TensorFlow学习笔记12-word2vec模型

为什么学习word2word2vec模型?

该模型用来学习文字的向量表示。图像和音频可以直接处理原始像素点和音频中功率谱密度的强度值，
把它们直接编码成向量数据集。但在"自然语言处理"中，对语句中的单词(Word)进行编码，无法提供
不同词汇之间的关联信息。这种"独立的、离散的"符号将导致数据稀疏，训练模型时将必须寻求更多
数据。word2vec旨在克服上述问题。

向量空间模型(VSMs)将语义近似的词汇映射为相邻的数据点，它假设出现于上下文情景中的词汇有相
类似的语义。采用该假设的研究方法分为：1. 基于计数的方法(计算词汇与邻近词汇在语料库中共同
出现的频率，并将其映射到小型且稠密的向量中)；2. 预测方法(直接从词汇的邻近词汇进行预测，利
用已学习到的小型且稠密的嵌套向量)。

Word2vec是一种可以进行高效率词嵌套学习的预测模型。其两种变体分别为：连续词袋模型（CBOW）
及Skip-Gram模型。从算法角度看，这两种方法非常相似，其区别为CBOW根据源词上下文词汇（'the
cat sits on the'）来预测目标词汇（例如，‘mat’），而Skip-Gram模型做法相反，它通过目标
词汇来预测源词汇。Skip-Gram模型采取CBOW的逆过程的动机在于：CBOW算法对于很多分布式信息
进行了平滑处理（例如将一整段上下文信息视为一个单一观察量）。很多情况下，对于小型的数据
集，这一处理是有帮助的。相形之下，Skip-Gram模型将每个“上下文-目标词汇”的组合视为一个新
观察量，这种做法在大型数据集中会更为有效。本教程余下部分将着重讲解Skip-Gram模型。

概率化语言模型

通常使用极大似然法 (ML) 进行训练，其中通过 softmax function 来最大化当提供前一个单词
(或几个单词构成的)上下文环境h(代表 "history")中，后一个单词的概率(代表 "target"):

[egin{aligned} P(w_t|h)&=softmax(score(w_t,h)) \ &=frac{exp(score(w_t,h))}{sum_{Word w' in Vocab}exp {score(w',h)}} end{aligned} ]

然而这个方法实际执行起来开销非常大，因为在每一步训练迭代中,我们需要去计算并正则化当前上下
文环境 h 中所有其他单词 w' 的概率得分。为了避免对概率模型中的所有单词进行计算，使用二分
类器(逻辑回归)在同一个上下文环境h中从k虚构的(噪声)单词(w')中区分出真正的目标单词(w_t)
。
所以其损失函数为

[J=log Q_{ heta}(D=1|w_t,h)+k_{w'sim P_{noise}}E[log Q_{ heta}(D=0|w',h)] ]

其中(Q_{ heta}(D=1|w_t,h))是数据集在上下文h，根据所学习的嵌套向量( heta)，目标单词(w)
使用逻辑回归计算得到的概率。当真实目标单词被分配到较高的概率，同时噪声单词分配到的概率很低时，
目标函数才会达到最大值。

Skip-gram模型

下面看实践。

• 数据集：the quick brown fox jumped over the lazy dog
• 定义"目标单词前一个和后一个单词作为上下文"(窗口为1)，得到数据集为：([the, brown], quick), ([quick, fox], brown), ([brown, jumped], fox), ...
• Skip-gram模型中将目标单词和上下文颠倒，得到数据集：(quick, the), (quick, brown), (brown, quick), (brown, fox), ...
• 本例中对每一个样本或batch_size(16 <= batch_size <= 512)很小的样本集(一句话或几句话)执行随机梯度下降(SGD)。

例如根据quick预测the时，随机选取了一个噪声单词为sheep，则损失函数为

[J_t=log Q_{ heta}(D=1|the,quick)+log Q_{ heta}(D=0|sheep,quick)] ]

计算(frac{partial J}{partial heta})并更新嵌套参数( heta)，将(J)最大化，直到把
真实单词和噪声单词很好地区分开。

完整代码：

# Copyright 2015 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""Basic word2vec example."""

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import collections
import math
import os
import random
import zipfile
import sys

import numpy as np
from six.moves import urllib
from six.moves import xrange  # pylint: disable=redefined-builtin
import tensorflow as tf

# Step 1: Download the data.
url = 'http://mattmahoney.net/dc/'

# 下载文件
def maybe_download(filename, expected_bytes):
  """Download a file if not present, and make sure it's the right size."""
  if not os.path.exists(filename):
    def _progress(count, block_size,total_size):
      sys.stdout.write('
>> Downloading %s %.1f%%' %(filename,float(count*block_size)/float(total_size)*100.0))
      sys.stdout.flush()
    filename, _ = urllib.request.urlretrieve(url + filename, filename,_progress)
  print()
  statinfo = os.stat(filename)
  if statinfo.st_size == expected_bytes:
    print('Found and verified', filename)
  else:
    print(statinfo.st_size)
    raise Exception(
        'Failed to verify ' + filename + '. Can you get to it with a browser?')
  return filename

filename = maybe_download('text8.zip', 31344016)


# 解压并读取文件
def read_data(filename):
  """Extract the first file enclosed in a zip file as a list of words."""
  with zipfile.ZipFile(filename) as f:
    data = tf.compat.as_str(f.read(f.namelist()[0])).split()
  return data

vocabulary = read_data(filename)
print('Data size', len(vocabulary))

# Step 2: Build the dictionary and replace rare words with UNK token.
vocabulary_size = 50000

# 建立数据集，words是所有单词的列表，n_words是想建的字典中单词的个数
def build_dataset(words, n_words):
  """Process raw inputs into a dataset."""
  #将所有低频单词设为UNK，个数先设为-1
  count = [['UNK', -1]]
  #将words集合中的单词按频数排序，将频率最高的前n_words-1个单词以及他们的出现的个数按顺序输出到count中，将频数排在n_words-1之后的单词设为UNK。同时，count的规律为索引越小，单词出现的频率越高
  count.extend(collections.Counter(words).most_common(n_words - 1))
  #建一个字典dict
  dictionary = dict()
  for word, _ in count:
    #对count中所有单词进行编号，赋予ID，由0开始，保存在字典dict中
    dictionary[word] = len(dictionary)
  #建一个列表
  data = list()
  unk_count = 0

  #对原words列表中的单词使用字典中的ID进行编号，即将单词转换成整数，储存在data列表中，同时对UNK进行计数
  for word in words:
    if word in dictionary:
      index = dictionary[word]
    else:
      index = 0  # dictionary['UNK']
      unk_count += 1
    data.append(index)
  #记录UNK个数
  count[0][1] = unk_count
  #将dictionary中的数据反转，即可以通过ID找到对应的单词，保存在reversed_dictionary中
  reversed_dictionary = dict(zip(dictionary.values(), dictionary.keys()))
  return data, count, dictionary, reversed_dictionary

data, count, dictionary, reverse_dictionary = build_dataset(vocabulary,
                                                            vocabulary_size)
del vocabulary  # Hint to reduce memory.

#输出频数最高的前5个单词
print('Most common words (+UNK)', count[:5])

print('Sample data', data[:10], [reverse_dictionary[i] for i in data[:10]])

data_index = 0

# Step 3: Function to generate a training batch for the skip-gram model.

#这个函数的功能是对数据data中的每个单词，分别与前一个单词和后一个单词生成一个batch，即[data[1],data[0]]和[data[1],data[2]]，其中当前单词data[1]存在batch中，前后单词存在labels中
def generate_batch(batch_size, num_skips, skip_window):
  global data_index                       #全局索引，在data中的位置
  assert batch_size % num_skips == 0
  assert num_skips <= 2 * skip_window
  batch = np.ndarray(shape=(batch_size), dtype=np.int32)   #建一个batch大小的数组，保存任意单词
  labels = np.ndarray(shape=(batch_size, 1), dtype=np.int32)  #建一个（batch，1）大小的二位数组，保存任意单词前一个或者后一个单词，从而形成一个pair
  span = 2 * skip_window + 1  # #窗的大小，为3，结构为[ skip_window target skip_window ]
  buffer = collections.deque(maxlen=span) #建立一个结构为双向队列的缓冲区，大小不超过3
  if data_index + span > len(data):  #如果索引超过了数据长度，则重新从数据头部开始
    data_index = 0
  buffer.extend(data[data_index:data_index + span])   #将数据index到index+3段赋值给buffer，大小刚好为span
  data_index += span  #将index向后移3位          -----------------------------------------------------------------（1）
  for i in range(batch_size // num_skips):   #128//2 四舍五入
    target = skip_window  # 将target赋值为1，即当前单词
    targets_to_avoid = [skip_window]      #将target存入targets_to_avoid中，避免重复存入
    for j in range(num_skips):
      while target in targets_to_avoid:            #选出还没出现在targets_to_avoid中的单词索引
        target = random.randint(0, span - 1)
      targets_to_avoid.append(target)               #存入targets_to_avoid
      batch[i * num_skips + j] = buffer[skip_window]    #在batch中存入当前单词
      labels[i * num_skips + j, 0] = buffer[target]      #在labels中存入当前单词前一个单词或者后一个单词
    if data_index == len(data):          #  如果到达数据尾部
      buffer[:] = data[:span]           #重新开始，将数据前三位存入buffer中，也就是说，是从数据第二个单词开始的
      data_index = span
    else:
      buffer.append(data[data_index])       #如果没有越界，则在buffer尾部插入一个新单词，同时挤出buffer中第一个单词，相当于是span的范围向后移了一位
      data_index += 1              #当前单词的索引向后移一位
  # Backtrack a little bit to avoid skipping words in the end of a batch
  data_index = (data_index + len(data) - span) % len(data)  #避免循环结束后刚好停在data尾部，以防下次运行该函数向后移动三位index时越界
  return batch, labels

batch, labels = generate_batch(batch_size=8, num_skips=2, skip_window=1)
for i in range(8):
  print(batch[i], reverse_dictionary[batch[i]],
        '->', labels[i, 0], reverse_dictionary[labels[i, 0]])

# Step 4: Build and train a skip-gram model.

batch_size = 128
embedding_size = 128  # Dimension of the embedding vector.
skip_window = 1       # How many words to consider left and right.
num_skips = 2         # How many times to reuse an input to generate a label.

# We pick a random validation set to sample nearest neighbors. Here we limit the
# validation samples to the words that have a low numeric ID, which by
# construction are also the most frequent.
valid_size = 16     # Random set of words to evaluate similarity on.
valid_window = 100  # Only pick dev samples in the head of the distribution.
valid_examples = np.random.choice(valid_window, valid_size, replace=False)
num_sampled = 64    # Number of negative examples to sample.

graph = tf.Graph()

with graph.as_default():

  # Input data.
  # 输入一个batch的训练数据，是当前单词在字典中的索引id
  train_inputs = tf.placeholder(tf.int32, shape=[batch_size])
  # 输入一个batch的训练数据的标签，是当前单词前一个或者后一个单词在字典中的索引id
  train_labels = tf.placeholder(tf.int32, shape=[batch_size, 1])
  #从字典前100个单词，即频率最高的前100个单词中，随机选出16个单词，将它们的id储存起来，作为验证集
  valid_dataset = tf.constant(valid_examples, dtype=tf.int32)

  # Ops and variables pinned to the CPU because of missing GPU implementation
  with tf.device('/cpu:0'):
    # Look up embeddings for inputs.
    # 初始化字典中每个单词的embeddings，值为-1到1的均匀分布
    embeddings = tf.Variable(
        tf.random_uniform([vocabulary_size, embedding_size], -1.0, 1.0))
    #找到训练数据对应的embeddings
    embed = tf.nn.embedding_lookup(embeddings, train_inputs)

    # Construct the variables for the NCE loss
    #初始化训练参数
    nce_weights = tf.Variable(
        tf.truncated_normal([vocabulary_size, embedding_size],
                            stddev=1.0 / math.sqrt(embedding_size))
    #初始化偏置
    nce_biases = tf.Variable(tf.zeros([vocabulary_size]))

  # Compute the average NCE loss for the batch.
  # tf.nce_loss automatically draws a new sample of the negative labels each
  # time we evaluate the loss.
  '''
  算法非常简单，根据词频或者类似词频的概率选出64个负采样v，联同正确的输入w（都是词的id），用它们在nce_weights
  对应的向量组成一个训练子集mu。
  对于训练子集中各个元素mu(i)，如果是w或者m(i)==w(w这里是输入对应的embedding)，loss(i)=log(sigmoid(w*mu(i)))
                        如果是负采样，则loss(i)=log(1-sigmoid(w*mu(i)))
  然后将所有loss加起来作为总的loss，loss越小越相似（余弦定理）
  用总的loss对各个参数求导数，来更新nce_weight以及输入的embedding
  '''
  loss = tf.reduce_mean(
      tf.nn.nce_loss(weights=nce_weights,
                     biases=nce_biases,
                     labels=train_labels,
                     inputs=embed,
                     num_sampled=num_sampled,
                     num_classes=vocabulary_size))

  # Construct the SGD optimizer using a learning rate of 1.0.
  optimizer = tf.train.GradientDescentOptimizer(1.0).minimize(loss)

  # Compute the cosine similarity between minibatch examples and all embeddings.
  #对embedding进行归一化
  norm = tf.sqrt(tf.reduce_sum(tf.square(embeddings), 1, keep_dims=True))
  normalized_embeddings = embeddings / norm
  #找到验证集中的id对应的embedding
  valid_embeddings = tf.nn.embedding_lookup(
      normalized_embeddings, valid_dataset)
  #判断验证集和整个归一化的embedding的相似性
  similarity = tf.matmul(
      valid_embeddings, normalized_embeddings, transpose_b=True)

  # Add variable initializer.
  init = tf.global_variables_initializer()

# Step 5: Begin training.
num_steps = 100001

with tf.Session(graph=graph) as session:
  # We must initialize all variables before we use them.
  init.run()
  print('Initialized')

  average_loss = 0
  for step in xrange(num_steps):
    #生成一个batch的训练数据
    batch_inputs, batch_labels = generate_batch(
        batch_size, num_skips, skip_window)
    feed_dict = {train_inputs: batch_inputs, train_labels: batch_labels}

    # We perform one update step by evaluating the optimizer op (including it
    # in the list of returned values for session.run()
    _, loss_val = session.run([optimizer, loss], feed_dict=feed_dict)
    average_loss += loss_val
    #求移动平均loss
    if step % 2000 == 0:
      if step > 0:
        average_loss /= 2000
      # The average loss is an estimate of the loss over the last 2000 batches.
      print('Average loss at step ', step, ': ', average_loss)
      average_loss = 0

    # Note that this is expensive (~20% slowdown if computed every 500 steps)
    if step % 10000 == 0:
      #每10000步评估一下验证集和整个embeddings的相似性
      #结果是验证集中每个词和字典中所有词的相似性
      sim = similarity.eval()
      #对于验证集里面的每一个词
      for i in xrange(valid_size):
        #根据id找回词
        valid_word = reverse_dictionary[valid_examples[i]]
        #因为两个向量相乘，值越小越相似（余弦定理），这里找出前8个最相似的词
        top_k = 8  # number of nearest neighbors
        #排序后输出值最小的前8个的id
        nearest = (-sim[i, :]).argsort()[1:top_k + 1]
        log_str = 'Nearest to %s:' % valid_word
        for k in xrange(top_k):
          #根据id找到对应的word
          close_word = reverse_dictionary[nearest[k]]
          log_str = '%s %s,' % (log_str, close_word)
        print(log_str)
  final_embeddings = normalized_embeddings.eval()

# Step 6: Visualize the embeddings.


def plot_with_labels(low_dim_embs, labels, filename='tsne.png'):
  assert low_dim_embs.shape[0] >= len(labels), 'More labels than embeddings'
  plt.figure(figsize=(18, 18))  # in inches
  for i, label in enumerate(labels):
    x, y = low_dim_embs[i, :]
    plt.scatter(x, y)
    plt.annotate(label,
                 xy=(x, y),
                 xytext=(5, 2),
                 textcoords='offset points',
                 ha='right',
                 va='bottom')

  plt.savefig(filename)

try:
  # pylint: disable=g-import-not-at-top
  from sklearn.manifold import TSNE
  import matplotlib.pyplot as plt

  tsne = TSNE(perplexity=30, n_components=2, init='pca', n_iter=5000, method='exact')
  plot_only = 500
  low_dim_embs = tsne.fit_transform(final_embeddings[:plot_only, :])
  labels = [reverse_dictionary[i] for i in xrange(plot_only)]
  plot_with_labels(low_dim_embs, labels)

except ImportError:
  print('Please install sklearn, matplotlib, and scipy to show embeddings.')

当前最新版本的tutorial已经更新了版本(我还没跑，你可以试试)。