论文分享-- >word2Vec论文总结

博客内容将首发在微信公众号"跟我一起读论文啦啦"，上面会定期分享机器学习、深度学习、数据挖掘、自然语言处理等高质量论文，欢迎关注！

一直以来，对$word2vec$，以及对 $tensorflow$ 里面的$wordEmbedding$底层实现原理一直模糊不清，由此决心阅读$word2Vec$的两篇原始论文，$EfficientEstimationofWordRepresentationsinVectorSpace$，$DistributedRepresentationsofWordsandPhrasesandtheirCompositionality$，看完以后还是有点半懂半不懂的感觉，于是又结合网上的一些比较好的讲解（Word2Vec Tutorial - The Skip-Gram Model)，以及开源的实现代码理解了一遍，在此总结一下。

下面主要以 $skip-gram$ 模型来介绍$word2Vec$。

word2vec工作流程

1. $word2Vec$只是一个 三层 的神经网络。
2. 喂给模型一个$word$，然后用来预测它周边的词。
3. 然后去掉最后一层，只保存$input\_layer$ 和 $hidden\_layer$。
4. 从词表中选取一个词，喂给模型，在$hidden\_layer$ 将会给出该词的$embeddingrepesentation$。

import numpy as np
import tensorflow as tf
corpus_raw = 'He is the king . The king is royal . She is the royal  queen '
# convert to lower case
corpus_raw = corpus_raw.lower()

上述代码非常简单和易懂，现在我们需要获取$inputoutputpair$，假设我们现在有这样一个任务，喂给模型一个词，我们需要获取它周边的词，举例来说，就是获取该词前$n$个和后$n$个词，那么这个$n$就是代码中的$window\_size$，例如下图：

注意：如果这个词是一个句子的开头或结尾，$window$ 忽略窗外的词。

我们需要对文本数据进行一个简单的预处理，创建一个$word2int$的字典和$int2word$的字典。

words = []
for word in corpus_raw.split():
    if word != '.': # because we don't want to treat . as a word
        words.append(word)
words = set(words) # so that all duplicate words are removed
word2int = {}
int2word = {}
vocab_size = len(words) # gives the total number of unique words
for i,word in enumerate(words):
    word2int[word] = i
    int2word[i] = word

来看看这个字典有啥效果：

print(word2int['queen'])
-> 42 (say)
print(int2word[42])
-> 'queen'

好，现在可以获取训练数据啦

data = []
WINDOW_SIZE = 2
for sentence in sentences:
    for word_index, word in enumerate(sentence):
        for nb_word in sentence[max(word_index - WINDOW_SIZE, 0) : min(word_index + WINDOW_SIZE, len(sentence)) + 1] : 
            if nb_word != word:
                data.append([word, nb_word])

上述代码就是切句子，然后切词，得出的一个个训练样本$[word,nb\_word]$，其中$word$就是模型输入，$nb\_word$就是该词周边的某个单词。

把$data$打印出来看看？

print(data)
[['he', 'is'],
 ['he', 'the'],
 ['is', 'he'],
 ['is', 'the'],
 ['is', 'king'],
 ['the', 'he'],
 ['the', 'is'],
 ['the', 'king'],
.
.
.
]

现在我们有了训练数据了，但是需要将它转成模型可读可理解的形式，这时，上面的$word2int$字典的作用就来了。

来，我们更进一步的对$word$进行处理，并使其转成$one-hot$向量

i.e., 
say we have a vocabulary of 3 words : pen, pineapple, apple
where 
word2int['pen'] -> 0 -> [1 0 0]
word2int['pineapple'] -> 1 -> [0 1 0]
word2int['apple'] -> 2 -> [0 0 1]

那么为啥是$one-hot$特征呢？稍后将解释。

# function to convert numbers to one hot vectors
def to_one_hot(data_point_index, vocab_size):
    temp = np.zeros(vocab_size)
    temp[data_point_index] = 1
    return temp
x_train = [] # input word
y_train = [] # output word
for data_word in data:
    x_train.append(to_one_hot(word2int[ data_word[0] ], vocab_size))
    y_train.append(to_one_hot(word2int[ data_word[1] ], vocab_size))
# convert them to numpy arrays
x_train = np.asarray(x_train)
y_train = np.asarray(y_train)

利用$tensorflow$建立模型

# making placeholders for x_train and y_train
x = tf.placeholder(tf.float32, shape=(None, vocab_size))
y_label = tf.placeholder(tf.float32, shape=(None, vocab_size))

由上图，我们可以看出，我们将$input$转换成$embedding\_representation$，并且将$vocabSize$维度降低到设定的$embedding\_dim$。

EMBEDDING_DIM = 5 # you can choose your own number
W1 = tf.Variable(tf.random_normal([vocab_size, EMBEDDING_DIM]))
b1 = tf.Variable(tf.random_normal([EMBEDDING_DIM])) #bias
hidden_representation = tf.add(tf.matmul(x,W1), b1)

接下来，我们需要使用$softmax$函数来预测该$word$周边的词。

W2 = tf.Variable(tf.random_normal([EMBEDDING_DIM, vocab_size]))
b2 = tf.Variable(tf.random_normal([vocab_size]))
prediction = tf.nn.softmax(tf.add( tf.matmul(hidden_representation, W2), b2))

所以整体的过程如下：

input_one_hot  --->  embedded repr. ---> predicted_neighbour_prob
predicted_prob will be compared against a one hot vector to correct it.

好了，来看看怎么训这个模型

sess = tf.Session()
init = tf.global_variables_initializer()
sess.run(init) #make sure you do this!
# define the loss function:
cross_entropy_loss = tf.reduce_mean(-tf.reduce_sum(y_label * tf.log(prediction), reduction_indices=[1]))
# define the training step:
train_step = tf.train.GradientDescentOptimizer(0.1).minimize(cross_entropy_loss)
n_iters = 10000
# train for n_iter iterations
for _ in range(n_iters):
    sess.run(train_step, feed_dict={x: x_train, y_label: y_train})
    print('loss is : ', sess.run(cross_entropy_loss, feed_dict={x: x_train, y_label: y_train}))

在训的过程中，你可以看到$loss$的变化：

loss is :  2.73213
loss is :  2.30519
loss is :  2.11106
loss is :  1.9916
loss is :  1.90923
loss is :  1.84837
loss is :  1.80133
loss is :  1.76381
loss is :  1.73312
loss is :  1.70745
loss is :  1.68556
loss is :  1.66654
loss is :  1.64975
loss is :  1.63472
loss is :  1.62112
loss is :  1.6087
loss is :  1.59725
loss is :  1.58664
loss is :  1.57676
loss is :  1.56751
loss is :  1.55882
loss is :  1.55064
loss is :  1.54291
loss is :  1.53559
loss is :  1.52865
loss is :  1.52206
loss is :  1.51578
loss is :  1.50979
loss is :  1.50408
loss is :  1.49861
.
.
.

最终$loss$会收敛，即使其$accuracy$不能达到很高的水平，我们并不$care$这点，我们最终的目的是获取较好的$W1$和$b1$，也就是$hidden\_repesentation$。

为什么是$one-hot$？

当我们用$one-hot$向量乘以$W1$时，获取的是$W1$矩阵的某一行，所以$W1$扮演的是一个$lookuptable$。

在我们这个代码例子中，可以看看$"queen"$在$Ｗ1$中的$repesetation$。

print(vectors[ word2int['queen'] ])
# say here word2int['queen'] is 2
-> 
[-0.69424796 -1.67628145  3.07313657 -1.14802659 -1.2207377 ]

给定一个向量，我们可以获取与其最近的向量

def euclidean_dist(vec1, vec2):
    return np.sqrt(np.sum((vec1-vec2)**2))

def find_closest(word_index, vectors):
    min_dist = 10000 # to act like positive infinity
    min_index = -1
    query_vector = vectors[word_index]
    for index, vector in enumerate(vectors):
        if euclidean_dist(vector, query_vector) < min_dist and not np.array_equal(vector, query_vector):
            min_dist = euclidean_dist(vector, query_vector)
            min_index = index
    return min_index

我们来看看，与$"king"、"queen"、"royal"$最近的词：

print(int2word[find_closest(word2int['king'], vectors)])
print(int2word[find_closest(word2int['queen'], vectors)])
print(int2word[find_closest(word2int['royal'], vectors)])
->
queen
king
he

进阶

上面总结的主要是第一篇论文$EfficientEstimationofWordRepresentationsinVectorSpace$内的内容，虽然只是一个三层的神经网络，但是在海量训练数据的情况下，需要极大的计算资源来支撑整个过程，举例来说，我们设定的$embedding\_size=300$时，而$vocab\_size=10,000$时，这时$W1$矩阵的维度就达到了$10,000\ast 300=3million$！！，这个时候再用$SGD$来优化训练过程就显得十分缓慢，但是有时候你必须使用大量的数据来训练模型来避免过拟合。论文$DistributedRepresentationsofWordsandPhrasesandtheirCompositionality$介绍了几种解决办法。

• 采用下采样来降低训练样本数量
  在$tensorflow$里面实现的$word2Vec$，$vocab\_szie$并不是所有的$word$的数量，而且先统计了所有$word$的出现频次，然后选取出现频次最高的前$50000$的词作为词袋。具体操作请看代码 tensorflow/examples/tutorials/word2vec/word2vec_basic.py，其余的词用$unk$代替。

• 采用一种所谓的"负采样"的操作，这种操作每次可以让一个样本只更新权重矩阵中一小部分，减小训练过程中的计算压力。
  举例来说：一个$inputoutputpair$ 如：$(“fox”,“quick”)$，由上面的分析可知，其$truelabel$为一个$one-hot$向量，并且该向量只是在$quick$的位置为1，其余的位置均为0，并且该向量的长度为$vocabsize$，由此每个样本都缓慢能更新权重矩阵，而"负采样"操作只是随机选择其余的部分$word$，使得其在$truelabel$的位置为0，那么我们只更新对应位置的权重。例如我们如果选择负采样数量为５，则选取５个其余的$word$，使其对应的$output$为0，这个时候$output$只是６个神经元，本来我们一次需要更新$300\ast 10,000$参数，进行负采样操作以后只需要更新$300\ast 6＝1800$个参数。

• Hierarchical Softmax 是NLP中常用方法，详情可以查看Hierarchical Softmax 。其主要思想是以词频构建Huffman树，树的叶子节点为词表中的词，相应的高频词距离根结点更近。当需要计算生成某个词的概率时，不需要对所有词进行概率计算，而是选择在Huffman树中从根结点到该词所在结点的路径进行计算，得到生成该词的概率，时间复杂度从 O(N) 降低到 O(logN)（N个结点，则树的深度logN）

个人总结

• seq2seq模型，输入处都会乘以$embedding\_matrix$，输出处都会乘以$embedding\_matri{x}^T$，这两个embedding矩阵有时会共享，有时则不会。我认为$word2Vec$ 其实就是$seq2seq$ 模型的原型，只不过应用到了不同的复杂场景中，根据场景需要，在内部加了$Attention$ 等机制，大致框架依然是$word2Vec$。
• $word2Vec$ 是当前自然语言处理领域的最基础知识，深刻理解$word2vec$ 原理非常重要。

个人感觉$word2Vec$了解到这个程度差不多了。

完整代码：

import tensorflow as tf
import numpy as np

corpus_raw = 'He is the king . The king is royal . She is the royal  queen '

# convert to lower case
corpus_raw = corpus_raw.lower()

words = []
for word in corpus_raw.split():
    if word != '.': # because we don't want to treat . as a word
        words.append(word)

words = set(words) # so that all duplicate words are removed
word2int = {}
int2word = {}
vocab_size = len(words) # gives the total number of unique words

for i,word in enumerate(words):
    word2int[word] = i
    int2word[i] = word

# raw sentences is a list of sentences.
raw_sentences = corpus_raw.split('.')
sentences = []
for sentence in raw_sentences:
    sentences.append(sentence.split())

WINDOW_SIZE = 2

data = []
for sentence in sentences:
    for word_index, word in enumerate(sentence):
        for nb_word in sentence[max(word_index - WINDOW_SIZE, 0) : min(word_index + WINDOW_SIZE, len(sentence)) + 1] : 
            if nb_word != word:
                data.append([word, nb_word])

# function to convert numbers to one hot vectors
def to_one_hot(data_point_index, vocab_size):
    temp = np.zeros(vocab_size)
    temp[data_point_index] = 1
    return temp

x_train = [] # input word
y_train = [] # output word

for data_word in data:
    x_train.append(to_one_hot(word2int[ data_word[0] ], vocab_size))
    y_train.append(to_one_hot(word2int[ data_word[1] ], vocab_size))

# convert them to numpy arrays
x_train = np.asarray(x_train)
y_train = np.asarray(y_train)

# making placeholders for x_train and y_train
x = tf.placeholder(tf.float32, shape=(None, vocab_size))
y_label = tf.placeholder(tf.float32, shape=(None, vocab_size))

EMBEDDING_DIM = 5 # you can choose your own number
W1 = tf.Variable(tf.random_normal([vocab_size, EMBEDDING_DIM]))
b1 = tf.Variable(tf.random_normal([EMBEDDING_DIM])) #bias
hidden_representation = tf.add(tf.matmul(x,W1), b1)

W2 = tf.Variable(tf.random_normal([EMBEDDING_DIM, vocab_size]))
b2 = tf.Variable(tf.random_normal([vocab_size]))
prediction = tf.nn.softmax(tf.add( tf.matmul(hidden_representation, W2), b2))


sess = tf.Session()
init = tf.global_variables_initializer()
sess.run(init) #make sure you do this!

# define the loss function:
cross_entropy_loss = tf.reduce_mean(-tf.reduce_sum(y_label * tf.log(prediction), reduction_indices=[1]))

# define the training step:
train_step = tf.train.GradientDescentOptimizer(0.1).minimize(cross_entropy_loss)

n_iters = 10000
# train for n_iter iterations

for _ in range(n_iters):
    sess.run(train_step, feed_dict={x: x_train, y_label: y_train})
    print('loss is : ', sess.run(cross_entropy_loss, feed_dict={x: x_train, y_label: y_train}))

vectors = sess.run(W1 + b1)

def euclidean_dist(vec1, vec2):
    return np.sqrt(np.sum((vec1-vec2)**2))

def find_closest(word_index, vectors):
    min_dist = 10000 # to act like positive infinity
    min_index = -1
    query_vector = vectors[word_index]
    for index, vector in enumerate(vectors):
        if euclidean_dist(vector, query_vector) < min_dist and not np.array_equal(vector, query_vector):
            min_dist = euclidean_dist(vector, query_vector)
            min_index = index
    return min_index


from sklearn.manifold import TSNE

model = TSNE(n_components=2, random_state=0)
np.set_printoptions(suppress=True)
vectors = model.fit_transform(vectors) 

from sklearn import preprocessing

normalizer = preprocessing.Normalizer()
vectors =  normalizer.fit_transform(vectors, 'l2')

print(vectors)

import matplotlib.pyplot as plt


fig, ax = plt.subplots()
print(words)
for word in words:
    print(word, vectors[word2int[word]][1])
    ax.annotate(word, (vectors[word2int[word]][0],vectors[word2int[word]][1] ))
plt.show()