Improving Temporal Relation Extraction with Training Instance Augmentation

Abstract

Temporal relation extraction is important for understanding the ordering of events in narrative text. We describe a method for increasing the number of high-quality training instances available to a temporal relation extraction task, with an adaptation to different annotation styles in the clinical domain by taking advantage of the Unified Medical Language System (UMLS). This method notably improves clinical temporal relation extraction, works beyond featurizing or duplicating the same information, can generalize between-argument signals in a more effective and robust fashion. We also report a new state-of-the-art result, which is a two point improvement over the best Clinical TempEval 2016 system.