Research On Named Entity Recognition Method Based On Enhanced Characters

As an important basic task in the field of natural language processing,named entity recognition(NER)is used in many fields such as machine translation,automatic question answering,text summarization,and information retrieval.Traditional entity recognition methods based on rules and statistics need to rely on artificially designed features and are less versatile.The deep learning method can automatically learn samples through neural networks to obtain effective features,and has gradually become the mainstream method for named entity recognition tasks since its introduction.In the existing research on named entity recognition based on deep learning,Word2 Vec,Glove and other distributed word vectors contain the semantic information of the words,and are sent to the neural network model as the vector representation of the words,which can obtain good results.However,distributed word vectors cannot solve problems such as out of vocabulary,rare words,and misspellings.Secondly,distributed word vectors do not contain morphological information such as word suffixes,upper and lower case,etc.,and the character-level knowledge is of great significance to the improvement of task performance.In order to integrate character-level knowledge,existing entity recognition methods usually need to add external resources such as linguistic features,and the process is too complicated.Therefore,this paper studies how to use character-level knowledge to improve the performance of entity recognition tasks without using external resources,automatically extract character-level word vectors of words through neural network models,and implement end-to-end training and testing.The research has important theoretical significance and application value.This paper focuses on how to use character-level knowledge to propose three character enhancement schemes.The main research contents and innovations are as follows:(1)A character-enhanced named entity recognition method based on long chains of characters is proposed.Sentences are sent as long chains of characters into the bidirectional gated recurrent unit network for encoding.The character-level word vectors containing context information and rich features are obtained by word boundaries.(2)A character-enhanced named entity recognition method incorporating selfattention mechanism is proposed.The self-attention mechanism assigns different weights to the convolutional in-word character features,and extracts the features that contribute more to label prediction,so as to obtain better character-level word vectors.(3)A character-enhanced named entity recognition method integrated into the capsule network is proposed.In this paper,the capsule network is first used to acquire character-level word vectors.The local features obtained by CNN convolution are converted into vector capsules through low-level character-level capsules,and then higher-level features between characters and words are extracted through dynamic routing algorithms to obtain word vectors containing character-level knowledge.In view of the above-mentioned methods,this paper conducts experiments on the Co NLL2003 public data set.The experimental results show that the three character enhancement models proposed are significantly better than the baseline model that does not incorporate character information,and the effects are different.The use of characterlevel knowledge can bring significant improvements in entity recognition performance.In addition,based on the character-enhanced named entity recognition model incorporating self-attention mechanism,this paper uses the pre-trained language model BERT to further improve system performance.The named entity recognition system with the pre-trained model BERT achieved an advanced result of 93.13%.