Using query patterns to learn the duration of events

Abstract

We present the first approach to learning the durations of events without annotated training data, employing web query patterns to infer duration distributions. For example, we learn that “war” lasts years or decades, while “look ” lasts seconds or minutes. Learning aspectual information is an important goal for computational semantics and duration information may help enable rich document understanding. We first describe and improve a supervised baseline that relies on event duration annotations. We then show how web queries for linguistic patterns can help learn the duration of events without labeled data, producing fine-grained duration judgments that surpass the supervised system. We evaluate on the TimeBank duration corpus, and also investigate how an event’s participants (arguments) effect its duration using a corpus collected through Amazon’s Mechanical Turk. We make available a new database of events and their duration distributions for use in research involving the temporal and aspectual properties of events. 1