Relation, Event, and Temporal Extraction › Summary

• Relations link entities into typed tuples. The schema's direction, symmetry, multiplicity, and inventory determine the represented statements. Closed-schema and open extraction make different choices about allowed predicates and later normalization.
• Distant supervision uses knowledge-base facts to label relation tuples whose features are aggregated across co-mentions or to project labels onto sentences. When labels are projected onto sentences, a sentence can receive a relation label it does not express. In the multi-instance learning setup above, the bag label is observed. Which sentences express the relation remains latent. Average pooling and selective attention illustrate two aggregation choices.
• Neural relation-classification methods expose candidate arguments in different ways, including position embeddings, piecewise max pooling, and entity markers. A question-answering formulation expresses a relation as a slot-filling question. Assess generalization for the model, templates, and data used in the evaluation. The number of queries depends on how the candidates and templates are enumerated.
• Event records represent described situations with triggers, role-filled arguments, and attributes. Trigger ambiguity involves word-sense and schema decisions. Words that name arguments can help identify the intended type. Joint models can also use predicted trigger and role information together. The JRNN comparison above reports lower argument-role F1 than trigger-classification F1. Nested event-valued arguments and references such as "the deal" are distinct schema and coreference decisions.
• TimeML/TIMEX3 uses ISO 8601 conventions and extensions; ISO-TimeML is standardized in ISO 24617-1. Anchoring distinguishes reference requirements: fully specified absolute dates need no reference date, deictic expressions use a temporal perspective such as document creation time, and relative expressions use a contextually supplied reference time. Reichenbach's three times explain the past-perfect ordering in the announcement example.
• Allen's interval algebra gives thirteen relations with exact endpoint conditions for positive-duration intervals and represents their extent explicitly. Representing each event by only its start point loses distinctions involving interval ends. General satisfiability is NP-complete. Path consistency is incomplete in the unrestricted algebra; small instances and tractable restricted classes can also be checked exactly.
• Document context is needed when the evidence or reference lies beyond a sentence. Cross-sentence arguments, coreference, entity linking, and context selection affect the design. Exhaustively enumerating ordered pairs of distinct mentions takes $O(n^2)$ candidates for $n$ mentions.
• Evaluation should include per-class scores and a simple baseline for the stated prediction unit. A majority-class baseline fits a declared single-label task. Consider label noise and shifts in the evaluation distribution. Mention-level and entity-level scores use different prediction units; a comparison must state both protocols.
• For LLM-based extraction, validate the schema, retain copied evidence, and recompute normalized values when the derivation is explicit. Substring matching can detect selected strings absent from the document. Source support for the relation or event needs a separate judgment. Recomputing an announcement date does not supply an unknown acquisition date.

Entity Linking and Knowledge Graph Construction shows how canonical identifiers let us join evidence across documents even when the same entity appears under different names.