1. 分词 word segmentation
国内有jieba 分词
2. Named Entity Recognition
-
训练自己的Model
How can I train my own NER model
https://nlp.stanford.edu/software/crf-faq.html#a
C:my_studyMLNLPstanford-ner-2018-02-27>java -cp stanford-ner.jar edu.stanford.nlp.ie.crf.CRFClassifier -prop chinese.meal.fpp.prop
Invoked on Thu Mar 22 16:34:06 CST 2018 with arguments: -prop chinese.meal.fpp.prop
usePrevSequences=true
useClassFeature=true
useTypeSeqs2=true
useSequences=true
wordShape=chris2useLC
useTypeySequences=true
useDisjunctive=true
noMidNGrams=true
serializeTo=ner-model.ser.gz
maxNGramLeng=6
useNGrams=true
usePrev=true
useNext=true
maxLeft=1
trainFile=chinese.meal.fpp.tsv
map=word=0,answer=1
useWord=true
useTypeSeqs=true
numFeatures = 564
Time to convert docs to feature indices: 0.0 seconds
numClasses: 5 [0=O,1=TIME,2=QUANTITY,3=UNIT,4=FOOD]
numDocuments: 1
numDatums: 56
numFeatures: 564
Time to convert docs to data/labels: 0.0 seconds
numWeights: 6460
QNMinimizer called on double function of 6460 variables, using M = 25.
An explanation of the output:
Iter The number of iterations
evals The number of function evaluations
SCALING <D> Diagonal scaling was used; <I> Scaled Identity
LINESEARCH [## M steplength] Minpack linesearch
1-Function value was too high
2-Value ok, gradient positive, positive curvature
3-Value ok, gradient negative, positive curvature
4-Value ok, gradient negative, negative curvature
[.. B] Backtracking
VALUE The current function value
TIME Total elapsed time
|GNORM| The current norm of the gradient
{RELNORM} The ratio of the current to initial gradient norms
AVEIMPROVE The average improvement / current value
EVALSCORE The last available eval score
Iter ## evals ## <SCALING> [LINESEARCH] VALUE TIME |GNORM| {RELNORM} AVEIMPROVE EVALSCORE
Iter 1 evals 1 <D> [M 1.000E-1] 9.068E2 0.04s |4.550E1| {4.995E-1} 0.000E0 -
Iter 2 evals 2 <D> [M 1.000E0] 6.222E2 0.05s |3.525E1| {3.870E-1} 2.287E-1 -
Iter 3 evals 3 <D> [M 1.000E0] 2.386E2 0.07s |5.406E1| {5.935E-1} 9.334E-1 -
Iter 4 evals 4 <D> [M 1.000E0] 9.082E1 0.08s |1.571E1| {1.724E-1} 2.246E0 -
Iter 5 evals 5 <D> [M 1.000E0] 7.031E1 0.10s |1.181E1| {1.297E-1} 2.379E0 -
Iter 6 evals 6 <D> [M 1.000E0] 5.308E1 0.11s |1.025E1| {1.125E-1} 2.681E0 -
Iter 7 evals 7 <D> [1M 2.740E-1] 2.988E1 0.14s |7.586E0| {8.328E-2} 4.193E0 -
Iter 8 evals 9 <D> [1M 1.292E-1] 2.234E1 0.16s |6.471E0| {7.105E-2} 4.949E0 -
Iter 9 evals 11 <D> [1M 1.801E-1] 1.615E1 0.18s |5.573E0| {6.118E-2} 6.127E0 -
Iter 10 evals 13 <D> [1M 1.815E-1] 1.218E1 0.24s |4.477E0| {4.915E-2} 7.346E0 -
Iter 11 evals 15 <D> [1M 3.119E-1] 8.873E0 0.30s |4.694E0| {5.154E-2} 6.912E0 -
Iter 12 evals 17 <D> [1M 4.760E-1] 6.621E0 0.31s |2.092E0| {2.296E-2} 3.504E0 -
Iter 13 evals 19 <D> [M 1.000E0] 6.093E0 0.32s |1.906E0| {2.092E-2} 1.390E0 -
Iter 14 evals 20 <D> [M 1.000E0] 5.844E0 0.33s |9.067E-1| {9.955E-3} 1.103E0 -
Iter 15 evals 21 <D> [M 1.000E0] 5.721E0 0.33s |5.774E-1| {6.339E-3} 8.279E-1 -
Iter 16 evals 22 <D> [M 1.000E0] 5.660E0 0.34s |3.535E-1| {3.881E-3} 4.279E-1 -
Iter 17 evals 23 <D> [M 1.000E0] 5.640E0 0.35s |1.946E-1| {2.137E-3} 2.961E-1 -
Iter 18 evals 24 <D> [M 1.000E0] 5.632E0 0.36s |7.832E-2| {8.599E-4} 1.868E-1 -
Iter 19 evals 25 <D> [M 1.000E0] 5.631E0 0.38s |3.559E-2| {3.907E-4} 1.163E-1 -
Iter 20 evals 26 <D> [M 1.000E0] 5.631E0 0.39s |2.149E-2| {2.359E-4} 5.758E-2 -
Iter 21 evals 27 <D> [M 1.000E0] 5.631E0 0.41s |1.027E-2| {1.128E-4} 1.758E-2 -
Iter 22 evals 28 <D> [M 1.000E0] 5.631E0 0.42s |3.631E-3| {3.986E-5} 8.218E-3 -
Iter 23 evals 29 <D> [M 1.000E0] 5.631E0 0.44s |1.629E-3| {1.789E-5} 3.791E-3 -
Iter 24 evals 30 <D> [M 1.000E0] 5.631E0 0.45s |9.548E-4| {1.048E-5} 1.596E-3 -
Iter 25 evals 31 <D> [M 1.000E0] 5.631E0 0.45s |5.724E-4| {6.284E-6} 5.196E-4 -
Iter 26 evals 32 <D> [M 1.000E0] 5.631E0 0.47s |1.578E-4| {1.732E-6} 1.686E-4 -
QNMinimizer terminated due to average improvement: | newest_val - previous_val | / |newestVal| < TOL
Total time spent in optimization: 0.49s
CRFClassifier training ... done [0.6 sec].
Serializing classifier to ner-model.ser.gz... done.
2. 使用训练好的Model来evaluate 一下，看看效果怎么样.
C:my_studyMLNLPstanford-ner-2018-02-27>java -cp stanford-ner.jar edu.stanford.nlp.ie.crf.CRFClassifier -loadClassifier ner-model.ser.gz -testFile chinese.meal.fpp.test.tsv Invoked on Thu Mar 22 16:30:48 CST 2018 with arguments: -loadClassifier ner-model.ser.gz -testFile chinese.meal.fpp.test.tsv testFile=chinese.meal.fpp.test.tsv loadClassifier=ner-model.ser.gz Loading classifier from ner-model.ser.gz ... done [0.1 sec]. 我 O O 今天 O O 晚上 TIME TIME 吃 O O 了 O O 两 QUANTITY QUANTITY 盘 UNIT UNIT 回锅肉 FOOD FOOD CRFClassifier tagged 8 words in 1 documents at 88.89 words per second. Entity P R F1 TP FP FN FOOD 1.0000 1.0000 1.0000 1 0 0 QUANTITY 1.0000 1.0000 1.0000 1 0 0 TIME 1.0000 1.0000 1.0000 1 0 0 UNIT 1.0000 1.0000 1.0000 1 0 0 Totals 1.0000 1.0000 1.0000 4 0 0
还不错哦！
Ref:
1. Standford NLP NER: https://nlp.stanford.edu/software/CRF-NER.html