Is temporal annotation the same as adding file timestamps?

No. File metadata describes the document, while temporal annotation examines time information expressed within it.

Must annotation be manual?

No. It can be manual, automatic or a reviewed combination. The quality requirements still depend on the intended use.

Does BEFORE imply causation?

No. An event preceding another does not establish that it caused the later event.

AI Glossary / Natural language and speech / Intermediate

Temporal annotation

Definition

Temporal annotation is the addition of structured labels describing time-related information in language. Depending on the annotation scheme, it can identify temporal expressions, events, temporal signals and relationships such as one event occurring before another.

What it means

A document may mention when something happened without placing its sentences in chronological order. It may also describe a plan, a canceled event or a duration whose starting point appears elsewhere. Temporal annotation makes these distinctions available for systematic analysis.

Finding a date is only one part of the task. “The inspection followed the repair” expresses an ordering even though no calendar date appears. “The repair may begin tomorrow” combines a time expression with uncertainty about an event.

An annotation represents what the text supports under stated conventions. It should not silently turn a prediction into a completed event or fill in an exact date that the document never establishes.

How it works

A project specifies the annotation units and relation inventory. Annotators or automated tools mark text spans, assign identifiers and connect related items. A scheme can separate event mentions from particular event instances and their attributes.

TimeML provides tags for events and temporal expressions, with temporal links connecting events and times. Its TIMEX3 categories include dates, times, durations and recurring sets. A reference time can help interpret expressions such as “yesterday.”

Normalising a time expression and linking an event to it are distinct decisions. So are identifying a relation and deciding whether it is explicitly stated or inferred. Good guidelines describe how to handle uncertainty and missing information.

Quality checks cover span boundaries, attribute values, link endpoints and consistency. Independent annotation and adjudication reveal where the instructions or source text permit disagreement. Software-valid markup alone does not establish a correct temporal interpretation.

Three practical examples

An incident-review project annotates reports so investigators can compare the order in which problems were noticed and repairs were attempted.
A news-research tool separates a document's publication date from the dates of the events it describes.
A service-history application distinguishes a planned maintenance visit from a visit reported as completed.

Think of it this way

Think of adding labeled pins and connections to a timeline. A pin may have a precise date, an approximate interval or only a relationship to another pin.

The analogy has a limit: not every event belongs at a known point, and some events in language are hypothetical. A useful annotation can preserve those gaps rather than forcing every mention into a definite chronological slot.

Worked example

Anchoring and ordering a short report

Imagine a report dated 2026-10-14 that says: “The pump failed yesterday. The crew completed the repair today, after the inspection.” Assume these relative expressions refer to the report's stated date.

Annotate “yesterday” as the date 2026-10-13 and “today” as 2026-10-14. Link the failure to the first date and the completed repair to the second. The phrase “after the inspection” supports inspection BEFORE repair.

The text does not explicitly supply the inspection's calendar date. It also does not establish whether the inspection happened before or after the failure. Both an earlier inspection and an inspection between failure and repair can fit the words given.

Consequently, the annotation should keep that ordering unresolved unless the project's inference policy and additional evidence justify it. There are three event mentions but only two date expressions in this example. Counting date strings would miss the undated inspection and its relation to repair.

Illustrative case study

A fictional equipment-rental company builds timelines from free-text service notes. Its initial system attaches every event to the note's upload date. Older handwritten notes entered in bulk then appear to describe a burst of simultaneous repairs.

The team separates upload metadata from dates stated in the narrative and records the anchor used for relative expressions. It also marks “scheduled” and “cancelled” visits distinctly from completed work.

Reviewers can inspect the supporting phrase for each timeline item. The company retains unresolved entries for clarification instead of manufacturing a complete chronology. The case shows why provenance and uncertainty are essential parts of a useful temporal record.

Common misconceptions and mistakes

Treating document order as event order. Narratives can move backwards and forwards in time.
Equating creation, publication and event time. They describe different things.
Assuming every event has an explicit date. Relative ordering may be the only available evidence.
Turning future or negated events into completed facts. Temporal structure must preserve relevant modality and polarity.
Interpreting a missing link as a negative relation. The relation may simply be unknown or outside the annotation scope.

Business use

Temporal annotation supports event search, timeline construction, document comparison and training datasets for time-aware language systems. Its usefulness depends on a clear link between the annotation scheme and the business question.

If the question is whether work preceded a delivery, collecting every visible date without event relationships will be insufficient. Conversely, a project needing only date extraction may not require a complete event graph.

Choose a documented scheme and record any local extensions. Evaluate relative-date anchoring separately from span detection and relation classification. Preserve original text and source metadata so an incorrect normalized date can be traced to its cause. Review conflicting annotations before relying on an automatically assembled sequence.

Check your understanding

A note says “The delivery occurred after the inspection,” but neither event has a date. Is there useful temporal information to annotate?

Answer: Yes. Inspection BEFORE delivery is a supported ordering. Exact dates remain unspecified.

Related terms

Temporal expressions
Event extraction
Information extraction
Temporal expression
TimeML
temporal relation
document creation time
annotation guideline

Learn next

Read temporal expression and event extraction. Annotate a short report with separate columns for source phrase, event status, time anchor and supported relations. Leave genuinely unresolved relationships blank or explicitly unknown.

Event extraction

Sources

Pustejovsky and colleagues, TimeBank 1.2 documentation, April 2006: https://timeml.github.io/site/timebank/documentation-1.2.html. TimeML 1.2.1 specification, October 2005: https://timeml.github.io/timeMLdocs/TimeML_1.2.1a.html. These versioned primary references explain the annotation components; no claim about the latest ISO standard is made. The pump report and rental-company case are invented.

Previous: Temporal anchor Next: Temporal Bias