Study On Named Entity Recognition For Chinese Specific Domains Based On Deep Learning

The rapid development of informational society has brought us more and more convenience that has also produced a large amount of text information simultaneously.How to analyze and mine the unstructured text information into processable and understandable language form for computer to achieve knowledge acquisition and expression is a core goal of Natural Language Processing(NLP)research and one of the themes of artificial intelligence,which is of great significance for the realization of machine cognitive intelligence.Named entity recognition(NER)as an important part of NLP research,aims at the recognition of language components with specific meanings such as names,locations and organizations in various texts.It is essential for high-level natural language processing technologies like dialogue & interactive systems and automatic question & answering.Previous work mainly focused on the above-mentioned common types of entities,which is relatively mature,while in specific domains,especially in Chinese specific domains,entity recognition is still in development with a lack of studies.Due to the scarcity of annotated corpus,with a large number of terminology and rare words,it is more difficult to model and represent the texts of Chinese specific domains.In view of this,this thesis mainly focuses on the above challenges.Specifically,for the difficulty of modeling and representation caused by data scarcity,this thesis taps the potential of large-scale natural language data,regards itself as naturally annotated data,uses language model to distill knowledge,and constructs new types of network architecture to achieve more effective modeling and tagging.To sum up,the main contributions of this thesis are as follows:Firstly,Bidirectional Long-Short Term Memory(BiLSTM)and Conditional Random Fields(CRF)integrated model called BiLSTM-CRF is introduced in detail,which is the current mainstream method in the named entity recognition field.On this basis,an end-to-end entity recognition method based on Hierarchical BiLSTM-CRF model is proposed from the aspect of network structure optimization and improvement.And experiments are carried out on three Chinese specific domain entity recognition datasets to explore the effect of deep-level network on entity recognition in feature extraction.The results prove that appropriate deep-level network structure is conducive to capturing the better feature representation of semantic level,thus improving the entity recognition effectiveness of the model.Secondly,the principle of deep context language model BERT(Bidirectional Encoder Representations from Transformers)is summarized.It uses super-large-scale natural language text corpus for bidirectional unsupervised pre-training and fine-tuning combined with specific downstream tasks.Compared with the traditional shallow language model,BERT can dynamically adjust the embedding representation of the same word according to the corresponding context,which solves the problem of polysemy well and thus has strong contextual language representation ability.Inspired by this,the thesis proposes an end-to-end entity recognition method based on BERT-CRF model from the improvement of semantic level and validates it on the above datasets.The results prove that when the pre-training data and the explicit language expression in the experimental datasets share great commonality,the recognition effect of the model is improved obviously.On the basis of the above,considering that BiLSTM-CRF model has good sequence tagging performance and has been used as a benchmark model for a long time,an improved end-to-end entity recognition method based on BERT-BiLSTM-CRF model is proposed.Subsequent experiments prove that the fusion effect of the mentioned is remarkable.It not only achieves the best recognition results on the above three Chinese specific domain entity recognition datasets,but also surpasses the previous models on the existing multiple open datasets.