[2L1] [General Session] 9. NLP / IR

[2L1-02] Convolutional Neural Network for Event Knowledge Extraction

〇Kai Ishikawa¹, Hiroya Takamura², Manabu Okumura² (1. NEC Corporation, 2. Tokyo Institute of Technology)

Keywords:Knowledge Base Population, event detection, Convolutional Neural Network

This paper describes a knowledge extraction system using a convolutional neural network applied to a shared task of NIST TAC KBP Event Nugget Detection. Our system based on conventional document classification system using a convolutional neural network. We found two major problems when we apply the method to knowledge extraction task: 1) Noise and context trade-off problem by using single fixed window to generate input token sequence. 2) Difficulty of handling some multiword expressions separately appeared in the context with convolution process focusing on neighboring words. We proposed two methods corresponding to the problems mentioned. As a result of evaluation using NIST TAC KBP2016 official evaluation dataset and tool kit, we found that our proposed method using both MW method and PME method achieved 35.85% in F1 score that outperforms the best performance of the task participants in KBP2016.