深度有趣 | 13 词向量的训练

简介

使用TensorFlow实现中文词向量的训练，并完成一些简单的语义任务

回顾

在全栈课程中介绍过如何使用gensim训练中文词向量，即词嵌入（Word Embedding）

• http://study.163.com/course/courseLearn.htm?courseId=1003520028&lessonId=1004013764
• https://note.youdao.com/share/?id=2155cf875395e84d92ef80baeae7c3c0&type=notebook#/WEB5c4ea5da9d01cbfa7cd3e5bd9a748ac9

若是没有gensim则安装

pip install gensim

准备好语料，例如中文维基百科分词语料

加载库

# -*- coding: utf-8 -*-

from gensim.models import Word2Vec
from gensim.models.word2vec import LineSentence
import time

训练模型并保存，在个人笔记本上训练共耗时1403秒

t0 = int(time.time())
sentences = LineSentence('wiki.zh.word.text')
model = Word2Vec(sentences, size=128, window=5, min_count=5, workers=4)
print('训练耗时 %d s' % (int(time.time()) - t0))
model.save('gensim_128')

加载模型并使用

model = Word2Vec.load('gensim_128')
# 相关词
items = model.wv.most_similar('数学')
for i, item in enumerate(items):
	print(i, item[0], item[1])
# 语义类比
print('=' * 20)
items = model.wv.most_similar(positive=['纽约', '中国'], negative=['北京'])
for i, item in enumerate(items):
	print(i, item[0], item[1])
# 不相关词
print('=' * 20)
print(model.wv.doesnt_match(['早餐', '午饭', '晚餐', '手机']))
# 计算相关度
print('=' * 20)
print(model.wv.similarity('男人', '女人'))

原理

词向量是对词语的一种表示（representation）

• 有了词向量以后，就能够将一句话表示成一个向量序列，即一个二维Tensor
• 若是是多个长度相等的句子，则能够表示为一个三维Tensor

说白了，词向量就是一个二维矩阵，维度为V*d，V是词的总个数，d是词向量的维度

One-Hot将每一个词语表示为一个V维向量，仅当前词语对应的维度为1，其余维度为0

词嵌入将One-Hot表示的高维稀疏向量，映射为该词语对应的，低维稠密实值的词向量

词向量的训练主要有两种方法

• CBOW（Continuous Bag-of-Words）：根据上下文词语预测当前词
• Skip-Gram：根据当前词预测上下文词语

这里咱们主要讲一下Skip-Gram的原理

输入为一个词对应的整数id或One-Hot表示，通过Embedding层后获得对应的词向量，通过一层映射和softmax处理后，获得每一个词对应的输出几率

因为词汇表每每很是大，几万、几十万甚至几百万，所以直接在整个词汇表上进行多分类将会致使很是大的计算量

一个有效的解决方法是Negative Sampling，即每次随机采样一些负样本

假设词汇表大小为5W，对于某个输入词，已知对应的正确输出词，再随机从词汇表中选择N个词，这N个词恰好是正确输出词的几率很是低，所以能够认为是负样本

• 给你一张狗狗图片，判断出对应的种类名称
• 给你五张狗狗图片，判断出每一张是不是哈士奇

这样一来，就把一个5W分类的多分类问题，变成了N个二分类问题，一样提供了可学习的梯度，而且大大下降了计算量

在具体实现中，可使用Noise-Contrastive Estimation（NCE）做为损失函数，在TensorFlow中使用tf.nn.nce_loss()便可

实现

加载库和语料，一共254419行

# -*- coding: utf-8 -*-

import pickle
import numpy as np
import tensorflow as tf
import collections
from tqdm import tqdm

with open('wiki.zh.word.text', 'rb') as fr:
    lines = fr.readlines()
print('共%d行' % len(lines))
print(lines[0].decode('utf-8'))

一共有148134974个词

lines = [line.decode('utf-8') for line in lines]
words = ' '.join(lines)
words = words.replace('\n', '').split(' ')
print('共%d个词' % len(words))

定义词典

vocab_size = 50000
vocab = collections.Counter(words).most_common(vocab_size - 1)

词频统计

count = [['UNK', 0]]
count.extend(vocab)
print(count[:10])

词和id之间的相互映射

word2id = {}
id2word = {}
for i, w in enumerate(count):
    word2id[w[0]] = i
    id2word[i] = w[0]
print(id2word[100], word2id['数学'])

将语料转为id序列，一共有22385926个UNK

data = []
for i in tqdm(range(len(lines))):
    line = lines[i].strip('\n').split(' ')
    d = []
    for word in line:
        if word in word2id:
            d.append(word2id[word])
        else:
            d.append(0)
            count[0][1] += 1
    data.append(d)
print('UNK数量%d' % count[0][1])

准备训练数据

X_train = []
Y_train = []
window = 3
for i in tqdm(range(len(data))):
    d = data[i]
    for j in range(len(d)):
        start = j - window
        end = j + window
        if start < 0:
            start = 0
        if end >= len(d):
            end = len(d) - 1
        
        while start <= end:
            if start == j:
                start += 1
                continue
            else:
                X_train.append(d[j])
                Y_train.append(d[start])
                start += 1
X_train = np.squeeze(np.array(X_train))
Y_train = np.squeeze(np.array(Y_train))
Y_train = np.expand_dims(Y_train, -1)
print(X_train.shape, Y_train.shape)

定义模型参数

batch_size = 128
embedding_size = 128
valid_size = 16
valid_range = 100
valid_examples = np.random.choice(valid_range, valid_size, replace=False)
num_negative_samples = 64

定义模型

X = tf.placeholder(tf.int32, shape=[batch_size], name='X')
Y = tf.placeholder(tf.int32, shape=[batch_size, 1], name='Y')
valid = tf.placeholder(tf.int32, shape=[None], name='valid')

embeddings = tf.Variable(tf.random_uniform([vocab_size, embedding_size], -1.0, 1.0))
embed = tf.nn.embedding_lookup(embeddings, X)

nce_weights = tf.Variable(tf.truncated_normal([vocab_size, embedding_size], stddev=1.0 / np.sqrt(embedding_size)))
nce_biases = tf.Variable(tf.zeros([vocab_size]))

loss = tf.reduce_mean(tf.nn.nce_loss(weights=nce_weights, biases=nce_biases, labels=Y, inputs=embed, num_sampled=num_negative_samples, num_classes=vocab_size))

optimizer = tf.train.AdamOptimizer().minimize(loss)

将词向量归一化，并计算和给定词之间的类似度

norm = tf.sqrt(tf.reduce_sum(tf.square(embeddings), axis=1, keep_dims=True))
normalized_embeddings = embeddings / norm

valid_embeddings = tf.nn.embedding_lookup(normalized_embeddings, valid)
similarity = tf.matmul(valid_embeddings, normalized_embeddings, transpose_b=True)

训练模型

sess = tf.Session()
sess.run(tf.global_variables_initializer())

offset = 0
losses = []
for i in tqdm(range(1000000)):
    if offset + batch_size >= X_train.shape[0]:
        offset = (offset + batch_size) % X_train.shape[0]
        
    X_batch = X_train[offset: offset + batch_size]
    Y_batch = Y_train[offset: offset + batch_size]
    
    _, loss_ = sess.run([optimizer, loss], feed_dict={X: X_batch, Y: Y_batch})
    losses.append(loss_)
    
    if i % 2000 == 0 and i > 0:
        print('Iteration %d Average Loss %f' % (i, np.mean(losses)))
        losses = []
        
    if i % 10000 == 0:
        sim = sess.run(similarity, feed_dict={valid: valid_examples})
        for j in range(valid_size):
            valid_word = id2word[valid_examples[j]]
            top_k = 5
            nearests = (-sim[j, :]).argsort()[1: top_k + 1]
            s = 'Nearest to %s:' % valid_word
            for k in range(top_k):
                s += ' ' + id2word[nearests[k]]
            print(s)
            
    offset += batch_size

保存模型、最终词向量、映射字典

saver = tf.train.Saver()
saver.save(sess, './tf_128')

final_embeddings = sess.run(normalized_embeddings)
with open('tf_128.pkl', 'wb') as fw:
    pickle.dump({'embeddings': final_embeddings, 'word2id': word2id, 'id2word': id2word}, fw, protocol=4)

在单机上使用训练好的模型和词向量

加载库和获得的词向量、映射字典

# -*- coding: utf-8 -*-

import tensorflow as tf
import numpy as np
from sklearn.manifold import TSNE
import matplotlib.pyplot as plt
import pickle

with open('tf_128.pkl', 'rb') as fr:
    data = pickle.load(fr)
    final_embeddings = data['embeddings']
    word2id = data['word2id']
    id2word = data['id2word']

获取频次最高的前200个非单字词，对其词向量进行tSNE降维可视化

word_indexs = []
count = 0
plot_only = 200
for i in range(1, len(id2word)):
    if len(id2word[i]) > 1:
        word_indexs.append(i)
        count += 1
        if count == plot_only:
            break

tsne = TSNE(perplexity=30, n_components=2, init='pca', n_iter=5000)
two_d_embeddings = tsne.fit_transform(final_embeddings[word_indexs, :])
labels = [id2word[i] for i in word_indexs]

plt.figure(figsize=(15, 12))
for i, label in enumerate(labels):
    x, y = two_d_embeddings[i, :]
    plt.scatter(x, y)
    plt.annotate(label, (x, y), ha='center', va='top', fontproperties='Microsoft YaHei')
plt.savefig('词向量降维可视化.png')

能够看到，语义相关的词确实都处于相近的位置

能够加载TensorFlow模型，给valid指定一些词对应的id以获取类似词

sess = tf.Session()
sess.run(tf.global_variables_initializer())

saver = tf.train.import_meta_graph('tf_128.meta')
saver.restore(sess, tf.train.latest_checkpoint('.'))
graph = tf.get_default_graph()
valid = graph.get_tensor_by_name('valid:0')
similarity = graph.get_tensor_by_name('MatMul_1:0')

word = '数学'
sim = sess.run(similarity, feed_dict={valid: [word2id[word]]})
top_k = 10
nearests = (-sim[0, :]).argsort()[1: top_k + 1]
s = 'Nearest to %s:' % word
for k in range(top_k):
    s += ' ' + id2word[nearests[k]]
print(s)

和数学最相关的10个词

Nearest to 数学: 理论 物理学 应用 物理 科学 化学 定义 哲学 生物学 天文学

使用词向量完成其余语义任务

# 计算相关度
def cal_sim(w1, w2):
    return np.dot(final_embeddings[word2id[w1]], final_embeddings[word2id[w2]])
print(cal_sim('男人', '女人'))

# 相关词
word = '数学'
sim = [[id2word[i], cal_sim(word, id2word[i])] for i in range(len(id2word))]
sim.sort(key=lambda x:x[1], reverse=True)
top_k = 10
for i in range(top_k):
    print(sim[i + 1])

# 不相关词
def find_mismatch(words):
    vectors = [final_embeddings[word2id[word]] for word in words]
    scores = {word: np.mean([cal_sim(word, w) for w in words]) for word in words}
    scores = sorted(scores.items(), key=lambda x:x[1])
    return scores[0][0]
print(find_mismatch(['早餐', '午饭', '晚餐', '手机']))

参考

• Efficient Estimation of Word Representations in Vector Space：https://arxiv.org/abs/1301.3781
• Distributed Representations of Words and Phrases and their Compositionality：https://papers.nips.cc/paper/5021-distributed-representations-of-words-and-phrases-and-their-compositionality.pdf
• Vector Representations of Words：https://www.tensorflow.org/tutorials/word2vec

视频讲解课程

深度有趣（一）