On Event Detection in Scientific Papers: A Multi-Domain Dataset

Anonymous

Abstract: Given the growing number of scientific papers, automatic information extraction in scientific documents is important for efficient knowledge update and discovery. A key component in scientific papers involves rhetorical activities/events to convey new knowledge and convince readers of the correctness. This work explores a new information extraction problem for scientific documents, aiming to identify event trigger words of rhetorical events/activities, i.e., event detection (ED). To promote future research in this area, we present SciEvent, the first and new dataset for event detection in scientific documents. SciEvent annotates scientific papers of four different domains (i.e., computer science, biology, physics, and mathematics) using 8 popular event types. Our experiments on SciEvent demonstrate the challenges of scientific ED for existing models and call for further research effort in this area. We will publicly release SciEvent to facilitate future research.