Research On Named Entity Recognition With Deep Learning

Named Entity Recoginition(NER)is a quite important basic task in Natural Language Processing(NLP)and a basic technology for many high-level applications of NLP.Traditional methods for solving NER are based on rules and statistics.The rule-based methods need strong linguistic knowledge and lack universality.The statistical-based methods get rid of the dependence on linguistic knowledge,but they still require a lot of handcrafted features.Deep Learning which can learn features by itself requires neither strong linguistic knowledge nor a lagre number of handcrafted features.It has been widely used in various fields of NLP in recent years.The Deep Learning-based NER models has surpassed traditional methods without too many additional features.There is still much space for research and exploration on NER with Deep Learning.After in-depth study on NER with Deep Learning,we found that there are two deficiencies:1)There is little research on the improvement of the Deep Learning-based NER models by introducing syntactic information.2)Deep Learning-based NER models are mainly on sentence level for now,which cannot solve the tagging inconsistency problem.For the deficiencies above,we firstly analyzed several key models and found some key factors that affected the sentence-level NER models through multiple sets of comparative experiments.Then we tried two methods for introducing syntactic information: linear encoding for Constituent Parsing Tree and syntactic Graph Convolutional Networks(GCNs)for Dependency Parsing Tree.Moreover,we improved the syntactic GCNs by synthesizing two kinds of syntactic GCNs and achieved certain improvement.Finally we proposed a model combining syntactic GCNs with document-level attention to slove the tagging inconsistency problem.This model,which had better performance and universality,achived 91.00% and 92.36% F1 score on the test dataset of CHEMDNER and Bio Creative V CDR corpus respectively without the domain dictionary.