ELMo词向量用于中文

    <p>10.10更新:ELMo已经由哈工大组用PyTorch重写了,并且提供了中文的预训练好的language model,可以直接使用。</p>

ELMo于今年二月由AllenNLP提出,与word2vec或GloVe不同的是其动态词向量的思想,其本质即通过训练language model,对于一句话进入到language model获得不同的词向量。根据实验可得,使用了Elmo词向量之后,许多NLP任务都有了大幅的提高。

论文:Deep contextualized word representations

AllenNLP一共release了两份ELMo的代码,一份是Pytorch版本的,另一份是Tensorflow版本的。Pytorch版本的只开放了使用预训练好的词向量的接口,但没有给出自己训练的接口,因此无法使用到中文语料中。Tensorflow版本有提供训练的代码,因此本文记录如何将ELMo用于中文语料中,但本文只记录使用到的部分,而不会分析全部的代码。

需求:
使用预训练好的词向量作为句子表示直接传入到RNN中(也就是不使用代码中默认的先过CNN),在训练完后,将模型保存,在需要用的时候load进来,对于一个特定的句子,首先将其转换成预训练的词向量,传入language model之后最终得到ELMo词向量。

准备工作:

  1. 将中文语料分词
  2. 训练好GloVe词向量或者word2vec
  3. 下载bilm-tf代码
  4. 生成词表 vocab_file (训练的时候要用到)
  5. optional:阅读Readme
  6. optional:通读bilm-tf的代码,对代码结构有一定的认识

思路:

  1. 将预训练的词向量读入
  2. 修改bilm-tf代码
    1. option部分
    2. 添加给embedding weight赋初值
    3. 添加保存embedding weight的代码
  3. 开始训练,获得checkpoint和option文件
  4. 运行脚本,获得language model的weight文件
  5. 将embedding weight保存为hdf5文件形式
  6. 运行脚本,将语料转化成ELMo embedding。

训练GloVe或word2vec

可参见我以前的博客或者网上的教程。
注意到,如果要用gensim导入GloVe训好的词向量,需要在开头添加num_word embedding_dim。 如:

获得vocab词表文件

注意到,词表文件的开头必须要有<S> </S> <UNK>,且大小写敏感。并且应当按照单词的词频降序排列。可以通过手动添加这三个特殊符号。
如:

代码:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
model=gensim.models.KeyedVectors.load_word2vec_format(
fname='/home/zhlin/GloVe/vectors.txt',binary=False
)
words=model.vocab
with open('vocab.txt','w') as f:
f.write('<S>')
f.write('\n'
f.write('</S>')
f.write('\n')
f.write('<UNK>')
f.write('\n') # bilm-tf 要求vocab有这三个符号,并且在最前面
for word in words:
f.write(word)
f.write('\n')

修改bilm-tf代码

注意到,在使用该代码之前,需要安装好相应的环境。

如果使用的是conda作为默认的Python解释器,强烈建议使用conda安装,否则可能会出现一些莫名的错误。

1
2
3
conda install tensorflow-gpu=1.4
conda install h5py
python setup.py install #应在bilm-tf的文件夹下执行该指令

然后再运行测试代码,通过说明安装成功。

修改train_elmo.py

bin文件夹下的train_elmo.py是程序的入口。
主要修改的地方:

  1. load_vocab的第二个参数应该改为None
  2. n_gpus CUDA_VISIBLE_DEVICES 根据自己需求改
  3. n_train_tokens 可改可不改,影响的是输出信息。要查看自己语料的行数,可以通过wc -l corpus.txt 查看。
  4. option的修改,将char_cnn部分都注释掉,其他根据自己需求修改

如:

修改LanguageModel类

由于我需要传入预训练好的GloVe embedding,那么还需要修改embedding部分,这部分在bilm文件夹下的training.py,进入到LanguageModel类中_build_word_embeddings函数中。注意到,由于前三个是<S> </S> <UNK>,而这三个字符在GloVe里面是没有的,因此这三个字符的embedding应当在训练的时候逐渐学习到,而正因此 embedding_weightstrainable应当设为True

如:

修改train函数

添加代码,使得在train函数的最后保存embedding文件。

训练并获得weights文件

训练需要语料文件corpus.txt,词表文件vocab.txt。

训练

cd到bilm-tf文件夹下,运行

1
2
3
4
5
export CUDA_VISIBLE_DEVICES=4
nohup python -u bin/train_elmo.py \
--train_prefix='/home/zhlin/bilm-tf/corpus.txt' \
--vocab_file /home/zhlin/bilm-tf/glove_embedding_vocab8.10/vocab.txt \
--save_dir /home/zhlin/bilm-tf/try >bilm_out.txt 2>&1 &

根据实际情况设定不同的值和路径。

运行情况:

PS:运行过程中可能会有warning:

‘list’ object has no attribute ‘name’
WARNING:tensorflow:Error encountered when serializing lstm_output_embeddings.
Type is unsupported, or the types of the items don’t match field type in CollectionDef.

应该不用担心,还是能够继续运行的,后面也不受影响。

在等待了相当长的时间后,在save_dir文件夹内生成了几个文件,其中checkpoint和options是关键,checkpoint能够进一步生成language model的weights文件,而options记录language model的参数。

获得language model的weights

接下来运行bin/dump_weights.py将checkpoint转换成hdf5文件。

1
2
3
nohup python -u  /home/zhlin/bilm-tf/bin/dump_weights.py  \
--save_dir /home/zhlin/bilm-tf/try \
--outfile /home/zhlin/bilm-tf/try/weights.hdf5 >outfile.txt 2>&1 &

其中save_dir是checkpoint和option文件保存的地址。

接下来等待程序运行:

最终获得了想要的weights和option:

将语料转化成ELMo embedding

由于我们有了vocab_file、与vocab_file一一对应的embedding h5py文件、以及language model的weights.hdf5和options.json。
接下来参考usage_token.py将一句话转化成ELMo embedding。

参考代码:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
import tensorflow as tf
import os
from bilm import TokenBatcher, BidirectionalLanguageModel, weight_layers, \
dump_token_embeddings

# Our small dataset.
raw_context = [
'这 是 测试 .',
'好的 .'
]
tokenized_context = [sentence.split() for sentence in raw_context]
tokenized_question = [
['这', '是', '什么'],
]

vocab_file='/home/zhlin/bilm-tf/glove_embedding_vocab8.10/vocab.txt'
options_file='/home/zhlin/bilm-tf/try/options.json'
weight_file='/home/zhlin/bilm-tf/try/weights.hdf5'
token_embedding_file='/home/zhlin/bilm-tf/glove_embedding_vocab8.10/vocab_embedding.hdf5'

## Now we can do inference.
# Create a TokenBatcher to map text to token ids.
batcher = TokenBatcher(vocab_file)

# Input placeholders to the biLM.
context_token_ids = tf.placeholder('int32', shape=(None, None))
question_token_ids = tf.placeholder('int32', shape=(None, None))

# Build the biLM graph.
bilm = BidirectionalLanguageModel(
options_file,
weight_file,
use_character_inputs=False,
embedding_weight_file=token_embedding_file
)

# Get ops to compute the LM embeddings.
context_embeddings_op = bilm(context_token_ids)
question_embeddings_op = bilm(question_token_ids)

elmo_context_input = weight_layers('input', context_embeddings_op, l2_coef=0.0)
with tf.variable_scope('', reuse=True):
# the reuse=True scope reuses weights from the context for the question
elmo_question_input = weight_layers(
'input', question_embeddings_op, l2_coef=0.0
)

elmo_context_output = weight_layers(
'output', context_embeddings_op, l2_coef=0.0
)
with tf.variable_scope('', reuse=True):
# the reuse=True scope reuses weights from the context for the question
elmo_question_output = weight_layers(
'output', question_embeddings_op, l2_coef=0.0
)


with tf.Session() as sess:
# It is necessary to initialize variables once before running inference.
sess.run(tf.global_variables_initializer())

# Create batches of data.
context_ids = batcher.batch_sentences(tokenized_context)
question_ids = batcher.batch_sentences(tokenized_question)

# Compute ELMo representations (here for the input only, for simplicity).
elmo_context_input_, elmo_question_input_ = sess.run(
[elmo_context_input['weighted_op'], elmo_question_input['weighted_op']],
feed_dict={context_token_ids: context_ids,
question_token_ids: question_ids}
)

print(elmo_context_input_,elmo_context_input_)

可以修改代码以适应自己的需求。

Reference

https://github.com/allenai/bilm-tf

</div>

猜你喜欢

转载自blog.csdn.net/weixin_37947156/article/details/83146349