Event time representation, propagation and prediction in temporal information extraction

Abstract

Temporal information extraction is a challenging task due to the inherent ambiguity of language. Event time plays an important role in temporal information extraction, which can help ground events into a timeline and can help other temporal information extraction tasks such as temporal relation extraction. However, explicit information for event time is not often expressed in a document. In this thesis, we first focus on a new event time representation that adopts the 4-tuple temporal representation proposed in the TAC-KBP temporal slot filling to resolve the uncertainty and sparsity problem. We then propose a graph neural network-based method to propagate local time information over constructed event graphs. We also study the event time in temporal relation extraction. We predict relative timestamps for events from event-event relation annotations and use those timestamps as additional features for training a temporal relation extraction system. We use the Stack-Propagation framework to jointly train the timestamps prediction and temporal relation extraction task. Finally, we demonstrate two knowledge extraction systems that have integrated the temporal information extraction models and show their effectiveness.