Named Entity Recognition using Hybrid Neural Network

Abstract

In this thesis, we developed methods to recognize Named Entities (NEs) in the general-domain documents by using Neural Networks. This method consists of two steps. First step is pre-training vectors using Word and Character-level Embedding. Second step is recognizing NEs using Bidirectional Recurrent Neural Network (Bi-RNN) and Convolutional Neural Network (CNN). By using Bi-RNN and CNN sequentially, we extracted important features from both inside and outside of the context window. Also, we built the model with Gated Recurrent Unit (GRU) and compared it to Long Short-Term Memory (LSTM). Experimental results show that hybrid Neural Network increases F1 score much higher than the single Neural Network, and GRU performs slightly better than LSTM. We proved that hybrid method is more effectively than single method. The proposed method can be helpful to recognize the NEs in the general-domain documents.