Event Argument Identification on Dependency Graphs with Bidirectional LSTMs


Event Extraction is a difficult information extraction task. In the ACE 2005 scheme it consists of two subtasks: (1) finding the words which most clearly express an event on the lexical surface (trigger detection) and (2) identifying the roles entities and attributes play in the event (argument identification).

My talk will consist of two parts.

The first is a dry run for IJCNLP. I will present work on event argument identification. Our key findings are that argument identification performance varies greatly across argument types. The variance cannot be explained by the amount of training data available. We identify distance as a more important factor -- the farther away an argument is to its trigger, the more difficult it is to identify. We propose a neural system which operates on shortest dependency paths to increase argument identification performance.

In the second part I extend the argument identification system to a full event extractor operating on dependency trees. This is work in progress.