates and quoted-message boundaries. It marks unresolved references for review and separates recurring sessions from one-off bookings.

A later test includes “for two hours” without a start time. The system now records the duration without creating an unsupported appointment. The practical improvement is preserving the information actually present, not pretending that every phrase supplies a complete calendar event.

Common misconceptions and mistakes

Treating all temporal expressions as exact timestamps. Durations, recurrences and partial dates require different structures.
Resolving “today” against processing time by default. The correct anchor may be the author's or quoted speaker's date.
Assuming a month always equals a fixed number of days. Calendar arithmetic depends on the dates and convention.
Confusing “in three days” with “for three days.” The first can locate a future time; the second normally expresses duration.
Dropping approximation words such as “about.” Doing so introduces unsupported precision.

Business use

Temporal-expression handling is useful in scheduling, document search, service records and timeline preparation. The practical requirement is often to make a proposed interpretation reviewable before it affects a workflow.

Specify the default date convention and time-zone handling, and identify which source supplies the reference time. Keep a visible distinction between extracted text and confirmed structured values. Test messages containing quotations, missing years, approximate dates and recurring schedules.

Measure normalization errors separately from failure to detect an expression. A system can locate “next Friday” perfectly while choosing the wrong week. Similarly, one correct date should not compensate for an incorrect duration elsewhere in the same message when evaluating a scheduling task.

Check your understanding

A note says “The workshop lasts two days” but gives no date. Should an extractor schedule it for today and tomorrow?

Answer: No. The phrase establishes a duration, not a start date. Scheduling would require an additional supported anchor.

Related terms

Temporal annotation
Duration
Temporal anchor
Event extraction
date normalisation
recurrence
calendar arithmetic

Learn next

Read temporal annotation and date normalization. Sort ten time phrases into dates, times, durations and recurrences, then mark exactly which extra context each phrase needs for a reliable interpretation.

Temporal annotation

Sources

TimeML 1.2.1, TIMEX3 specification: https://timeml.github.io/timeMLdocs/TimeML_1.2.1a.html. Saurí and Pustejovsky, temporal-expression annotation guidelines in the LDC Spanish TimeBank documentation: https://catalog.ldc.upenn.edu/docs/LDC2012T12/annotationGuidelines_timex3_SP.pdf. The cited versioned materials support expression types and anchoring. English examples and calendar calculations are original teaching illustrations.

Previous: Temporal-difference learning with gradient correction Next: Temporal logic