Temporal Relation Extraction

The core mechanism of TRE involves classifying the temporal link between pairs of events or an event and a time expression. This is not merely a binary classification; it requires mapping to a defined set of relations.

• Allen's Interval Algebra provides the foundational 13 relations (e.g., before, meets, overlaps, during).
• TimeML standard simplifies this for text, using links like TLINK (Temporal Link), SLINK (Subordination Link), and ALINK (Aspectual Link).
• A model must distinguish between strict ordering (X before Y) and vague containment (X happened sometime last week).

To build a timeline, events must be anchored to a fixed point. The Document Creation Time (DCT) serves as the origin point for the narrative's timeline.

• Explicit anchoring: 'The meeting happened on Tuesday.'
• Relative anchoring: 'The crash occurred three days after the launch.'
• Vague anchoring: 'The company previously announced layoffs.'

TRE systems must resolve these anchors to calculate the temporal distance between events and the DCT, often using temporal normalization to map expressions like 'last quarter' to concrete date ranges.

Temporal relations are often explicitly signaled by specific lexical triggers. High-precision TRE systems rely heavily on identifying these temporal signals.

• Prepositions: 'The stock crashed after the CEO resigned.'
• Subordinating Conjunctions: 'While the system was rebooting, the attack occurred.'
• Discourse Markers: 'Subsequently, the board voted on the merger.'
• Tense & Aspect: The grammatical structure itself signals ordering ('had eaten' vs. 'ate').

Ignoring these signals forces the model to rely on vague statistical correlations, drastically reducing accuracy.

TRE is not just about isolated pairs; the goal is to construct a globally coherent temporal graph from a document. This requires transitive closure reasoning.

• If Event A is before Event B, and Event B is before Event C, the system must infer A is before C.
• Integer Linear Programming (ILP) is often used as a post-processing step to enforce global consistency, resolving conflicts where a local classifier might predict A before B and B before A simultaneously.
• The output is a directed acyclic graph (DAG) representing the chronological flow of the narrative.

The progression of TRE technology is benchmarked against standardized datasets that define the annotation schema.

• TimeBank 1.2: The foundational corpus annotated with TimeML, containing news articles with gold-standard temporal links.
• TempEval-3: A shared task that standardized evaluation across multiple languages and domains, focusing on extracting events, time expressions, and the relations between them.
• MATRES: A more recent benchmark focusing on multi-axis temporal relations, requiring models to understand the start and end points of events, not just vague ordering.

TRE provides critical value in high-stakes domains where the sequence of events determines liability or diagnosis.

• Clinical Texts: Extracting the timeline of symptoms, medications, and procedures from electronic health records to understand disease progression. 'Patient took aspirin before the onset of chest pain.'
• Legal Discovery: Analyzing deposition transcripts to identify contradictions in witness timelines or to establish a definitive chain of events leading to an incident.
• Financial Compliance: Monitoring earnings call transcripts to detect if forward-looking statements were made before or after material non-public information was disclosed.