Relation, Event, and Temporal Extraction › Key Parameters

The key parameters for this extraction pipeline are:

• TRIGGERS maps the selected lemmas to relation types. The loop looks up each lowercased token lemma without a part-of-speech filter.
• PRONOUNS selects which supplied strings the toy resolver attempts to resolve. Membership does not establish anaphora.
• ROLE_PATTERN: The regular expression for possessive-appositive role constructions. Controls which surface patterns yield chief_executive_of relations.
• MULTIPLIER: The scale table for money mentions. Controls how numeric amounts are converted to base units.
• DCT: The document creation time. Supplies the reference date for this example's deictic expression; quoted speech and other temporal perspectives can require their own reference times.
• SCHEMA: The requested relation and event type inventory with roles. Supplies the prompt's allowed labels; explicit validation or constrained decoding is needed to enforce them.

Adding entries to TRIGGERS can recover new relations and introduce false positives; whether recall increases depends on the added entries and data. The loop uses the acquisition branch for acquire and sends every other mapped label to the founding branch. Add an extraction branch if a new label needs different argument roles. Adding to PRONOUNS lets the resolver search for an antecedent when a supplied head matches a newly added entry. The heuristic can still choose the wrong antecedent. ROLE_PATTERN defines which surface constructions match. The MULTIPLIER table needs the scales used by the supported money expressions. Incorrect reference dates can change deictic normalization, and schema versions should be recorded when aggregating outputs produced under different inventories.