Can a Student Outperform a Teacher? Deep Learning-based Named Entity Recognition using Automatic Labeling of the Global Terrorism Database
Il-Hwan Kim, William M. Pottenger, Vincent Behe · 2018
In this paper, we propose a deep learning-based named entity recognition system (DL-NER) for the counter- terrorism domain. In addition, we propose a training method where an existing rule-based system is used to label terrorism incident datasets such as the Global Terrorism Database (GTD) and the Worldwide Incident Tracking System (WITS), and the resulting labeled datasets are used to train DL-NER. We report improved performance of DL-NER in comparison to the rule-based system that was used to label the training datasets for DL-NER. We also observed better performance of DL-NER than Stanford NER when DL-NER was trained using CoNLL03 and GTD datasets labeled by Stanford NER. Thus we conclude that an existing rule-based NER system may be used to work around the issue of the high knowledge engineering cost of developing sufficient training data for deep learning approaches to NER.