Deep Learning With Minimal Training Data: TurkuNLP Entry in the BioNLP Shared Task 2016

Abstract

<p>We present the TurkuNLP entry to the BioNLP Shared Task 2016 Bacteria Biotopes event extraction (BB3-event) subtask. We propose a deep learning-based approach to event extraction using a combination of several Long Short-Term Memory (LSTM) networks over syntactic dependency graphs. Features for the proposed neural network are generated based on the shortest path connecting the two candidate entities in the dependency graph. We further detail how this network can be efficiently trained to have good generalization performance even when only a very limited number of training examples are available and part-of-speech (POS) and dependency type feature representations must be learned from scratch. Our method ranked second among the entries to the shared task, achieving an F-score of 52.1% with 62.3% precision and 44.8% recall<br /></p>