Temporal information extraction encompasses the identification, normalisation, and reasoning about temporal information in text. The task involves three core subtasks: recognising temporal expressions (dates, times, durations, and frequencies), normalising them to standard calendar values, and determining the temporal relations between events and between events and times. Understanding temporal structure is essential for applications such as timeline construction, question answering about when events occurred, clinical narrative processing, and historical text analysis.
Temporal Expression Recognition and Normalisation
<TIMEX3 type="DATE" value="2024-03-15">March 15, 2024</TIMEX3>
EVENT: event-denoting expressions
<EVENT class="OCCURRENCE">acquired</EVENT>
TLINK: temporal relations between events/times
TLINK(e₁, e₂, BEFORE) — e₁ occurs before e₂
Temporal expression recognition identifies spans of text that denote temporal references, while normalisation maps these spans to standard representations. The expression "last Tuesday" must be resolved relative to the document creation time, "the 1990s" must be mapped to a specific decade, and "three days after the surgery" requires anchoring to the time of the surgery event. The TimeML annotation language and its associated TIMEX3 tag define a standard representation for temporal expressions, classifying them as dates, times, durations, or sets and assigning ISO 8601 values. Systems such as SUTime and HeidelTime achieve F1 scores above 85% for temporal expression extraction and above 80% for value normalisation.
Temporal Relation Classification
Temporal relation classification determines how events and times are ordered relative to one another. The TimeML framework defines 13 temporal relation types, including BEFORE, AFTER, SIMULTANEOUS, INCLUDES, IS_INCLUDED, BEGINS, ENDS, and their inverses. Classifying temporal relations is challenging because the information may be implicit (requiring world knowledge or pragmatic reasoning) and because the space of possible relations between all event-time pairs in a document is quadratic. The TempEval shared tasks (2007, 2010, 2013) established benchmarks for temporal relation classification, though inter-annotator agreement for this task is notably lower than for most NLP tasks, reflecting the genuine ambiguity of temporal reasoning in natural language.
Temporal information extraction is particularly critical in the clinical domain, where understanding the temporal relationships between symptoms, diagnoses, treatments, and outcomes is essential for patient care. The i2b2 clinical NLP challenges and the Clinical TempEval shared tasks have driven research on extracting temporal information from clinical narratives, including admission notes, discharge summaries, and radiology reports. Clinical temporal reasoning must handle domain-specific expressions ("postoperative day 2," "q.i.d.") and complex event-time relations that are often left implicit in clinical writing.
Timeline construction synthesises temporal information extraction results into coherent chronological orderings of events. Given a document or a collection of documents about an entity or topic, timeline construction produces an ordered sequence of events with their associated times. This task requires not only accurate temporal extraction but also event coreference resolution (determining when different mentions refer to the same event) and temporal reasoning (inferring relations not explicitly stated using transitivity and other temporal logic rules). Allen's interval algebra provides the formal framework for temporal reasoning, defining 13 possible relations between two time intervals that are mutually exclusive and jointly exhaustive.