深度學(xué)習(xí)簡(jiǎn)介
深度學(xué)習(xí)的資料很多，這里就不展開(kāi)了講，本文就介紹中文NLP的序列標(biāo)注工作的一般方法。
機(jī)器學(xué)習(xí)與深度學(xué)習(xí)
簡(jiǎn)單來(lái)說(shuō)，機(jī)器學(xué)習(xí)就是根據(jù)樣本(即數(shù)據(jù))學(xué)習(xí)得到一個(gè)模型，再根據(jù)這個(gè)模型預(yù)測(cè)的一種方法。
ML算法很多，Naive Bayes樸素貝葉斯、Decision Tree決策樹(shù)、Support Vector Machine支持向量機(jī)、Logistic Regression邏輯回歸、Conditional Random Field條件隨機(jī)場(chǎng)等。
而深度學(xué)習(xí)，簡(jiǎn)單來(lái)說(shuō)是一種有多層隱層的感知機(jī)。
DL也分很多模型，但一般了解Convolution Neural Network卷積神經(jīng)網(wǎng)絡(luò)、Recurrent Neural Network循環(huán)神經(jīng)網(wǎng)絡(luò)就夠了(當(dāng)然都要學(xué)，這里是指前期學(xué)習(xí)階段可以側(cè)重這兩個(gè))。
異同：ML是一種淺層學(xué)習(xí)，一般來(lái)說(shuō)都由人工設(shè)計(jì)特征，而DL則用pre-training或者無(wú)監(jiān)督學(xué)習(xí)來(lái)抽取特征表示，再使用監(jiān)督學(xué)習(xí)來(lái)訓(xùn)練預(yù)測(cè)模型（當(dāng)然不全都是這樣）。
本文主要用于介紹DL在中文NLP的應(yīng)用，所以采用了使用最為簡(jiǎn)單、方便的DL框架keras來(lái)開(kāi)發(fā)，它是構(gòu)建于兩個(gè)非常受歡迎的DL框架theano和tensorflow之上的上層應(yīng)用框架。
NLP簡(jiǎn)介
Natural Language Process自然語(yǔ)言處理又分為NLU自然語(yǔ)言理解和NLG自然語(yǔ)言生成。而分詞、詞性標(biāo)注、實(shí)體識(shí)別、依存分析則是NLP的基礎(chǔ)工作，它們都可以理解為一種序列標(biāo)注工作。
序列標(biāo)注工作簡(jiǎn)介
詞向量簡(jiǎn)介
Word Embedding詞向量方法，用實(shí)數(shù)向量來(lái)表示一個(gè)詞的方法，是對(duì)One-hot Representation的一種優(yōu)化。優(yōu)點(diǎn)是低維，而且可以方便的用數(shù)學(xué)距離衡量詞的詞義相似度，缺點(diǎn)是詞一多，模型就有點(diǎn)大，所以又有工作提出了Char Embedding方法，這種方法訓(xùn)練出來(lái)的模型很小，但丟失了很多的語(yǔ)義信息，所以又有基于分詞信息的字向量的研究工作。
中文NLP序列標(biāo)注之CWS
CWS簡(jiǎn)介
Chinese Word Segmentation中文分詞是中文NLP的基礎(chǔ)，一般來(lái)說(shuō)中文分詞有兩種方法，一種是基于詞典的方法，一種是基于ML或者DL的方法。CWS的發(fā)展可以參考漫話中文分詞，簡(jiǎn)單來(lái)說(shuō)基于詞典的方法實(shí)現(xiàn)簡(jiǎn)單、速度快，但是對(duì)歧義和未登錄詞沒(méi)有什么好的辦法，而基于ML和DL的方法實(shí)現(xiàn)復(fù)雜、速度較慢，但是可以較好地應(yīng)對(duì)歧義和OOV(Out-Of-Vocabulary)。
基于詞典的方法應(yīng)用最廣的應(yīng)該是正向最大匹配，而基于ML的CWS效果比較好的算法是CRF，本文主要介紹基于DL的方法，但在實(shí)際應(yīng)用中應(yīng)該合理的結(jié)合兩種方法。
標(biāo)注集與評(píng)估方法
這里采用B(Begin字為詞的起始)、M(Middle字為詞的中間)、E(End字為詞的結(jié)束)、S(Single單字詞)標(biāo)注集，訓(xùn)練預(yù)料和評(píng)估工具采用SIGHAN中的方法，具體可以參考我的另一篇文章SIGHAN測(cè)評(píng)中文分詞的方法與指標(biāo)介紹。
模型
原理是采用bi-directional LSTM模型訓(xùn)練后對(duì)句子進(jìn)行預(yù)測(cè)得到一個(gè)標(biāo)注的概率，再使用Viterbi算法尋找最優(yōu)的標(biāo)注序列。在分詞的工作中不需要加入詞向量，提升效果不明顯。
實(shí)現(xiàn)
預(yù)處理
#!/usr/bin/env python
#-*- coding: utf-8 -*-
#2016年 03月 03日 星期四 11:01:05 CST by Demobin
import json
import h5py
import string
import codecs
corpus_tags = ['S', 'B', 'M', 'E']
def saveCwsInfo(path, cwsInfo):
'''保存分詞訓(xùn)練數(shù)據(jù)字典和概率'''
print('save cws info to %s'%path)
fd = open(path, 'w')
(initProb, tranProb), (vocab, indexVocab) = cwsInfo
j = json.dumps((initProb, tranProb))
fd.write(j + '\n')
for char in vocab:
fd.write(char.encode('utf-8') + '\t' + str(vocab[char]) + '\n')
fd.close()
def loadCwsInfo(path):
'''載入分詞訓(xùn)練數(shù)據(jù)字典和概率'''
print('load cws info from %s'%path)
fd = open(path, 'r')
line = fd.readline()
j = json.loads(line.strip())
initProb, tranProb = j[0], j[1]
lines = fd.readlines()
fd.close()
vocab = {}
indexVocab = [0 for i in range(len(lines))]
for line in lines:
rst = line.strip().split('\t')
if len(rst) < 2: continue
char, index = rst[0].decode('utf-8'), int(rst[1])
vocab[char] = index
indexVocab[index] = char
return (initProb, tranProb), (vocab, indexVocab)
def saveCwsData(path, cwsData):
'''保存分詞訓(xùn)練輸入樣本'''
print('save cws data to %s'%path)
#采用hdf5保存大矩陣效率最高
fd = h5py.File(path,'w')
(X, y) = cwsData
fd.create_dataset('X', data = X)
fd.create_dataset('y', data = y)
fd.close()
def loadCwsData(path):
'''載入分詞訓(xùn)練輸入樣本'''
print('load cws data from %s'%path)
fd = h5py.File(path,'r')
X = fd['X'][:]
y = fd['y'][:]
fd.close()
return (X, y)
def sent2vec2(sent, vocab, ctxWindows = 5):
charVec = []
for char in sent:
if char in vocab:
charVec.append(vocab[char])
else:
charVec.append(vocab['retain-unknown'])
#首尾padding
num = len(charVec)
pad = int((ctxWindows - 1)/2)
for i in range(pad):
charVec.insert(0, vocab['retain-padding'] )
charVec.append(vocab['retain-padding'] )
X = []
for i in range(num):
X.append(charVec[i:i + ctxWindows])
return X
def sent2vec(sent, vocab, ctxWindows = 5):
chars = []
for char in sent:
chars.append(char)
return sent2vec2(chars, vocab, ctxWindows = ctxWindows)
def doc2vec(fname, vocab):
'''文檔轉(zhuǎn)向量'''
#一次性讀入文件，注意內(nèi)存
fd = codecs.open(fname, 'r', 'utf-8')
lines = fd.readlines()
fd.close()
#樣本集
X = []
y = []
#標(biāo)注統(tǒng)計(jì)信息
tagSize = len(corpus_tags)
tagCnt = [0 for i in range(tagSize)]
tagTranCnt = [[0 for i in range(tagSize)] for j in range(tagSize)]
#遍歷行
for line in lines:
#按空格分割
words = line.strip('\n').split()
#每行的分詞信息
chars = []
tags = []
#遍歷詞
for word in words:
#包含兩個(gè)字及以上的詞
if len(word) > 1:
#詞的首字
chars.append(word[0])
tags.append(corpus_tags.index('B'))
#詞中間的字
for char in word[1:(len(word) - 1)]:
chars.append(char)
tags.append(corpus_tags.index('M'))
#詞的尾字
chars.append(word[-1])
tags.append(corpus_tags.index('E'))
#單字詞
else:
chars.append(word)
tags.append(corpus_tags.index('S'))
#字向量表示
lineVecX = sent2vec2(chars, vocab, ctxWindows = 7)
#統(tǒng)計(jì)標(biāo)注信息
lineVecY = []
lastTag = -1
for tag in tags:
#向量
lineVecY.append(tag)
#lineVecY.append(corpus_tags[tag])
#統(tǒng)計(jì)tag頻次
tagCnt[tag] += 1
#統(tǒng)計(jì)tag轉(zhuǎn)移頻次
if lastTag != -1:
tagTranCnt[lastTag][tag] += 1
#暫存上一次的tag
lastTag = tag
X.extend(lineVecX)
y.extend(lineVecY)
#字總頻次
charCnt = sum(tagCnt)
#轉(zhuǎn)移總頻次
tranCnt = sum([sum(tag) for tag in tagTranCnt])
#tag初始概率
initProb = []
for i in range(tagSize):
initProb.append(tagCnt[i]/float(charCnt))
#tag轉(zhuǎn)移概率
tranProb = []
for i in range(tagSize):
p = []
for j in range(tagSize):
p.append(tagTranCnt[i][j]/float(tranCnt))
tranProb.append(p)
return X, y, initProb, tranProb
def genVocab(fname, delimiters = [' ', '\n']):
#一次性讀入文件，注意內(nèi)存
fd = codecs.open(fname, 'r', 'utf-8')
data = fd.read()
fd.close()
vocab = {}
indexVocab = []
#遍歷
index = 0
for char in data:
#如果為分隔符則無(wú)需加入字典
if char not in delimiters and char not in vocab:
vocab[char] = index
indexVocab.append(char)
index += 1
#加入未登陸新詞和填充詞
vocab['retain-unknown'] = len(vocab)
vocab['retain-padding'] = len(vocab)
indexVocab.append('retain-unknown')
indexVocab.append('retain-padding')
#返回字典與索引
return vocab, indexVocab
def load(fname):
print 'train from file', fname
delims = [' ', '\n']
vocab, indexVocab = genVocab(fname)
X, y, initProb, tranProb = doc2vec(fname, vocab)
print len(X), len(y), len(vocab), len(indexVocab)
print initProb
print tranProb
return (X, y), (initProb, tranProb), (vocab, indexVocab)
if __name__ == '__main__':
load('~/work/corpus/icwb2/training/msr_training.utf8')
模型
#!/usr/bin/env python
#-*- coding: utf-8 -*-
#2016年 03月 03日 星期四 11:01:05 CST by Demobin
import numpy as np
import json
import h5py
import codecs
from dataset import cws
from util import viterbi
from sklearn.model_selection import train_test_split
from keras.preprocessing import sequence
from keras.optimizers import SGD, RMSprop, Adagrad
from keras.utils import np_utils
from keras.models import Sequential,Graph, model_from_json
from keras.layers.core import Dense, Dropout, Activation, TimeDistributedDense
from keras.layers.embeddings import Embedding
from keras.layers.recurrent import LSTM, GRU, SimpleRNN
from gensim.models import Word2Vec
def train(cwsInfo, cwsData, modelPath, weightPath):
(initProb, tranProb), (vocab, indexVocab) = cwsInfo
(X, y) = cwsData
train_X, test_X, train_y, test_y = train_test_split(X, y , train_size=0.9, random_state=1)
train_X = np.array(train_X)
train_y = np.array(train_y)
test_X = np.array(test_X)
test_y = np.array(test_y)
outputDims = len(cws.corpus_tags)
Y_train = np_utils.to_categorical(train_y, outputDims)
Y_test = np_utils.to_categorical(test_y, outputDims)
batchSize = 128
vocabSize = len(vocab) + 1
wordDims = 100
maxlen = 7
hiddenDims = 100
w2vModel = Word2Vec.load('model/sougou.char.model')
embeddingDim = w2vModel.vector_size
embeddingUnknown = [0 for i in range(embeddingDim)]
embeddingWeights = np.zeros((vocabSize + 1, embeddingDim))
for word, index in vocab.items():
if word in w2vModel:
e = w2vModel[word]
else:
e = embeddingUnknown
embeddingWeights[index, :] = e
#LSTM
model = Sequential()
model.add(Embedding(output_dim = embeddingDim, input_dim = vocabSize + 1,
input_length = maxlen, mask_zero = True, weights = [embeddingWeights]))
model.add(LSTM(output_dim = hiddenDims, return_sequences = True))
model.add(LSTM(output_dim = hiddenDims, return_sequences = False))
model.add(Dropout(0.5))
model.add(Dense(outputDims))
model.add(Activation('softmax'))
model.compile(loss = 'categorical_crossentropy', optimizer = 'adam')
result = model.fit(train_X, Y_train, batch_size = batchSize,
nb_epoch = 20, validation_data = (test_X,Y_test), show_accuracy=True)
j = model.to_json()
fd = open(modelPath, 'w')
fd.write(j)
fd.close()
model.save_weights(weightPath)
return model
def loadModel(modelPath, weightPath):
fd = open(modelPath, 'r')
j = fd.read()
fd.close()
model = model_from_json(j)
model.load_weights(weightPath)
return model
# 根據(jù)輸入得到標(biāo)注推斷
def cwsSent(sent, model, cwsInfo):
(initProb, tranProb), (vocab, indexVocab) = cwsInfo
vec = cws.sent2vec(sent, vocab, ctxWindows = 7)
vec = np.array(vec)
probs = model.predict_proba(vec)
#classes = model.predict_classes(vec)
prob, path = viterbi.viterbi(vec, cws.corpus_tags, initProb, tranProb, probs.transpose())
ss = ''
for i, t in enumerate(path):
ss += '%s/%s '%(sent[i], cws.corpus_tags[t])
ss = ''
word = ''
for i, t in enumerate(path):
if cws.corpus_tags[t] == 'S':
ss += sent[i] + ' '
word = ''
elif cws.corpus_tags[t] == 'B':
word += sent[i]
elif cws.corpus_tags[t] == 'E':
word += sent[i]
ss += word + ' '
word = ''
elif cws.corpus_tags[t] == 'M':
word += sent[i]
return ss
def cwsFile(fname, dstname, model, cwsInfo):
fd = codecs.open(fname, 'r', 'utf-8')
lines = fd.readlines()
fd.close()
fd = open(dstname, 'w')
for line in lines:
rst = cwsSent(line.strip(), model, cwsInfo)
fd.write(rst.encode('utf-8') + '\n')
fd.close()
def test():
print 'Loading vocab...'
cwsInfo = cws.loadCwsInfo('./model/cws.info')
cwsData = cws.loadCwsData('./model/cws.data')
print 'Done!'
print 'Loading model...'
#model = train(cwsInfo, cwsData, './model/cws.w2v.model', './model/cws.w2v.model.weights')
#model = loadModel('./model/cws.w2v.model', './model/cws.w2v.model.weights')
model = loadModel('./model/cws.model', './model/cws.model.weights')
print 'Done!'
print '-------------start predict----------------'
#s = u'為寂寞的夜空畫上一個(gè)月亮'
#print cwsSent(s, model, cwsInfo)
cwsFile('~/work/corpus/icwb2/testing/msr_test.utf8', './msr_test.utf8.cws', model, cwsInfo)
if __name__ == '__main__':
test()
viterbi算法
#!/usr/bin/python
# -*- coding: utf-8 -*-
#2016年 01月 28日 星期四 17:14:03 CST by Demobin
def _print(hiddenstates, V):
s = " " + " ".join(("%7d" % i) for i in range(len(V))) + "\n"
for i, state in enumerate(hiddenstates):
s += "%.5s: " % state
s += " ".join("%.7s" % ("%f" % v[i]) for v in V)
s += "\n"
print(s)
#標(biāo)準(zhǔn)viterbi算法，參數(shù)為觀察狀態(tài)、隱藏狀態(tài)、概率三元組(初始概率、轉(zhuǎn)移概率、觀察概率)
def viterbi(obs, states, start_p, trans_p, emit_p):
lenObs = len(obs)
lenStates = len(states)
V = [[0.0 for col in range(lenStates)] for row in range(lenObs)]
path = [[0 for col in range(lenObs)] for row in range(lenStates)]
#t = 0時(shí)刻
for y in range(lenStates):
#V[0][y] = start_p[y] * emit_p[y][obs[0]]
V[0][y] = start_p[y] * emit_p[y][0]
path[y][0] = y
#t > 1時(shí)
for t in range(1, lenObs):
newpath = [[0.0 for col in range(lenObs)] for row in range(lenStates)]
for y in range(lenStates):
prob = -1
state = 0
for y0 in range(lenStates):
#nprob = V[t - 1][y0] * trans_p[y0][y] * emit_p[y][obs[t]]
nprob = V[t - 1][y0] * trans_p[y0][y] * emit_p[y][t]
if nprob > prob:
prob = nprob
state = y0
#記錄最大概率
V[t][y] = prob
#記錄路徑
newpath[y][:t] = path[state][:t]
newpath[y][t] = y
path = newpath
prob = -1
state = 0
for y in range(lenStates):
if V[lenObs - 1][y] > prob:
prob = V[lenObs - 1][y]
state = y
#_print(states, V)
return prob, path[state]
def example():
#隱藏狀態(tài)
hiddenstates = ('Healthy', 'Fever')
#觀察狀態(tài)
observations = ('normal', 'cold', 'dizzy')
#初始概率
'''
Healthy': 0.6, 'Fever': 0.4
'''
start_p = [0.6, 0.4]
#轉(zhuǎn)移概率
'''
Healthy' : {'Healthy': 0.7, 'Fever': 0.3},
Fever' : {'Healthy': 0.4, 'Fever': 0.6}
'''
trans_p = [[0.7, 0.3], [0.4, 0.6]]
#發(fā)射概率/輸出概率/觀察概率
'''
Healthy' : {'normal': 0.5, 'cold': 0.4, 'dizzy': 0.1},
Fever' : {'normal': 0.1, 'cold': 0.3, 'dizzy': 0.6}
'''
emit_p = [[0.5, 0.4, 0.1], [0.1, 0.3, 0.6]]
return viterbi(observations,
hiddenstates,
start_p,
trans_p,
emit_p)
if __name__ == '__main__':
print(example())
中文NLP序列標(biāo)注之POS
預(yù)處理
#!/usr/bin/env python
# -*- coding: utf-8 -*-
#2016年 03月 03日 星期四 11:01:05 CST by Demobin
import h5py
import json
import codecs
mappings = {
#人民日?qǐng)?bào)標(biāo)注集：863標(biāo)注集
'w': 'wp',
't': 'nt',
'nr': 'nh',
'nx': 'nz',
'nn': 'n',
'nzz': 'n',
'Ng': 'n',
'f': 'nd',
's': 'nl',
'Vg': 'v',
'vd': 'v',
'vn': 'v',
'vnn': 'v',
'ad': 'a',
'an': 'a',
'Ag': 'a',
'l': 'i',
'z': 'a',
'mq': 'm',
'Mg': 'm',
'Tg': 'nt',
'y': 'u',
'Yg': 'u',
'Dg': 'd',
'Rg': 'r',
'Bg': 'b',
'pn': 'p',
}
tags_863 = {
'a' : [0, '形容詞'],
'b' : [1, '區(qū)別詞'],
'c' : [2, '連詞'],
'd' : [3, '副詞'],
'e' : [4, '嘆詞'],
'g' : [5, '語(yǔ)素字'],
'h' : [6, '前接成分'],
'i' : [7, '習(xí)用語(yǔ)'],
'j' : [8, '簡(jiǎn)稱'],
'k' : [9, '后接成分'],
'm' : [10, '數(shù)詞'],
'n' : [11, '名詞'],
'nd': [12, '方位名詞'],
'nh': [13, '人名'],
'ni': [14, '團(tuán)體、機(jī)構(gòu)、組織的專名'],
'nl': [15, '處所名詞'],
'ns': [16, '地名'],
'nt': [17, '時(shí)間名詞'],
'nz': [18, '其它專名'],
'o' : [19, '擬聲詞'],
'p' : [20, '介詞'],
'q' : [21, '量詞'],
'r' : [22, '代詞'],
'u' : [23, '助詞'],
'v' : [24, '動(dòng)詞'],
'wp': [25, '標(biāo)點(diǎn)'],
'ws': [26, '字符串'],
'x' : [27, '非語(yǔ)素字'],
}
def genCorpusTags():
s = ''
features = ['b', 'm', 'e', 's']
for tag in tags:
for f in features:
s += '\'' + tag + '-' + f + '\'' + ','
print s
corpus_tags = [
'nh-b','nh-m','nh-e','nh-s','ni-b','ni-m','ni-e','ni-s','nl-b','nl-m','nl-e','nl-s','nd-b','nd-m','nd-e','nd-s','nz-b','nz-m','nz-e','nz-s','ns-b','ns-m','ns-e','ns-s','nt-b','nt-m','nt-e','nt-s','ws-b','ws-m','ws-e','ws-s','wp-b','wp-m','wp-e','wp-s','a-b','a-m','a-e','a-s','c-b','c-m','c-e','c-s','b-b','b-m','b-e','b-s','e-b','e-m','e-e','e-s','d-b','d-m','d-e','d-s','g-b','g-m','g-e','g-s','i-b','i-m','i-e','i-s','h-b','h-m','h-e','h-s','k-b','k-m','k-e','k-s','j-b','j-m','j-e','j-s','m-b','m-m','m-e','m-s','o-b','o-m','o-e','o-s','n-b','n-m','n-e','n-s','q-b','q-m','q-e','q-s','p-b','p-m','p-e','p-s','r-b','r-m','r-e','r-s','u-b','u-m','u-e','u-s','v-b','v-m','v-e','v-s','x-b','x-m','x-e','x-s'
]
def savePosInfo(path, posInfo):
'''保存分詞訓(xùn)練數(shù)據(jù)字典和概率'''
print('save pos info to %s'%path)
fd = open(path, 'w')
(initProb, tranProb), (vocab, indexVocab) = posInfo
j = json.dumps((initProb, tranProb))
fd.write(j + '\n')
for char in vocab:
fd.write(char.encode('utf-8') + '\t' + str(vocab[char]) + '\n')
fd.close()
def loadPosInfo(path):
'''載入分詞訓(xùn)練數(shù)據(jù)字典和概率'''
print('load pos info from %s'%path)
fd = open(path, 'r')
line = fd.readline()
j = json.loads(line.strip())
initProb, tranProb = j[0], j[1]
lines = fd.readlines()
fd.close()
vocab = {}
indexVocab = [0 for i in range(len(lines))]
for line in lines:
rst = line.strip().split('\t')
if len(rst) < 2: continue
char, index = rst[0].decode('utf-8'), int(rst[1])
vocab[char] = index
indexVocab[index] = char
return (initProb, tranProb), (vocab, indexVocab)
def savePosData(path, posData):
'''保存分詞訓(xùn)練輸入樣本'''
print('save pos data to %s'%path)
#采用hdf5保存大矩陣效率最高
fd = h5py.File(path,'w')
(X, y) = posData
fd.create_dataset('X', data = X)
fd.create_dataset('y', data = y)
fd.close()
def loadPosData(path):
'''載入分詞訓(xùn)練輸入樣本'''
print('load pos data from %s'%path)
fd = h5py.File(path,'r')
X = fd['X'][:]
y = fd['y'][:]
fd.close()
return (X, y)
def sent2vec2(sent, vocab, ctxWindows = 5):
charVec = []
for char in sent:
if char in vocab:
charVec.append(vocab[char])
else:
charVec.append(vocab['retain-unknown'])
#首尾padding
num = len(charVec)
pad = int((ctxWindows - 1)/2)
for i in range(pad):
charVec.insert(0, vocab['retain-padding'] )
charVec.append(vocab['retain-padding'] )
X = []
for i in range(num):
X.append(charVec[i:i + ctxWindows])
return X
def sent2vec(sent, vocab, ctxWindows = 5):
chars = []
words = sent.split()
for word in words:
#包含兩個(gè)字及以上的詞
if len(word) > 1:
#詞的首字
chars.append(word[0] + '_b')
#詞中間的字
for char in word[1:(len(word) - 1)]:
chars.append(char + '_m')
#詞的尾字
chars.append(word[-1] + '_e')
#單字詞
else:
chars.append(word + '_s')
return sent2vec2(chars, vocab, ctxWindows = ctxWindows)
def doc2vec(fname, vocab):
'''文檔轉(zhuǎn)向量'''
#一次性讀入文件，注意內(nèi)存
fd = codecs.open(fname, 'r', 'utf-8')
lines = fd.readlines()
fd.close()
#樣本集
X = []
y = []
#標(biāo)注統(tǒng)計(jì)信息
tagSize = len(corpus_tags)
tagCnt = [0 for i in range(tagSize)]
tagTranCnt = [[0 for i in range(tagSize)] for j in range(tagSize)]
#遍歷行
for line in lines:
#按空格分割
words = line.strip('\n').split()
#每行的分詞信息
chars = []
tags = []
#遍歷詞
for word in words:
rst = word.split('/')
if len(rst) <= 0:
print word
continue
word, tag = rst[0], rst[1].decode('utf-8')
if tag not in tags_863:
tag = mappings[tag]
#包含兩個(gè)字及以上的詞
if len(word) > 1:
#詞的首字
chars.append(word[0] + '_b')
tags.append(corpus_tags.index(tag + '-' + 'b'))
#詞中間的字
for char in word[1:(len(word) - 1)]:
chars.append(char + '_m')
tags.append(corpus_tags.index(tag + '-' + 'm'))
#詞的尾字
chars.append(word[-1] + '_e')
tags.append(corpus_tags.index(tag + '-' + 'e'))
#單字詞
else:
chars.append(word + '_s')
tags.append(corpus_tags.index(tag + '-' + 's'))
#字向量表示
lineVecX = sent2vec2(chars, vocab, ctxWindows = 7)
#統(tǒng)計(jì)標(biāo)注信息
lineVecY = []
lastTag = -1
for tag in tags:
#向量
lineVecY.append(tag)
#lineVecY.append(corpus_tags[tag])
#統(tǒng)計(jì)tag頻次
tagCnt[tag] += 1
#統(tǒng)計(jì)tag轉(zhuǎn)移頻次
if lastTag != -1:
tagTranCnt[lastTag][tag] += 1
#暫存上一次的tag
lastTag = tag
X.extend(lineVecX)
y.extend(lineVecY)
#字總頻次
charCnt = sum(tagCnt)
#轉(zhuǎn)移總頻次
tranCnt = sum([sum(tag) for tag in tagTranCnt])
#tag初始概率
initProb = []
for i in range(tagSize):
initProb.append(tagCnt[i]/float(charCnt))
#tag轉(zhuǎn)移概率
tranProb = []
for i in range(tagSize):
p = []
for j in range(tagSize):
p.append(tagTranCnt[i][j]/float(tranCnt))
tranProb.append(p)
return X, y, initProb, tranProb
def vocabAddChar(vocab, indexVocab, index, char):
if char not in vocab:
vocab[char] = index
indexVocab.append(char)
index += 1
return index
def genVocab(fname, delimiters = [' ', '\n']):
#一次性讀入文件，注意內(nèi)存
fd = codecs.open(fname, 'r', 'utf-8')
lines = fd.readlines()
fd.close()
vocab = {}
indexVocab = []
#遍歷所有行
index = 0
for line in lines:
words = line.strip().split()
if words <= 0: continue
#遍歷所有詞
for word in words:
word, tag = word.split('/')
#包含兩個(gè)字及以上的詞
if len(word) > 1:
#詞的首字
char = word[0] + '_b'
index = vocabAddChar(vocab, indexVocab, index, char)
#詞中間的字
for char in word[1:(len(word) - 1)]:
char = char + '_m'
index = vocabAddChar(vocab, indexVocab, index, char)
#詞的尾字
char = word[-1] + '_e'
index = vocabAddChar(vocab, indexVocab, index, char)
#單字詞
else:
char = word + '_s'
index = vocabAddChar(vocab, indexVocab, index, char)
#加入未登陸新詞和填充詞
vocab['retain-unknown'] = len(vocab)
vocab['retain-padding'] = len(vocab)
indexVocab.append('retain-unknown')
indexVocab.append('retain-padding')
#返回字典與索引
return vocab, indexVocab
def load(fname):
print 'train from file', fname
delims = [' ', '\n']
vocab, indexVocab = genVocab(fname)
X, y, initProb, tranProb = doc2vec(fname, vocab)
print len(X), len(y), len(vocab), len(indexVocab)
print initProb
print tranProb
return (X, y), (initProb, tranProb), (vocab, indexVocab)
def test():
load('../data/pos.train')
if __name__ == '__main__':
test()
模型
#!/usr/bin/env python
#-*- coding: utf-8 -*-
#2016年 03月 03日 星期四 11:01:05 CST by Demobin
import numpy as np
import json
import h5py
import codecs
from dataset import pos
from util import viterbi
from sklearn.model_selection import train_test_split
from keras.preprocessing import sequence
from keras.optimizers import SGD, RMSprop, Adagrad
from keras.utils import np_utils
from keras.models import Sequential,Graph, model_from_json
from keras.layers.core import Dense, Dropout, Activation, TimeDistributedDense
from keras.layers.embeddings import Embedding
from keras.layers.recurrent import LSTM, GRU, SimpleRNN
from util import pChar
def train(posInfo, posData, modelPath, weightPath):
(initProb, tranProb), (vocab, indexVocab) = posInfo
(X, y) = posData
train_X, test_X, train_y, test_y = train_test_split(X, y , train_size=0.9, random_state=1)
train_X = np.array(train_X)
train_y = np.array(train_y)
test_X = np.array(test_X)
test_y = np.array(test_y)
outputDims = len(pos.corpus_tags)
Y_train = np_utils.to_categorical(train_y, outputDims)
Y_test = np_utils.to_categorical(test_y, outputDims)
batchSize = 128
vocabSize = len(vocab) + 1
wordDims = 100
maxlen = 7
hiddenDims = 100
w2vModel, vectorSize = pChar.load('model/pChar.model')
embeddingDim = int(vectorSize)
embeddingUnknown = [0 for i in range(embeddingDim)]
embeddingWeights = np.zeros((vocabSize + 1, embeddingDim))
for word, index in vocab.items():
if word in w2vModel:
e = w2vModel[word]
else:
print word
e = embeddingUnknown
embeddingWeights[index, :] = e
#LSTM
model = Sequential()
model.add(Embedding(output_dim = embeddingDim, input_dim = vocabSize + 1,
input_length = maxlen, mask_zero = True, weights = [embeddingWeights]))
model.add(LSTM(output_dim = hiddenDims, return_sequences = True))
model.add(LSTM(output_dim = hiddenDims, return_sequences = False))
model.add(Dropout(0.5))
model.add(Dense(outputDims))
model.add(Activation('softmax'))
model.compile(loss = 'categorical_crossentropy', optimizer = 'adam')
result = model.fit(train_X, Y_train, batch_size = batchSize,
nb_epoch = 20, validation_data = (test_X,Y_test), show_accuracy=True)
j = model.to_json()
fd = open(modelPath, 'w')
fd.write(j)
fd.close()
model.save_weights(weightPath)
return model
#Bi-directional LSTM
def loadModel(modelPath, weightPath):
fd = open(modelPath, 'r')
j = fd.read()
fd.close()
model = model_from_json(j)
model.load_weights(weightPath)
return model
# 根據(jù)輸入得到標(biāo)注推斷
def posSent(sent, model, posInfo):
(initProb, tranProb), (vocab, indexVocab) = posInfo
vec = pos.sent2vec(sent, vocab, ctxWindows = 7)
vec = np.array(vec)
probs = model.predict_proba(vec)
#classes = model.predict_classes(vec)
prob, path = viterbi.viterbi(vec, pos.corpus_tags, initProb, tranProb, probs.transpose())
ss = ''
words = sent.split()
index = -1
for word in words:
for char in word:
index += 1
ss += word + '/' + pos.tags_863[pos.corpus_tags[path[index]][:-2]][1].decode('utf-8') + ' '
#ss += word + '/' + pos.corpus_tags[path[index]][:-2] + ' '
return ss[:-1]
def posFile(fname, dstname, model, posInfo):
fd = codecs.open(fname, 'r', 'utf-8')
lines = fd.readlines()
fd.close()
fd = open(dstname, 'w')
for line in lines:
rst = posSent(line.strip(), model, posInfo)
fd.write(rst.encode('utf-8') + '\n')
fd.close()
def test():
print 'Loading vocab...'
#(X, y), (initProb, tranProb), (vocab, indexVocab) = pos.load('data/pos.train')
#posInfo = ((initProb, tranProb), (vocab, indexVocab))
#posData = (X, y)
#pos.savePosInfo('./model/pos.info', posInfo)
#pos.savePosData('./model/pos.data', posData)
posInfo = pos.loadPosInfo('./model/pos.info')
posData = pos.loadPosData('./model/pos.data')
print 'Done!'
print 'Loading model...'
#model = train(posInfo, posData, './model/pos.w2v.model', './model/pos.w2v.model.weights')
model = loadModel('./model/pos.w2v.model', './model/pos.w2v.model.weights')
#model = loadModel('./model/pos.model', './model/pos.model.weights')
print 'Done!'
print '-------------start predict----------------'
s = u'為 寂寞 的 夜空 畫 上 一個(gè) 月亮'
print posSent(s, model, posInfo)
#posFile('~/work/corpus/icwb2/testing/msr_test.utf8', './msr_test.utf8.pos', model, posInfo)
if __name__ == '__main__':
test()
中文NLP序列標(biāo)注之NER
預(yù)處理
模型
中文NLP序列標(biāo)注之DP
To be continue...
PS：全貼代碼有點(diǎn)長(zhǎng)，等我找時(shí)間再整理一下。
