NER – Named Entity Recognition Summary

1.  What is named entity recognition ？

Named entity recognition （Named Entity Recognition, abbreviation NER）, Also known as “ Proper name recognition ”, It refers to the recognition of entities with specific meaning in the text , Mainly including people's names 、 Place names 、 Organization name 、 Proper nouns, etc . Simple speak , It is to identify the boundaries and categories of entity references in natural texts .

2.  The history of named entity recognition

  Early rule based 、 The dictionary method is not detailed . At present, the most widely used method should be based on statistics （ The dependence on corpus is relatively large ）, Using large-scale corpus to learn the annotation model , To mark the positions .CRF yes NER The current mainstream model , Its objective function considers not only the input state characteristic function , It also includes the label transfer characteristic function . When the model is known , Calculate the predicted output sequence for the input sequence, that is, find the optimal sequence to maximize the objective function , It is a dynamic programming problem , have access to Viterbi Algorithm decoding to get the optimal tag sequence .CRF Its advantage is that it can make use of rich internal and contextual feature information in the process of labeling a location .

 3. BiLSTM-CRF

be applied to NER Medium biLSTM-CRF The model is mainly composed of Embedding layer （ There are mainly word vectors , Word vectors and some additional features ）, two-way LSTM layer , And finally CRF Layers make up . Experimental results show that biLSTM-CRF Having reached or exceeded the goal of CRF Model , Become the current deep learning based NER The most popular model in the method . In terms of characteristics , This model inherits the advantages of the deep learning method , No feature engineering required , Using word vectors and character vectors can achieve good results , If there are high-quality dictionary features , Can be further improved .

 

4. summary

       Combine neural networks with CRF Model combined CNN/RNN-CRF It's now NER The mainstream model of . about CNN And RNN, No one has an absolute advantage , Each has its merits . because RNN It has a natural sequence structure , therefore RNN-CRF More widely used . Based on neural network structure NER Method , It inherits the advantages of the deep learning method , No need for a lot of artificial features . Only word vector and word vector can reach the mainstream level , Adding high-quality dictionary features can further improve the effect . For a small number of labeled training sets , The migration study , Semi supervised learning should be the focus of future research .

5. Tool recommendation

5.1. Stanford NER

Named entity recognition system based on conditional random field developed by Stanford University , The system parameters are based on CoNLL、MUC-6、MUC-7 and ACE Named training corpus .

Address ：https://nlp.stanford.edu/software/CRF-NER.shtml

python Realized Github Address ：https://github.com/Lynten/stanford-corenlp

#  install ：pip install stanfordcorenlp
#  Domestic source installation ：pip install stanfordcorenlp -i https://pypi.tuna.tsinghua.edu.cn/simple
#  Use stanfordcorenlp Identify named entity classes 
#  Download model first , Download address ：https://nlp.stanford.edu/software/corenlp-backup-download.html
#  Chinese entity recognition 
from stanfordcorenlp import StanfordCoreNLP
zh_model = StanfordCoreNLP(r'stanford-corenlp-full-2018-02-27', lang='zh')
s_zh = ' I love natural language processing technology ！'
ner_zh = zh_model.ner(s_zh)
s_zh1 = ' I love tian 'anmen square in Beijing ！'
ner_zh1 = zh_model.ner(s_zh1)
print(ner_zh)
print(ner_zh1)

[(' I love ', 'O'), (' natural ', 'O'), (' Language ', 'O'), (' Handle ', 'O'), (' technology ', 'O'), ('！', 'O')]
[(' I love ', 'O'), (' Beijing ', 'STATE_OR_PROVINCE'), (' The tiananmen square ', 'FACILITY'), ('！', 'O')]


#  Entity recognition for English 
eng_model = StanfordCoreNLP(r'stanford-corenlp-full-2018-02-27')
s_eng = 'I love natural language processing technology!'
ner_eng = eng_model.ner(s_eng)
s_eng1 = 'I love Beijing Tiananmen!'
ner_eng1 = eng_model.ner(s_eng1)
print(ner_eng)
print(ner_eng1)

[('I', 'O'), ('love', 'O'), ('natural', 'O'), ('language', 'O'), ('processing', 'O'), ('technology', 'O'), ('!', 'O')]
[('I', 'O'), ('love', 'O'), ('Beijing', 'CITY'), ('Tiananmen', 'LOCATION'), ('!', 'O')]

5.2 MALLET

An open source package for statistical natural language processing developed by the University of Massachusetts , It can realize named entity recognition in the application of sequence annotation tool . Official address ：http://mallet.cs.umass.edu/

5.3 Hanlp

HanLP It's a series of models and algorithms NLP tool kit , It's dominated by big search and fully open source , The goal is to popularize the application of natural language processing in the production environment . Support named entity recognition . Github Address ：https://github.com/hankcs/pyhanlp

Official website ：http://hanlp.linrunsoft.com/

#  install ：pip install pyhanlp
#  Domestic source installation ：pip install pyhanlp  -i https://pypi.tuna.tsinghua.edu.cn/simple
#  adopt crf The algorithm recognizes entities 
from pyhanlp import *
#  Examples of transliteration of names 
CRFnewSegment = HanLP.newSegment("crf")
term_list = CRFnewSegment.seg(" I love tian 'anmen square in Beijing ！")
print(term_list)

[ I /r,  Love /v,  Beijing /ns,  The tiananmen square /ns, ！/w]

5.4  NLTK

NLTK It's an efficient Python Built platform , Used to process human natural language data .

Github Address ：https://github.com/nltk/nltk Official website ：http://www.nltk.org/

#  install ：pip install nltk
#  Domestic source installation ：pip install nltk  -i https://pypi.tuna.tsinghua.edu.cn/simple
import nltk
s = 'I love natural language processing technology!'
s_token = nltk.word_tokenize(s)
s_tagged = nltk.pos_tag(s_token)
s_ner = nltk.chunk.ne_chunk(s_tagged)
print(s_ner)

5.5 SpaCy

Industrial grade natural language processing tools , Unfortunately, Chinese is not supported . Gihub Address ： https://github.com/explosion/spaCy Official website ：https://spacy.io/

#  install ：pip install spaCy
#  Domestic source installation ：pip install spaCy  -i https://pypi.tuna.tsinghua.edu.cn/simple
import spacy
eng_model = spacy.load('en')
s = 'I want to Beijing learning natural language processing technology!'
#  Named entity recognition 
s_ent = eng_model(s)
for ent in s_ent.ents:
   print(ent, ent.label_, ent.label)

Beijing GPE 382

 6.6 Crfsuite

You can load your own data set to train CRF Entity recognition model .

Document address ：

https://sklearn-crfsuite.readthedocs.io/en/latest/?badge=latest

Code Uploaded ：https://github.com/yuquanle/StudyForNLP/blob/master/NLPbasic/NER.ipynb

reference ： https://mp.weixin.qq.com/s/7R3Y2-nD5fELPa9rtnK7Tg