Multilingual Named Entity Recognition using Hybrid Neural Networks

Named entity recognition is a significant subtask of information extraction. Most of the high performing NER systems model the task as a sequence labelling or a structured prediction problem (Lafferty et al., 2001; Finkel et al., 2005; Ratinov and Roth, 2009). In recent years, neural network based models obtain remarkable success in a wide range of natural language processing tasks, which outperforms the classical models in terms of accuracy (Collobert et al., 2011; Graves, 2012). The aim of this study is to systematically evaluate the performances of different neural networks for NER in various languages. We also investigate the impact of additional features and configurations. Three baseline models, including a feedforward network, a standard Bi-LSTM and a window-based Bi-LSTM are extensively tested with different feature and hyper-parameter settings on three data sets. The experimental results indicate that the neural network based models are generally robust and capable of achieving reasonable accuracies across different languages. In addition, the window-based Bi-LSTM is more robust than the standard Bi-LSTM when less information is available for the task. The effectiveness of the features depends on both the architecture of the model and the data set, except that all the models benefit greatly from the pre-trained word embeddings and the Conditional Random Fields (CRF) based interface. Overall, our best performing models are competitive when compared to the state-of-the-art NER systems.