使用深度学习进行中文自然语言处理之序列标注

深度学习简介

深度学习的资料很多，这里就不展开了讲，本文就介绍中文NLP的序列标注工作的一般方法。

机器学习与深度学习

简单来说，机器学习就是根据样本(即数据)学习得到一个模型，再根据这个模型预测的一种方法。
ML算法很多，Naive Bayes朴素贝叶斯、Decision Tree决策树、Support Vector Machine支持向量机、Logistic Regression逻辑回归、Conditional Random Field条件随机场等。
而深度学习，简单来说是一种有多层隐层的感知机。
DL也分很多模型，但一般了解Convolution Neural Network卷积神经网络、Recurrent Neural Network循环神经网络就够了(当然都要学，这里是指前期学习阶段可以侧重这两个)。
异同：ML是一种浅层学习，一般来说都由人工设计特征，而DL则用pre-training或者无监督学习来抽取特征表示，再使用监督学习来训练预测模型（当然不全都是这样）。
本文主要用于介绍DL在中文NLP的应用，所以采用了使用最为简单、方便的DL框架keras来开发，它是构建于两个非常受欢迎的DL框架theano和tensorflow之上的上层应用框架。

NLP简介

Natural Language Process自然语言处理又分为NLU自然语言理解和NLG自然语言生成。而分词、词性标注、实体识别、依存分析则是NLP的基础工作，它们都可以理解为一种序列标注工作。

序列标注工作简介

词向量简介

Word Embedding词向量方法，用实数向量来表示一个词的方法，是对One-hot Representation的一种优化。优点是低维，而且可以方便的用数学距离衡量词的词义相似度，缺点是词一多，模型就有点大，所以又有工作提出了Char Embedding方法，这种方法训练出来的模型很小，但丢失了很多的语义信息，所以又有基于分词信息的字向量的研究工作。

中文NLP序列标注之CWS

CWS简介

Chinese Word Segmentation中文分词是中文NLP的基础，一般来说中文分词有两种方法，一种是基于词典的方法，一种是基于ML或者DL的方法。CWS的发展可以参考漫话中文分词，简单来说基于词典的方法实现简单、速度快，但是对歧义和未登录词没有什么好的办法，而基于ML和DL的方法实现复杂、速度较慢，但是可以较好地应对歧义和OOV(Out-Of-Vocabulary)。
基于词典的方法应用最广的应该是正向最大匹配，而基于ML的CWS效果比较好的算法是CRF，本文主要介绍基于DL的方法，但在实际应用中应该合理的结合两种方法。

标注集与评估方法

这里采用B(Begin字为词的起始)、M(Middle字为词的中间)、E(End字为词的结束)、S(Single单字词)标注集，训练预料和评估工具采用SIGHAN中的方法，具体可以参考我的另一篇文章SIGHAN测评中文分词的方法与指标介绍。

模型

原理是采用bi-directional LSTM模型训练后对句子进行预测得到一个标注的概率，再使用Viterbi算法寻找最优的标注序列。在分词的工作中不需要加入词向量，提升效果不明显。

实现

预处理

#!/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):
    '''保存分词训练数据字典和概率'''
    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):
    '''载入分词训练数据字典和概率'''
    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):
    '''保存分词训练输入样本'''
    print('save cws data to %s'%path)
    #采用hdf5保存大矩阵效率最高
    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):
    '''载入分词训练输入样本'''
    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):
    '''文档转向量'''

    #一次性读入文件，注意内存
    fd = codecs.open(fname, 'r', 'utf-8')
    lines = fd.readlines()
    fd.close()

    #样本集
    X = []
    y = []

    #标注统计信息
    tagSize = len(corpus_tags)
    tagCnt = [0 for i in range(tagSize)]
    tagTranCnt = [[0 for i in range(tagSize)] for j in range(tagSize)]

    #遍历行
    for line in lines:
        #按空格分割
        words = line.strip('\n').split()
        #每行的分词信息
        chars = []
        tags = []
        #遍历词
        for word in words:
            #包含两个字及以上的词
            if len(word) > 1:
                #词的首字
                chars.append(word[0])
                tags.append(corpus_tags.index('B'))
                #词中间的字
                for char in word[1:(len(word) - 1)]:
                    chars.append(char)
                    tags.append(corpus_tags.index('M'))
                #词的尾字
                chars.append(word[-1])
                tags.append(corpus_tags.index('E'))
            #单字词
            else: 
                chars.append(word)
                tags.append(corpus_tags.index('S'))

        #字向量表示
        lineVecX = sent2vec2(chars, vocab, ctxWindows = 7)

        #统计标注信息
        lineVecY = []
        lastTag = -1
        for tag in tags:
            #向量
            lineVecY.append(tag)
            #lineVecY.append(corpus_tags[tag])
            #统计tag频次
            tagCnt[tag] += 1
            #统计tag转移频次
            if lastTag != -1:
                tagTranCnt[lastTag][tag] += 1
            #暂存上一次的tag
            lastTag = tag

        X.extend(lineVecX)
        y.extend(lineVecY)

    #字总频次
    charCnt = sum(tagCnt)
    #转移总频次
    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转移概率
    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']):
    
    #一次性读入文件，注意内存
    fd = codecs.open(fname, 'r', 'utf-8')
    data = fd.read()
    fd.close()

    vocab = {}
    indexVocab = []
    #遍历
    index = 0
    for char in data:
        #如果为分隔符则无需加入字典
        if char not in delimiters and char not in vocab:
            vocab[char] = index
            indexVocab.append(char)
            index += 1

    #加入未登陆新词和填充词
    vocab['retain-unknown'] = len(vocab)
    vocab['retain-padding'] = len(vocab)
    indexVocab.append('retain-unknown')
    indexVocab.append('retain-padding')
    #返回字典与索引
    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


# 根据输入得到标注推断
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'为寂寞的夜空画上一个月亮'
    #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)

#标准viterbi算法，参数为观察状态、隐藏状态、概率三元组(初始概率、转移概率、观察概率)
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时刻
    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时
    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
                    #记录最大概率
                    V[t][y] = prob
                    #记录路径
                    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():
    #隐藏状态
    hiddenstates = ('Healthy', 'Fever')
    #观察状态
    observations = ('normal', 'cold', 'dizzy')

    #初始概率
    '''
    Healthy': 0.6, 'Fever': 0.4
    '''
    start_p = [0.6, 0.4]
    #转移概率
    '''
    Healthy' : {'Healthy': 0.7, 'Fever': 0.3},
    Fever' : {'Healthy': 0.4, 'Fever': 0.6}
    '''
    trans_p = [[0.7, 0.3], [0.4, 0.6]]
    #发射概率/输出概率/观察概率
    '''
    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序列标注之POS

预处理

#!/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 = {
    #人民日报标注集：863标注集
            '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, '形容词'],
        'b' :    [1, '区别词'],
        'c' :    [2, '连词'],
        'd' :    [3, '副词'],
        'e' :    [4, '叹词'],
        'g' :    [5, '语素字'],
        'h' :    [6, '前接成分'],
        'i' :    [7, '习用语'],
        'j' :    [8, '简称'],
        'k' :    [9, '后接成分'],
        'm' :    [10, '数词'],
        'n' :    [11, '名词'],
        'nd':    [12, '方位名词'],
        'nh':    [13, '人名'],
        'ni':    [14, '团体、机构、组织的专名'],
        'nl':    [15, '处所名词'],
        'ns':    [16, '地名'],
        'nt':    [17, '时间名词'],
        'nz':    [18, '其它专名'],
        'o' :    [19, '拟声词'],
        'p' :    [20, '介词'],
        'q' :    [21, '量词'],
        'r' :    [22, '代词'],
        'u' :    [23, '助词'],
        'v' :    [24, '动词'],
        'wp':    [25, '标点'],
        'ws':    [26, '字符串'],
        'x' :    [27, '非语素字'],
    }

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):
    '''保存分词训练数据字典和概率'''
    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):
    '''载入分词训练数据字典和概率'''
    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):
    '''保存分词训练输入样本'''
    print('save pos data to %s'%path)
    #采用hdf5保存大矩阵效率最高
    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):
    '''载入分词训练输入样本'''
    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:
        #包含两个字及以上的词
        if len(word) > 1:
            #词的首字
            chars.append(word[0] + '_b')
            #词中间的字
            for char in word[1:(len(word) - 1)]:
                chars.append(char + '_m')
            #词的尾字
            chars.append(word[-1] + '_e')
        #单字词
        else: 
            chars.append(word + '_s')
    
    return sent2vec2(chars, vocab, ctxWindows = ctxWindows)

def doc2vec(fname, vocab):
    '''文档转向量'''

    #一次性读入文件，注意内存
    fd = codecs.open(fname, 'r', 'utf-8')
    lines = fd.readlines()
    fd.close()

    #样本集
    X = []
    y = []

    #标注统计信息
    tagSize = len(corpus_tags)
    tagCnt = [0 for i in range(tagSize)]
    tagTranCnt = [[0 for i in range(tagSize)] for j in range(tagSize)]

    #遍历行
    for line in lines:
        #按空格分割
        words = line.strip('\n').split()
        #每行的分词信息
        chars = []
        tags = []
        #遍历词
        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]
            #包含两个字及以上的词
            if len(word) > 1:
                #词的首字
                chars.append(word[0] + '_b')
                tags.append(corpus_tags.index(tag + '-' + 'b'))
                #词中间的字
                for char in word[1:(len(word) - 1)]:
                    chars.append(char + '_m')
                    tags.append(corpus_tags.index(tag + '-' + 'm'))
                #词的尾字
                chars.append(word[-1] + '_e')
                tags.append(corpus_tags.index(tag + '-' + 'e'))
            #单字词
            else: 
                chars.append(word + '_s')
                tags.append(corpus_tags.index(tag + '-' + 's'))

        #字向量表示
        lineVecX = sent2vec2(chars, vocab, ctxWindows = 7)

        #统计标注信息
        lineVecY = []
        lastTag = -1
        for tag in tags:
            #向量
            lineVecY.append(tag)
            #lineVecY.append(corpus_tags[tag])
            #统计tag频次
            tagCnt[tag] += 1
            #统计tag转移频次
            if lastTag != -1:
                tagTranCnt[lastTag][tag] += 1
            #暂存上一次的tag
            lastTag = tag

        X.extend(lineVecX)
        y.extend(lineVecY)

    #字总频次
    charCnt = sum(tagCnt)
    #转移总频次
    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转移概率
    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']):
    
    #一次性读入文件，注意内存
    fd = codecs.open(fname, 'r', 'utf-8')
    lines = fd.readlines()
    fd.close()

    vocab = {}
    indexVocab = []
    #遍历所有行
    index = 0
    for line in lines:
        words = line.strip().split()
        if words <= 0: continue
        #遍历所有词
        for word in words:
            word, tag = word.split('/')
            #包含两个字及以上的词
            if len(word) > 1:
                #词的首字
                char = word[0] + '_b'
                index = vocabAddChar(vocab, indexVocab, index, char)
                #词中间的字
                for char in word[1:(len(word) - 1)]:
                    char = char + '_m'
                    index = vocabAddChar(vocab, indexVocab, index, char)
                #词的尾字
                char = word[-1] + '_e'
                index = vocabAddChar(vocab, indexVocab, index, char)
            #单字词
            else: 
                char = word + '_s'
                index = vocabAddChar(vocab, indexVocab, index, char)

    #加入未登陆新词和填充词
    vocab['retain-unknown'] = len(vocab)
    vocab['retain-padding'] = len(vocab)
    indexVocab.append('retain-unknown')
    indexVocab.append('retain-padding')
    #返回字典与索引
    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


# 根据输入得到标注推断
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'为 寂寞 的 夜空 画 上 一个 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序列标注之NER

预处理

模型

中文NLP序列标注之DP

To be continue...
PS：全贴代码有点长，等我找时间再整理一下。