tensorflow 实践（一）使用神经网络做中文情感分析

本文使用哈工大做文本预处理； 两层隐层神经网络；
后注：不是标准的ann，做了去停用词和词性筛选，没有端到端。

# -*- coding: utf-8 -*-
# @bref :使用tensorflow做中文情感分析
import numpy as np
import tensorflow as tf
import random
from sklearn.feature_extraction.text import CountVectorizer
import os
import traceback

real_dir_path = os.path.split(os.path.realpath(__file__))[0]
pos_file = os.path.join(real_dir_path, 'data/pos_bak.txt')
neg_file = os.path.join(real_dir_path, 'data/neg_bak.txt')

#使用哈工大分词和词性标注
from pyltp import Segmentor, Postagger
seg = Segmentor()
seg.load('/root/git/ltp_data/cws.model')
poser = Postagger()
poser.load('/root/git/ltp_data/pos.model')
real_dir_path = os.path.split(os.path.realpath(__file__))[0] #文件所在路径
stop_words_file = os.path.join(real_dir_path, '../util/stopwords.txt')
#定义允许的词性
allow_pos_ltp = ('a', 'i', 'j', 'n', 'nh', 'ni', 'nl', 'ns', 'nt', 'nz', 'v', 'ws')

#分词、去除停用词、词性筛选
def cut_stopword_pos(s):
    words = seg.segment(''.join(s.split()))
    poses = poser.postag(words)
    stopwords = {}.fromkeys([line.rstrip() for line in open(stop_words_file)])
    sentence = []
    for i, pos in enumerate(poses):
        if (pos in allow_pos_ltp) and (words[i] not in stopwords):
            sentence.append(words[i])
    return sentence

def create_vocab(pos_file, neg_file):
    def process_file(file_path):
        with open(file_path, 'r') as f:
            v = []
            lines = f.readlines()
            for line in lines:
                sentence = cut_stopword_pos(line)
                v.append(' '.join(sentence))
            return v
    sent = process_file(pos_file)
    sent += process_file(neg_file)
    tf_v = CountVectorizer(max_df=0.9, min_df=1)
    tf = tf_v.fit_transform(sent)
    #print tf_v.vocabulary_
    return tf_v.vocabulary_.keys()

#获取词汇
vocab = create_vocab(pos_file, neg_file)

#依据词汇将评论转化为向量
def normalize_dataset(vocab):
    dataset = []
    # vocab:词汇表; review:评论; clf:评论对应的分类, [0, 1]表示负面评论,[1, 0]表示正面
    def string_to_vector(vocab, review, clf):
        words = cut_stopword_pos(review) # list of str
        features = np.zeros(len(vocab))
        for w in words:
            if w.decode('utf-8') in vocab:
                features[vocab.index(w.decode('utf-8'))] = 1
        return [features, clf]
    with open(pos_file, 'r') as f:
        lines = f.readlines()
        for line in lines:
            one_sample = string_to_vector(vocab, line, [1, 0])
            dataset.append(one_sample)
    with open(neg_file, 'r') as f:
        lines = f.readlines()
        for line in lines:
            one_sample = string_to_vector(vocab, line, [0, 1])
            dataset.append(one_sample)
    return dataset

dataset = normalize_dataset(vocab)
random.shuffle(dataset)  #打乱顺序

#取样本的10%作为测试数据
test_size = int(len(dataset) * 0.1)
dataset = np.array(dataset)
train_dataset = dataset[:-test_size]
test_dataset = dataset[-test_size:]
print 'test_size = {}'.format(test_size)
#print 'size of train_dataset is {}'.format(train_dataset)

#Feed-forward nueral network
#定义每个层有多少个神经元
n_input_layer = len(vocab)   #输入层每个神经元代表一个term

n_layer_1 = 1000  #hiden layer
n_layer_2 = 1000 # hiden layer
n_output_layer = 2

#定义待训练的神经网络
def neural_netword(data):
    #定义第一层神经元的w和b, random_normal定义服从正态分布的随机变量
    layer_1_w_b = {'w_':tf.Variable(tf.random_normal([n_input_layer, n_layer_1])), 'b_':tf.Variable(tf.random_normal([n_layer_1]))}
    layer_2_w_b = {'w_':tf.Variable(tf.random_normal([n_layer_1, n_layer_2])), 'b_':tf.Variable(tf.random_normal([n_layer_2]))}
    layer_output_w_b = {'w_':tf.Variable(tf.random_normal([n_layer_2, n_output_layer])), 'b_':tf.Variable(tf.random_normal([n_output_layer]))}

    layer_1 = tf.add(tf.matmul(data, layer_1_w_b['w_']), layer_1_w_b['b_'])
    layer_1 = tf.nn.relu(layer_1) #relu做激活函数
    layer_2 = tf.add(tf.matmul(layer_1, layer_2_w_b['w_']), layer_2_w_b['b_'])
    layer_2 = tf.nn.relu(layer_2)
    layer_output = tf.add(tf.matmul(layer_2, layer_output_w_b['w_']), layer_output_w_b['b_'])
    return layer_output

batch_size = 50
X = tf.placeholder('float', [None, n_input_layer])  #None表示样本数量任意; 每个样本纬度是term数量
Y = tf.placeholder('float')

#使用数据训练神经网络
def train_neural_network(X, Y):
    predict = neural_netword(X)
    #cost func是输出层softmax的cross entropy的平均值。 将softmax 放在此处而非nn中是为了效率.
    cost_func = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=predict, labels=Y))
    #设置优化器
    optimizer = tf.train.AdamOptimizer().minimize(cost_func)

    epochs = 13  #epoch本意是时代、纪, 这里是迭代周期
    with tf.Session() as session:
        session.run(tf.initialize_all_variables()) #初始化所有变量,包括w,b

        random.shuffle(train_dataset)
        train_x = train_dataset[:, 0] #每一行的features;
        train_y = train_dataset[:, 1] #每一行的label
        print 'size of train_x is {}'.format(len(train_x))
        for epoch in range(epochs):
            epoch_loss = 0 #每个周期的loss
            i = 0
            while i < len(train_x):
                start = i
                end = i + batch_size

                batch_x = train_x[start:end]
                batch_y = train_y[start:end]
                #run的第一个参数fetches可以是单个,也可以是多个。 返回值是fetches的返回值。
                #此处因为要打印cost,所以cost_func也在fetches中
                _, c = session.run([optimizer, cost_func], feed_dict={X:list(batch_x), Y:list(batch_y)})
                epoch_loss += c
                i = end
            print(epoch, ' : ', epoch_loss)

        #评估模型
        test_x = test_dataset[:, 0]
        test_y = test_dataset[:, 1]
        #argmax能给出某个tensor对象在某一维上的其数据最大值所在的索引值, 这里是索引值的list。tf.equal用于检测匹配,返回bool型的list
        correct = tf.equal(tf.argmax(predict, 1), tf.argmax(Y, 1))
        #tf.cast 可以将[True, False, True] 转化为[1, 0, 1]
        #reduce_mean用于在某一维上计算平均值, 未指定纬度则计算所有元素
        accurqcy = tf.reduce_mean(tf.cast(correct, 'float'))
        print('准确率: {}'.format(accurqcy.eval({X:list(test_x), Y:list(test_y)})))
        #等价: print session.run(accuracy, feed_dict={X:list(test_x), Y:list(test_y)})

train_neural_network(X, Y)

最终的执行显示：

size of train_x is 31612
(0, ' : ', 105508.38228607178)
(1, ' : ', 11773.463727131188)
(2, ' : ', 4551.4978754326503)
(3, ' : ', 3576.6907950473492)
(4, ' : ', 3144.6771814899175)
(5, ' : ', 2911.1803286887775)
(6, ' : ', 2691.8284285693276)
(7, ' : ', 2651.9982114042473)
(8, ' : ', 2882.4479921576026)
(9, ' : ', 2665.3818837262743)
(10, ' : ', 2551.3030235993206)
(11, ' : ', 2838.3546982686303)
(12, ' : ', 2770.5539811982608)
准确率: 0.828587830067