A Deep Neural Network Model for the Task of Named Entity Recognition

원문 초록

One of the most important factors which directly and significantly affects the quality of the neural sequence labeling is the selection and encoding the input features to generate rich semantic and grammatical representation vectors. In this paper, we propose a deep neural network model to address a particular task of sequence labeling problem, the task of Named Entity Recognition (NER). The model consists of three sub-networks to fully exploit character-level and capitalization features as well as word-level contextual representation. To show the ability of our model to generalize to different languages, we evaluated the model in Russian, Vietnamese, English and Chinese and obtained state-of-the-art performances: 91.10%, 94.43%, 91.22%, 92.95% of F-Measure on Gareev's dataset, VLSP-2016, CoNLL-2003 and MSRA datasets, respectively. Besides that, our model also obtained a good performance (about 70% of F1) with using only 100 samples for training and development sets.