Press "Enter" to skip to content

【NLP实战】tensorflow词向量训练实战

实战是学习一门技术最好的方式,也是深入了解一门技术唯一的方式。因此,NLP专栏计划推出一个实战专栏,让有兴趣的同学在看文章之余也可以自己动手试一试。

 

本篇介绍自然语言处理中最基础的词向量的训练。

 

作者&编辑 | 小Dream哥

 

1 语料准备

 

用于词向量训练的语料应该是已经分好词的语料,如下所示:

 

 

2 词向量训练

 

(1) 读取语料数据

 

读取数据的过程很简单,就是从压缩文件中读取上面显示的语料,得到一个列表。

 

def read_data(filename):

 

with zipfile.ZipFile(filename) as f:

 

data = tf.compat.as_str(f.read(f.namelist()[0])).split()

 

return data

 

(2) 根据语料,构建字典

 

构建字典几乎是所有NLP任务所必须的步骤。

 

def build_dataset(words):

 

count = [[‘UNK’, -1]]

 

count.extend(collections.Counter(words).most_common

 

(vocabulary_size – 1))

 

dictionary = dict()

 

for word, _ in count:

 

dictionary[word] = len(dictionary)

 

data = list()

 

unk_count = 0

 

data=[dictionary[word]  if  word in dictionary else 0 for word in words]

 

count[0][1] = unk_count

 

reverse_dictionary = dict(zip(dictionary.values(), dictionary.keys()))

 

return data, count, dictionary, reverse_dictionary

 

(3) 根据语料,获取一个batch的数据

 

这里需要解释一下,此次词向量的训练,采用的是skip gram的方式,即通过一个词,预测该 词附近的词。generate_batch函数中,skip_window表示取该词左边或右边多少个词,num_skips表示总共取多少个词。最后生成的batch数据,batch是num_skips*batch_size个词,label是中间的batch_size个词。

 

def generate_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)

 

span = 2 * skip_window + 1  # [ skip_window target skip_window ]

 

buffer = collections.deque(maxlen=span)

 

for _ in range(span):

 

buffer.append(data[data_index])

 

data_index = (data_index + 1)

 

for i in range(batch_size // num_skips):

 

target = skip_window

 

targets_to_avoid = [skip_window]

 

for j in range(num_skips):

 

while target in targets_to_avoid:

 

target = random.randint(0, span – 1)

 

targets_to_avoid.append(target)

 

batch[i * num_skips + j] = buffer[skip_window]

 

labels[i * num_skips + j, 0] = buffer[target]

 

buffer.append(data[data_index])

 

data_index = (data_index + 1) % len(data)

 

return batch, labels

 

(4) 用tensforslow训练词向量

 

首先,构造tensorflow运算图,主要包括以下几个步骤:

 

1.用palceholder先给训练数据占坑;

 

2.初始化词向量表,是一个|V|*embedding_size的矩阵,目标就是优化这个矩阵;

 

3.初始化权重;

 

4.构建损失函数,这里用NCE构建;

 

5.构建优化器;

 

6.构建变量初始化器

 

graph = tf.Graph()

 

with graph.as_default():

 

# input data

 

train_inputs = tf.placeholder(tf.int32, shape=[batch_size])

 

train_labels = tf.placeholder(tf.int32, shape=[batch_size, 1])

 

valid_dataset = tf.constant(valid_examples, dtype=tf.int32)

 

# operations and variables

 

embeddings = tf.Variable(tf.random_uniform([vocabulary_size, embedding_size], -1.0, 1.0))

 

embed = tf.nn.embedding_lookup(embeddings, train_inputs)

 

# construct the variables for the NCE loss

 

nce_weights = tf.Variable(tf.truncated_normal([vocabulary_size, embedding_size], stddev=1.0 / math.sqrt(embedding_size)))

 

nce_biases = tf.Variable(tf.zeros([vocabulary_size]))

 

ncs_loss_test=tf.nn.nce_loss(weights=nce_weights, biases=nce_biases,labels=train_labels, inputs=embed, num_sampled=num_sampled, num_classes=vocabulary_size)

 

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

 

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

 

normalized_embeddings = embeddings / norm

 

valid_embeddings = tf.nn.embedding_lookup(normalized_embeddings, valid_dataset)

 

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

 

# add variable initializer

 

init = tf.initialize_all_variables()

 

然后,开始训练词向量:

 

num_steps = 1000

 

with tf.Session(graph=graph) as session:

 

# we must initialize all variables before using them

 

init.run()

 

print(‘initialized.’)

 

# loop through all training steps and keep track of loss

 

average_loss = 0

 

for step in range(num_steps):

 

# generate a minibatch of training data

 

batch_inputs, batch_labels = generate_batch(batch_size, num_skips, skip_window)

 

feed_dict = {train_inputs: batch_inputs, train_labels: batch_labels}

 

# we perform a single update step by evaluating the optimizer operation (including it

 

_, loss_val,ncs_loss_ = session.run([optimizer, loss,ncs_loss_test], feed_dict=feed_dict)

 

average_loss += loss_val

 

final_embeddings = normalized_embeddings.eval()

 

print(final_embeddings)

 

(5) 保存词向量

 

将训练好的词向量写到文件中备用。

f

inal_embeddings = normalized_embeddings.eval()

print(final_embeddings)

fp=open(‘vector.txt’,’w’,encoding=’utf8′)

for k,v in reverse_dictionary.items():

t=tuple(final_embeddings[k])

s=”

for i in t:

i=str(i)

s+=i+” ”

fp.write(v+” “+s+”\n”)

fp.close()

 

最后,我们将词向量写到了vector.txt里面,得到了一份很大的词向量表,我们看看它长成什幺样子:

 

 

可以看到,词向量就是将每个中文词用一个向量来表示,整个词表及其词向量构成了这份词向量表。

 

这里留一个作业,读者可以自己试一下,从表中读取出来几个词的向量,计算出来他们的相似度,看训练出来的词向量质量如何。

 

至此本文介绍了如何利用tensorflow平台自己写代码,训练一份自己想要的词向量,代码在我们有三AI的github可以

 

https://github.com/longpeng2008/yousan.ai/tree/master/natural_language_processing

 

找到word2vec文件夹,执行python3 w2v_skip_gram.py就可以运行,训练词向量了。

 

总结

 

这里讲述了词向量的具体训练过程,相关的原理在我之前的系列文章里有详细的讲述,感兴趣的同学可以好好看一下:

 

Be First to Comment

发表评论

电子邮件地址不会被公开。 必填项已用*标注