Named Entity Recognition On Global Search

Named Entity Recognition(NER)is a basic task in natural language processing,it targets at recognizing some meaningful words or phrases from unprocessed text.NER is the antecedent task of many down-stream tasks and affects the effectiveness of these tasks.The popular NER deep learning models are mainly based on sequence tagging models.These models are limited by Markov hypothesis,which makes neural network only learning the dependence between tags rather than complete text sequence.The fault tolerance rate of annotation results is low,and tag-scheme are difficult to train with downstream tasks.In the light of the present problem,as our first innovation,we reference the one-stage object detection model in computer vision and propose the novel global search method to recognize named entities,which utilizing the continuity of named entities.This method assumes that words of the same entity have shared boundaries and the same type,and decouples NER into two tasks:entity word recognition task and boundary searching task.Entity word recognition task aims at extracting entity word from text and boundary searching task is to search the entity boundaries.Based on this method,we design the Entity Search Model(ESM).Meanwhile,we propose two new neural networks: Local-Guided Multi-Head-Attention machine(LGA)and Boundary Convolution Block(BCB)to build the entity word recognition module and boundary searching module in ESM to learning above two tasks.In order to prove the validity of the global search method as well as ESM,we experiment on Co NLL-2003 dataset and Onto Note5.0 dataset and compare our model with some current work to show our advancement.We also experiment on resume and Onto Note4.0 Chinese benchmarks to show ESM`s portability.Through the analysis of the experiments,we also discover the dependency between words and entities in the output features of ESM,which may provide more potential information for the downstream tasks.