Exploiting Document-level Temporal Rhythms for Event Temporal Status Identification

Previous research shows that it is a challenging task to determine the temporal statuses of event mentions relative to the document creation time because explicit temporal status cues, such as tense and aspect, are often lacking and an event mention's local context may be ambiguous. To further improve temporal status identification, we exploit the observation that document-level temporal rhythms reflective of story narrative structures exist as sequential patterns among the statuses of event mentions in a document. For example, a news article often starts by introducing the newsworthy event that may overlap with the document creation time, then describes precursory events, and closes by describing future implications. Experiments on the Richer Event Description and TimeBank corpora show that a simple neural network model aware of an event mention's position in a document significantly improves the performance of event temporal status identification. We also demonstrate that exploiting temporal rhythms enables data efficient transfer learning across document domains