词向量嵌入需要高效率处理大规模文本语料库。word2vec。简单方式，词送入独热编码(one-hot encoding)学习系统，长度为词汇表长度的向量，词语对应位置元素为1,其余元素为0。向量维数很高，无法刻画不同词语的语义关联。共生关系(co-occurrence)表示单词，解决语义关联，遍历大规模文本语料库，统计每个单词一定距离范围内的周围词汇，用附近词汇规范化数量表示每个词语。类似语境中词语语义相似。用PCA或类似方法降维出现向量(occurrence vector)，得到更稠密表示。性能好，追踪所有词汇共生矩阵，宽度、高度为词汇表长度。2013年,Mikolov、Tomas等提出上下文计算词表示方法，《Efficient estimation of word representations in vector space》(arXiv preprint arXiv:1301.3781(2013))。skip-gram模型，从随机表示开始，依据当前词语预测上下文词语简单分类器，误差通过分类器权值和词表示传播，对两者调整减少预测误差。大规模语料库训练模型表示赂量逼近压缩后共生向量。
数据集, 英文*转储文件包含所有页面完整修订历史，当前页面版本100GB，https://dumps.wikimedia.org/backup-index.html。
下载转储文件，提取页面词语。统计词语出现次数，构建常见词汇表。用词汇表对提取页面编码。逐行读取文件，结果立即写入磁盘。在不同步骤间保存检查点，避免程序崩溃重来。
__iter__遍历词语索引列表页面。encode获取字符串词语词汇索引。decode依据词汇索引返回字符串词语。_read_pages从*转储文件(压缩XML)提取单词，保存到页面文件，每个页面一行空格分隔的单词。bz2模块open函数读取文件。中间结果压缩处理。正则表达式捕捉任意连续字母序列或单独特殊字母。_build_vocabulary统计页面文件单词数，出现频率高词语写入文件。独热编码需要词汇表。词汇表索引编码。移除拼写错误、极不常见词语，词汇表只包含vocabulary_size - 1个最常见词语。所有不在词汇表词语<unk>标记，未出现单词词向量。
动态形成训练样本，组织到大批数据，分类器不占大量内存。skip-gram模型预测当前词语的上下文词语。遍历文本，当前词语数据，周围词语目标，创建训练样本。上下文尺寸R，每个单词生成2R样本，当前词左右各R个词。语义上下文，距离近重要，尽量少创建远上下文词语训练样本，范围[1,D=10]随机选择词上下文尺寸。依据skip-gram模型形成训练对。Numpy数组生成数值流批数据。
初始，单词表示为随机向量。分类器根据中层表示预测上下文单词当前表示。传播误差，微调权值、输入单词表示。MomentumOptimizer 模型优化，智能不足，效率高。
分类器是模型核心。噪声对比估计损失(noisecontrastive estimation loss)性能优异。softmax分类器建模。tf.nn.nce_loss 新随机向量负样本(对比样本)，近似softmax分类器。
训练模型结束，最终词向量写入文件。*语料库子集，普通CPU训练5小时，得到NumPy数组嵌入表示。完整语料库: https://dumps.wikimedia.org/enwiki/20160501/enwiki-20160501-pages-meta-current.xml.bz2 。AttrDict类等价Python dict，键可属性访问。
import bz2
import collections
import os
import re
from lxml import etree
from helpers import download
class Wikipedia:
TOKEN_REGEX = re.compile(r'[A-Za-z]+|[!?.:,()]')
def __init__(self, url, cache_dir, vocabulary_size=10000):
self._cache_dir = os.path.expanduser(cache_dir)
self._pages_path = os.path.join(self._cache_dir, 'pages.bz2')
self._vocabulary_path = os.path.join(self._cache_dir, 'vocabulary.bz2')
if not os.path.isfile(self._pages_path):
print('Read pages')
self._read_pages(url)
if not os.path.isfile(self._vocabulary_path):
print('Build vocabulary')
self._build_vocabulary(vocabulary_size)
with bz2.open(self._vocabulary_path, 'rt') as vocabulary:
print('Read vocabulary')
self._vocabulary = [x.strip() for x in vocabulary]
self._indices = {x: i for i, x in enumerate(self._vocabulary)}
def __iter__(self):
with bz2.open(self._pages_path, 'rt') as pages:
for page in pages:
words = page.strip().split()
words = [self.encode(x) for x in words]
yield words
@property
def vocabulary_size(self):
return len(self._vocabulary)
def encode(self, word):
return self._indices.get(word, 0)
def decode(self, index):
return self._vocabulary[index]
def _read_pages(self, url):
wikipedia_path = download(url, self._cache_dir)
with bz2.open(wikipedia_path) as wikipedia, \
bz2.open(self._pages_path, 'wt') as pages:
for _, element in etree.iterparse(wikipedia, tag='{*}page'):
if element.find('./{*}redirect') is not None:
continue
page = element.findtext('./{*}revision/{*}text')
words = self._tokenize(page)
pages.write(' '.join(words) + '\n')
element.clear()
def _build_vocabulary(self, vocabulary_size):
counter = collections.Counter()
with bz2.open(self._pages_path, 'rt') as pages:
for page in pages:
words = page.strip().split()
counter.update(words)
common = ['<unk>'] + counter.most_common(vocabulary_size - 1)
common = [x[0] for x in common]
with bz2.open(self._vocabulary_path, 'wt') as vocabulary:
for word in common:
vocabulary.write(word + '\n')
@classmethod
def _tokenize(cls, page):
words = cls.TOKEN_REGEX.findall(page)
words = [x.lower() for x in words]
return words
import tensorflow as tf
import numpy as np
from helpers import lazy_property
class EmbeddingModel:
def __init__(self, data, target, params):
self.data = data
self.target = target
self.params = params
self.embeddings
self.cost
self.optimize
@lazy_property
def embeddings(self):
initial = tf.random_uniform(
[self.params.vocabulary_size, self.params.embedding_size],
-1.0, 1.0)
return tf.Variable(initial)
@lazy_property
def optimize(self):
optimizer = tf.train.MomentumOptimizer(
self.params.learning_rate, self.params.momentum)
return optimizer.minimize(self.cost)
@lazy_property
def cost(self):
embedded = tf.nn.embedding_lookup(self.embeddings, self.data)
weight = tf.Variable(tf.truncated_normal(
[self.params.vocabulary_size, self.params.embedding_size],
stddev=1.0 / self.params.embedding_size ** 0.5))
bias = tf.Variable(tf.zeros([self.params.vocabulary_size]))
target = tf.expand_dims(self.target, 1)
return tf.reduce_mean(tf.nn.nce_loss(
weight, bias, embedded, target,
self.params.contrastive_examples,
self.params.vocabulary_size))
import collections
import tensorflow as tf
import numpy as np
from batched import batched
from EmbeddingModel import EmbeddingModel
from skipgrams import skipgrams
from Wikipedia import Wikipedia
from helpers import AttrDict
WIKI_DOWNLOAD_DIR = './wikipedia'
params = AttrDict(
vocabulary_size=10000,
max_context=10,
embedding_size=200,
contrastive_examples=100,
learning_rate=0.5,
momentum=0.5,
batch_size=1000,
)
data = tf.placeholder(tf.int32, [None])
target = tf.placeholder(tf.int32, [None])
model = EmbeddingModel(data, target, params)
corpus = Wikipedia(
'https://dumps.wikimedia.org/enwiki/20160501/'
'enwiki-20160501-pages-meta-current1.xml-p000000010p000030303.bz2',
WIKI_DOWNLOAD_DIR,
params.vocabulary_size)
examples = skipgrams(corpus, params.max_context)
batches = batched(examples, params.batch_size)
sess = tf.Session()
sess.run(tf.initialize_all_variables())
average = collections.deque(maxlen=100)
for index, batch in enumerate(batches):
feed_dict = {data: batch[0], target: batch[1]}
cost, _ = sess.run([model.cost, model.optimize], feed_dict)
average.append(cost)
print('{}: {:5.1f}'.format(index + 1, sum(average) / len(average)))
if index > 100000:
break
embeddings = sess.run(model.embeddings)
np.save(WIKI_DOWNLOAD_DIR + '/embeddings.npy', embeddings)
参考资料：
《面向机器智能的TensorFlow实践》
欢迎加我微信交流：qingxingfengzi
我的微信公众号:qingxingfengzigz
我老婆张幸清的微信公众号：qingqingfeifangz