Recovering event actions in privacy policies: a multi-layered approach to structured event extraction

Privacy policies describe data practices using verbal predicates, nominal expressions, and light-verb constructions, complicating structured event extraction. We present a multi-layer system that extends verbal predicate analysis with a nominal-event recovery module to extract structured event frames from both verbal and nominal expressions in privacy policies. Evaluated on 150 manually annotated privacy policy sentences, the full system achieves an action F1 score of 0.8641 and a field-level micro-F1 score of 0.6592. Disabling the nominal-event recovery module reduces action F1 to 0.8492. Enabling this module yields a net gain of 16 true positives at the cost of three additional false positives. Light-verb recovery has a small effect on overall action F1. These results indicate that nominal-event recovery improves recall and extends coverage beyond verbal predicates, while highlighting limitations in participant extraction and event linking. The extracted event frames provide an intermediate representation intended to support subsequent policy formalization and compliance verification, but their suitability for this task remains to be evaluated.