Event Extraction

The foundational step of identifying the word or phrase that most clearly signals an event's occurrence, typically a verb or nominalization.

• Trigger: The lexical anchor evoking the event (e.g., 'fired' in 'The CEO fired the COO').
• Classification: Triggers are categorized into event types like Personnel:End-Position or Conflict:Attack.
• Ambiguity: A single word can trigger different events based on context ('left' can mean Transport or End-Position).
• Non-verbal triggers: Nominalizations like 'resignation' or 'explosion' also serve as event triggers.

The process of identifying the entities involved in an event and assigning them specific semantic roles relative to the event trigger.

• Core Arguments: Essential participants like Agent (who caused the event), Patient (who was affected), and Instrument (what was used).
• Temporal Arguments: Roles specifying Time-within, Time-before, or Time-after the event.
• Location Arguments: Roles like Place or Destination grounding the event spatially.
• Frame-specific roles: A Transaction event requires a Buyer, Seller, and Goods, while an Attack event requires an Attacker and Target.

The task of determining when multiple textual mentions refer to the same real-world event, enabling narrative understanding.

• Cross-document coreference: Linking 'the explosion in Beirut' in one article to 'Tuesday's blast' in another.
• Anaphoric references: Resolving 'this deal' or 'the ceremony' back to a previously introduced event.
• Event hopper construction: Building a timeline of a single event's evolution across a document corpus.
• Challenges: Requires deep semantic understanding to distinguish between identical event types with different participants.

A closely related field that classifies the chronological ordering between events and temporal expressions using standards like TimeML.

• TLINKs: Temporal links establishing relations like BEFORE, AFTER, SIMULTANEOUS, or INCLUDES.
• Dense ordering: Creating a complete or partial timeline of all events in a narrative.
• Signal words: Lexical cues like 'after', 'subsequently', and 'meanwhile' provide strong temporal signals.
• Vagueness: Handling implicit or vague temporal relations where exact ordering is ambiguous.

Identifying the compositional structure of complex events, where a macro-event consists of multiple sub-events.

• Parent-child relations: A 'Business Merger' event may contain sub-events like 'Sign Agreement', 'Regulatory Approval', and 'Integration'.
• Script learning: Inferring prototypical event sequences (scripts) from large text corpora.
• Granularity: Models must handle events at different levels of abstraction simultaneously.
• Causal chains: Sub-events are often linked by causal relations, forming a directed acyclic graph of occurrences.

While both extract structured information, they model fundamentally different aspects of text semantics.

• Event Extraction: Focuses on dynamic occurrences with temporal properties (triggers, arguments, time).
• Relation Extraction: Focuses on static semantic relationships between entities (e.g., EMPLOYEE-OF, LOCATED-IN).
• Overlap: An event's arguments often participate in static relations, making the tasks complementary.
• Joint modeling: Modern systems increasingly perform both tasks simultaneously to leverage mutual context.