Named entity recognition is a significant subtask of natural language processing. Traditional methods require large numbers of feature engineering and handcrafted additional dictionaries to achieve high performance.
Nowadays, deep learning is widely applied for image and natural language processing domains. Through the deep neural network, features are automatically extracted from the training data, and this avoids most feature engineering. Some deep neural network based methods are also used for the modelling sequence labelled problem. Throughout these methods, the performance of named entity recognition can be improved.
In this thesis, we have mainly used the bidirectional LSTM model based on deep learning as an architecture, word embedding and Convolutional Neural Network (CNN) as a word-level and char-level feature extractor. The conditional random field (CRF) layer is used to output the predicted labels. We used the CoNLL 2003 dataset and the Broad Twitter Corpus dataset training and testing the Bi-LSTM-CNN-CRF model respectively and got satisfactory results. In order to evaluate the Bi-LSTM-CNN-CRF model, we compared it with three popular machine learning methods. The three popular machine learning methods include the Support Vector Machine (SVM), the Hidden Markov Model (HMM), and the Conditional Random Field (CRF). The results of the evaluation show that the Bi-LSTM-CNN-CRF model with only given labelled text and pre-training word embedding surpasses the three machine learning methods that employ handcrafted feature engineering.
