目录
1)前言
1.1 语言模型
1.2N-gram模型
1.3词向量表示
2)预备知识
2.1 sigmoid函数
2.2 逻辑回归
2.3贝叶斯公式
2.4 Huffman编码
3)神经网络概率语言模型
4)基于Hierarchial Sodtmax模型
4.1CBOW模型
4.2 Skip-gram模型
5)基于Negative Sampling的模型
5.1如何选取负样本
5.2 CBOW模型
5.3 Skip-gram模型
6)基于TensorFlow的word2vec实战
参考资料:
本章主要介绍了word2vec中的数学原理及基于TensorFlow的实现。
word2vec是Google在2013年开源推出的有个用于获取word vector的工具包,它简单,高效,因此引起了很多人关注。它是一个很简单的浅层结构,今天就来揭开的面纱。
从官方的介绍可以看出word2vec是一个将词表示为一个向量的工具,通过该向量表示,可以用来进行更深入的自然语言处理,比如机器翻译等。
为了理解word2vec的设计思想,我们有必要先学习一下自然语言处理的相关发展历程和基础知识。
P('我喜欢吃梨') = P('我') * P('喜欢'|'我') * P('吃'|'我','喜欢') * P('梨'|'我','喜欢','吃')
P('我喜欢吃力') = P('我') * P('喜欢'|'我') * P('吃力'|'我','喜欢')
// One-hot Representation 向量的维度是词表的大小,比如有10w个词,该向量的维度就是10w
v('足球') = [0 1 0 0 0 0 0 ......]
v('篮球') = [0 0 0 0 0 1 0 ......]
// Distributed Representation 向量的维度是某个具体的值如50
v('足球') = [0.26 0.49 -0.54 -0.08 0.16 0.76 0.33 ......]
v('篮球') = [0.31 0.54 -0.48 -0.01 0.28 0.94 0.38 ......]
这里介绍一下后面用到的知识点,包括sigmoid函数,逻辑回归,贝叶斯公式,Huffman树等。
sigmoid函数是神经网络常用的激活函数,定义为:
图像为:
导数形式为:
此时我们可以得到以下结论:
逻辑回归是我们第一个学习到的机器学习算法,这里只介绍逻辑回归的整体损失函数。
这里简单介绍霍夫曼树的构造:给定n个权值作为二叉树的n个叶子结点,可通过如下算法来构造。
2)在森林中选出两个根节点的权值最小的树合并,作为一颗新树的左右子树,且新树的根节点权值为其左右子树根节点权值之和。
3)从森林中删除选取的两棵树,并将新树加入森林。
4)重复(2)(3)步,直到森林中只剩一棵树为止,该树即为所求的霍夫曼树。
下面是一个例子:
下面是霍夫曼编码示意图:
现在开始正式介绍word2vec中用到的两个重要模型----CBOW模型和Skip-pram模型。下面是两个模型的示意图,两个模型都包含输入层、投影层、输出层。前者是在已知当前词的上下文的前提下预测当前词;后者恰恰相反,是在已知当前词的前提下预测上下文。
对于这两个模型,word2vec给出了两套框架,现在介绍基于Hierarchical Softmax的CBOW和Skip-gram模型。
CBOW模型全名为Continuous bags of words。之所以叫bags-of-words是因为输入层到投影层的操作由『拼接』变成了『叠加』,对于『叠加而言』,无所谓词的顺序,所以称为词袋模型。
对比CBOW模型结构和神经概率语言模型的模型结构,区别在于:
Skip-gram模型是已知当前词w,对其上下文中的词进行预测。
网络结构:
选取负样本需要按照一定的概率分布,Word2vec的作者们测试发现最佳的分布是3/4次幂的Unigram distribution。
来它认为语料库中所有的词出现的概率都是互相独立的。所以就是按照在语料库中随机选择,因此高频词被选中的概率大,低频词被选中的概率小,这也很符合逻辑。概率分布公式如下:
""" Word2Vec.
Implement Word2Vec algorithm to compute vector representations of words.
This example is using a small chunk of Wikipedia articles to train from.
References:
- Mikolov, Tomas et al. "Efficient Estimation of Word Representations
in Vector Space.", 2013.
Links:
- [Word2Vec] https://arxiv.org/pdf/1301.3781.pdf
Author: Aymeric Damien
Project: https://github.com/aymericdamien/TensorFlow-Examples/
"""
from __future__ import division, print_function, absolute_import
import collections
import os
import random
import urllib
import zipfile
import numpy as np
import tensorflow as tf
# Training Parameters
learning_rate = 0.1
batch_size = 128
num_steps = 3000000
display_step = 10000
eval_step = 200000
# Evaluation Parameters
eval_words = ['five', 'of', 'going', 'hardware', 'american', 'britain']
# Word2Vec Parameters
embedding_size = 200 # Dimension of the embedding vector
max_vocabulary_size = 50000 # Total number of different words in the vocabulary
min_occurrence = 10 # Remove all words that does not appears at least n times
skip_window = 3 # How many words to consider left and right
num_skips = 2 # How many times to reuse an input to generate a label
num_sampled = 64 # Number of negative examples to sample
# Download a small chunk of Wikipedia articles collection
url = 'http://mattmahoney.net/dc/text8.zip'
data_path = 'text8.zip'
if not os.path.exists(data_path):
print("Downloading the dataset... (It may take some time)")
filename, _ = urllib.urlretrieve(url, data_path)
print("Done!")
# Unzip the dataset file. Text has already been processed
with zipfile.ZipFile(data_path) as f:
text_words = f.read(f.namelist()[0]).lower().split()
# Build the dictionary and replace rare words with UNK token
count = [('UNK', -1)]
# Retrieve the most common words
count.extend(collections.Counter(text_words).most_common(max_vocabulary_size - 1))
# Remove samples with less than 'min_occurrence' occurrences
for i in range(len(count) - 1, -1, -1):
if count[i][1] < min_occurrence:
count.pop(i)
else:
# The collection is ordered, so stop when 'min_occurrence' is reached
break
# Compute the vocabulary size
vocabulary_size = len(count)
# Assign an id to each word
word2id = dict()
for i, (word, _)in enumerate(count):
word2id[word] = i
data = list()
unk_count = 0
for word in text_words:
# Retrieve a word id, or assign it index 0 ('UNK') if not in dictionary
index = word2id.get(word, 0)
if index == 0:
unk_count += 1
data.append(index)
count[0] = ('UNK', unk_count)
id2word = dict(zip(word2id.values(), word2id.keys()))
print("Words count:", len(text_words))
print("Unique words:", len(set(text_words)))
print("Vocabulary size:", vocabulary_size)
print("Most common words:", count[:10])
data_index = 0
# Generate training batch for the skip-gram model
def next_batch(batch_size, num_skips, skip_window):
global data_index
assert batch_size % num_skips == 0
assert num_skips <= 2 * skip_window
batch = np.ndarray(shape=(batch_size), dtype=np.int32)
labels = np.ndarray(shape=(batch_size, 1), dtype=np.int32)
# get window size (words left and right + current one)
span = 2 * skip_window + 1
buffer = collections.deque(maxlen=span)
if data_index + span > len(data):
data_index = 0
buffer.extend(data[data_index:data_index + span])
data_index += span
for i in range(batch_size // num_skips):
context_words = [w for w in range(span) if w != skip_window]
words_to_use = random.sample(context_words, num_skips)
for j, context_word in enumerate(words_to_use):
batch[i * num_skips + j] = buffer[skip_window]
labels[i * num_skips + j, 0] = buffer[context_word]
if data_index == len(data):
buffer.extend(data[0:span])
data_index = span
else:
buffer.append(data[data_index])
data_index += 1
# Backtrack a little bit to avoid skipping words in the end of a batch
data_index = (data_index + len(data) - span) % len(data)
return batch, labels
# Input data
X = tf.placeholder(tf.int32, shape=[None])
# Input label
Y = tf.placeholder(tf.int32, shape=[None, 1])
# Ensure the following ops & var are assigned on CPU
# (some ops are not compatible on GPU)
with tf.device('/cpu:0'):
# Create the embedding variable (each row represent a word embedding vector)
embedding = tf.Variable(tf.random_normal([vocabulary_size, embedding_size]))
# Lookup the corresponding embedding vectors for each sample in X
X_embed = tf.nn.embedding_lookup(embedding, X)
# Construct the variables for the NCE loss
nce_weights = tf.Variable(tf.random_normal([vocabulary_size, embedding_size]))
nce_biases = tf.Variable(tf.zeros([vocabulary_size]))
# Compute the average NCE loss for the batch
loss_op = tf.reduce_mean(
tf.nn.nce_loss(weights=nce_weights,
biases=nce_biases,
labels=Y,
inputs=X_embed,
num_sampled=num_sampled,
num_classes=vocabulary_size))
# Define the optimizer
optimizer = tf.train.GradientDescentOptimizer(learning_rate)
train_op = optimizer.minimize(loss_op)
# Evaluation
# Compute the cosine similarity between input data embedding and every embedding vectors
X_embed_norm = X_embed / tf.sqrt(tf.reduce_sum(tf.square(X_embed)))
embedding_norm = embedding / tf.sqrt(tf.reduce_sum(tf.square(embedding), 1, keepdims=True))
cosine_sim_op = tf.matmul(X_embed_norm, embedding_norm, transpose_b=True)
# Initialize the variables (i.e. assign their default value)
init = tf.global_variables_initializer()
with tf.Session() as sess:
# Run the initializer
sess.run(init)
# Testing data
x_test = np.array([word2id[w] for w in eval_words])
average_loss = 0
for step in xrange(1, num_steps + 1):
# Get a new batch of data
batch_x, batch_y = next_batch(batch_size, num_skips, skip_window)
# Run training op
_, loss = sess.run([train_op, loss_op], feed_dict={X: batch_x, Y: batch_y})
average_loss += loss
if step % display_step == 0 or step == 1:
if step > 1:
average_loss /= display_step
print("Step " + str(step) + ", Average Loss= " + \
"{:.4f}".format(average_loss))
average_loss = 0
# Evaluation
if step % eval_step == 0 or step == 1:
print("Evaluation...")
sim = sess.run(cosine_sim_op, feed_dict={X: x_test})
for i in xrange(len(eval_words)):
top_k = 8 # number of nearest neighbors
nearest = (-sim[i, :]).argsort()[1:top_k + 1]
log_str = '"%s" nearest neighbors:' % eval_words[i]
for k in xrange(top_k):
log_str = '%s %s,' % (log_str, id2word[nearest[k]])
print(log_str)
https://github.com/aymericdamien/TensorFlow-Examples/blob/master/examples/2_BasicModels/word2vec.py
https://blog.csdn.net/itplus/article/details/37969979
https://blog.csdn.net/mytestmy/article/details/26969149
https://www.jianshu.com/p/418f27df3968
http://www.hankcs.com/nlp/word2vec.html