Event Coreference

The fundamental distinction in event coreference is between a specific event instance and a generic event class.

• Event Instance: A singular occurrence anchored in time and space. Example: 'The assassination of Archduke Franz Ferdinand on June 28, 1914'.
• Event Class: A category of events sharing common properties. Example: 'Political assassinations' or 'Diplomatic visits'.

Systems must avoid erroneously linking a mention of a generic class to a specific historical instance. This requires fine-grained event typing and temporal reasoning to distinguish 'the protest' (specific) from 'a protest' (generic).

Event identity is intrinsically tied to its spatiotemporal coordinates. Two descriptions are coreferent only if they occupy the same temporal interval and geographic location.

Key mechanisms for anchoring:

• Timex Normalization: Parsing relative dates ('last Tuesday') into absolute ISO 8601 timestamps.
• Geoparsing: Resolving place names to precise latitude/longitude pairs using a gazetteer.
• Event Hopper Detection: Identifying when a narrative jumps forward or backward in time, preventing false coreference links across distinct occurrences of the same event type.

A system must recognize that 'the explosion at dawn' and 'the explosion at dusk' are distinct events despite sharing the same location and event type.

Event coreference operates at two distinct scopes, each with unique challenges:

Within-Document Coreference

• Resolves mentions in a single news article or narrative.
• Relies heavily on discourse structure, lexical cohesion, and pronominal references ('it happened', 'the incident').
• Example: Linking 'the merger was announced' with 'the deal closed on Friday' in the same press release.

Cross-Document Coreference

• Clusters mentions of the same event across an entire corpus.
• Requires robust entity linking to anchor participants and locations.
• Example: Aggregating all news reports from different outlets about the same natural disaster into a single event cluster for a knowledge base population task.

Coreferent event mentions must exhibit consistent participant structures. The same real-world entities must fill the same semantic roles across mentions.

Critical role analysis includes:

• Agent/Perpetrator: The entity causing the event.
• Patient/Theme: The entity undergoing the event.
• Instrument: The tool or method used.
• Location & Time: The spatiotemporal anchors.

Example: 'Police arrested the suspect' and 'The suspect was apprehended by authorities' are coreferent because the semantic roles are preserved despite syntactic variation (active/passive voice). A mismatch in participants, such as a different agent, signals a distinct event.

Coreferent event mentions rarely use identical wording. Systems must bridge significant lexical gaps and pragmatic inference.

Common variation patterns:

• Synonymy & Hypernymy: 'The acquisition' vs. 'The takeover' vs. 'The deal'.
• Metonymy: Referring to an event by its location ('Wall Street crashed') or date ('9/11 changed everything').
• Nominalization: 'They invaded' (verb) vs. 'The invasion' (noun).
• Implicit Reference: 'The verdict was read' implicitly references the preceding trial event.

Resolving these requires deep contextual embeddings and commonsense knowledge about event sequences and causal relationships.

Event coreference systems are evaluated using metrics adapted from entity coreference, each with distinct biases:

• MUC (Message Understanding Conference): Link-based; penalizes missing coreference links. Favors systems that over-merge.
• B³ (Bagga & Baldwin): Mention-based; computes precision and recall per mention. Balances over-merging and under-merging.
• CEAF (Constrained Entity-Alignment F-Measure): Entity-based; finds optimal one-to-one alignment between gold and predicted clusters. Penalizes large erroneous clusters.
• BLANC (BiLateral Assessment of Noun-phrase Coreference): Link-based; evaluates both coreference and non-coreference links to reward balanced decisions.

The CoNLL-2012 average (arithmetic mean of MUC, B³, and CEAF F1 scores) is the standard aggregate metric for shared tasks like OntoNotes.

The precursor task that identifies event triggers (words that signal an event occurrence) and classifies them into types like 'Attack,' 'Acquisition,' or 'Birth.' Systems like ACE (Automatic Content Extraction) define event schemas with specific argument roles. Event coreference then clusters these extracted event mentions.

• Trigger identification: finding 'fired' or 'resigned'
• Argument role labeling: who did what to whom
• Schema-based vs. open-domain extraction approaches

Identifies and classifies the temporal links between events in text—determining if one event happened before, after, or simultaneously with another. This is tightly coupled with event coreference because two mentions cannot refer to the same event if they occupy incompatible temporal positions.

• Uses TimeML annotation standards
• Resolves relative expressions ('last Tuesday')
• Provides hard constraints for coreference clustering

The automated process of extracting facts from unstructured text and inserting them into a structured knowledge base. Event coreference feeds KBP by ensuring that multiple reports about the same real-world occurrence are merged into a single canonical event node rather than creating duplicate entries.

• Cold start: building KBs from scratch
• Incremental: updating existing KBs with new facts
• Relies on accurate entity and event linking

Connects textual entity mentions to their unique entries in a knowledge base like Wikidata. Event coreference often depends on entity linking as a feature—knowing that 'Paris' refers to the city, not the person, helps disambiguate which event is being discussed. Collective entity linking jointly resolves all mentions for coherence.

• Candidate generation via surface form dictionaries
• Contextual disambiguation using neural rankers
• Provides semantic grounding for event arguments