[自然语言处理] word2vec自己训练词向量

word2vec代码（中文英文都可以训练）

import collections
import math
import random
import zipfile
import numpy as np
from six.moves import xrange
import tensorflow as tf


def read_data(filename):
    with zipfile.ZipFile(filename) as f:
        data = tf.compat.as_str(f.read(f.namelist()[0])).split()
    return data

words = read_data('word_embeddings/msr_unlabel.zip')
print('Data size', len(words))

vocabulary_size = 7000
def build_dataset(words, vocabulary_size):
    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
    for word in words:
        if word in dictionary:
            index = dictionary[word]
        else:
            index = 0  # dictionary['UNK']
            unk_count += 1
        data.append(index)

    count[0][1] = unk_count
    reverse_dictionary = dict(zip(dictionary.values(), dictionary.keys()))

    return data, count, dictionary, reverse_dictionary

data, count, dictionary, reverse_dictionary = build_dataset(words, vocabulary_size)

# 删除words引用
del words


#******************************   开始   ********************************************
data_index = 0
# Step 3: Function to generate a training batch for the skip-gram model.
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) % len(data)
    # 获取batch和labels
    for i in range(batch_size // num_skips):
        target = skip_window  # target label at the center of the buffer
        targets_to_avoid = [skip_window]
        # 循环2次，一个目标单词对应两个上下文单词
        for j in range(num_skips):
            while target in targets_to_avoid:
                # 可能先拿到前面的单词也可能先拿到后面的单词
                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)
    # Backtrack a little bit to avoid skipping words in the end of a batch
    # 回溯3个词。因为执行完一个batch的操作之后，data_index会往右多偏移span个位置
    data_index = (data_index + len(data) - span) % len(data)
    return batch, labels


batch_size = 128
embedding_size = 128  # 词向量维度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.
valid_size = 16  # Random set of words to evaluate similarity on.
valid_window = 100  # Only pick dev samples in the head of the distribution.
# 从0-100抽取16个整数，无放回抽样
valid_examples = np.random.choice(valid_window, valid_size, replace=False)
# 负采样样本数
num_sampled = 64  # Number of negative examples to sample.

# Step 4: Build and train a skip-gram model.
graph = tf.Graph()
with graph.as_default():
    # Input data.
    with tf.variable_scope('input'):
        train_inputs = tf.placeholder(tf.int32, shape=[batch_size],name='train_inputs')
        train_labels = tf.placeholder(tf.int32, shape=[batch_size, 1],name='train_labels')
        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'):
    # 词向量----------------------5万个词就是5万行，定义128维特征为128列************88
    # Look up embeddings for inputs.
    with tf.variable_scope('embedding'):
        embeddings = tf.Variable(
            tf.random_uniform([vocabulary_size, embedding_size], -1.0, 1.0),name='embedding')
        # embedding_lookup(params,ids)其实就是按照ids顺序返回params中的第ids行
        # 比如说，ids=[1,7,4],就是返回params中第1,7,4行。返回结果为由params的1,7,4行组成的tensor
        # 提取要训练的词-----------------------------------不是每次迭代5万个词，抽样迭代按批次就是按词的编号，把词的编号传进去
        embed = tf.nn.embedding_lookup(embeddings, train_inputs)

    with tf.variable_scope('net'):
        # Construct the variables for the noise-contrastive estimation(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]))

    # Compute the average NCE loss for the batch.
    with tf.variable_scope('loss'):
        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),name='loss')
        tf.summary.scalar('ece_loss',loss)

    # Construct the SGD optimizer using a learning rate of 1.0.
    optimizer = tf.train.GradientDescentOptimizer(1).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
    # 抽取一些常用词来测试余弦相似度
    # 如果输入的是64，那么对应的embedding是normalized_embeddings第64行的vector
    valid_embeddings = tf.nn.embedding_lookup(
        normalized_embeddings, valid_dataset)
    # valid_size == 16
    # [16,1] * [1*50000] = [16,50000]
    similarity = tf.matmul(
        valid_embeddings, normalized_embeddings, transpose_b=True)

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


num_steps = 20000
final_embeddings = []
# Step 5: 开始训练，启动session
with tf.Session(graph=graph) as session:
    print("启动session")
    merge = tf.summary.merge_all()
    init.run()
    train_writer = tf.summary.FileWriter('log')
    average_loss = 0

    for step in xrange(num_steps):
        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, summary_train = session.run([optimizer, loss, merge], feed_dict=feed_dict)
        average_loss += loss_val
        train_writer.add_summary(summary_train, step)

        # print("batch_inputs:%s  batch_labels:%s" % (batch_inputs,batch_labels))
        # batch_inputs矩阵 成对的标号   batch_labels 换行的标号      ？？


        # 每2000次迭代，打印损失值
        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

        # 每2000次迭代，随机抽一个词，并打印周围相似词
        if step % 2000 == 0:
            sim = similarity.eval()
            # 计算验证集的余弦相似度最高的词
            for i in xrange(valid_size):
                # 根据id拿到对应单词
                valid_word = reverse_dictionary[valid_examples[i]]
                top_k = 8  # number of nearest neighbors
                # 从大到小排序，排除自己本身，取前top_k个值
                nearest = (-sim[i, :]).argsort()[1:top_k + 1]
                log_str = "Nearest to %s:" % valid_word
                for k in xrange(top_k):
                    close_word = reverse_dictionary[nearest[k]]
                    log_str = "%s %s," % (log_str, close_word)
                print(log_str)

    # 训练结束得到的全部词的词向量矩阵
        final_embeddings = normalized_embeddings.eval()

    # 常规记录日志文件
    writer = tf.summary.FileWriter("log", session.graph)

# 保存词对应词向量的文件
e = open('word_embeddings/msrp_embeddings','w', encoding='utf-8')
e.write(str(vocabulary_size)+" "+str(embedding_size)+'\n')
for index in range(len(final_embeddings)):
    embedding_list = final_embeddings[index].tolist()
    # print(embedding_list)
    embedding_str = " ".join('%s' % id for id in embedding_list)
    e.write(str(reverse_dictionary[index])+" "+str(embedding_str)+'\n')

e.close()




# Step 6: Visualize the embeddings.降维画图
def plot_with_labels(low_dim_embs, labels,filename='word_embeddings/msrp_embeddings.png'):
    assert low_dim_embs.shape[0] >= len(labels), "More labels than embeddings"
    # 设置图片大小
    plt.figure(figsize=(15, 15))  # 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',
                     fontproperties = 'SimHei',
                     fontsize = 14,
                     ha='right',
                     va='bottom')
    plt.savefig(filename)

try:
    from sklearn.manifold import TSNE
    import matplotlib.pyplot as plt

    tsne = TSNE(perplexity=30, n_components=2, init='pca', n_iter=5000, method='exact')  # mac：method='exact'
    # 画500个点
    plot_only = 300
    #每个词reverse_dictionary对应每个词向量final_embeddings
    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 visualize embeddings.")

训练过程：

训练结果：

得到7000个词的128维的词向量表达（输出词的个数，维度，训练次数都可以自己根据需求设置）

词向量之间的多维空间距离，压平到二维平面。（达到距离近的词语义相近的效果）

训练次数增多效果会好些，但训练时间会长。

为了提升词嵌入的效果，也可使用预训练好的词向量，详见本篇博客：

预训练词向量中文*,英文斯坦福glove预训练的词向量下载

https://blog.csdn.net/sinat_41144773/article/details/89875130

结束。