Chinese Named Entity Recognition Based On Deep Transition Network

In daily life,people have created a huge amount of text data while using the Internet,which contains great value.An important step in mining the value of massive text data is named entity recognition.Named entity recognition technology(NER)is the recognition of specified entity,mainly including the name of person,location,organization,etc.It is the key step for computer to understand text.Besides,it is also the basic task of many NLP applications such as information extraction,information retrieval,question answering system.However,some deficiencies remain in the current Chinese named entity recognition technology as follow:(1)Current state-of-the-art systems for Chinese named entity recognition tasks are typically based on the family of Recurrent neural Networks.The shallow connections between consecutive hidden states of RNN and insufficient modeling of global information restrict the potential performance of those models.(2)Chinese NER is commonly based on word sequence,which needs the help of Chinese Word Segmentation technique,while ignoring the impact of noise data caused by segmentation errors on the quality of those models.In order to solve the problems above,we proposed Chinese NER based on deep transition network,Lattice DT for brevity.Specifically,the model deepen the state transition path at each position in a sentence,and further assign every token with global representation learned from the entire sentence.Furthermore,the model adopt char-sequence-based method to recognize name entity,introducing word vector,which is based on external corpus.Meanwhile,we integrate latent word vector into character-sequence to alleviate the noise caused by the Chinese word segmentation errors.To prove the effectiveness of the model,we performed extensive experiments with baseline models on different domain datasets including Onto Notes 4,MSRA,Weibo NER and Chinese resume NER dataset.The results show that our model Proposed in this thesis outperforms the state-of-the-art baselines in terms of Fl.