-
Notifications
You must be signed in to change notification settings - Fork 2
/
AddLattice.py
252 lines (209 loc) · 9.74 KB
/
AddLattice.py
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
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
# Author : Orange
# Coding : Utf-8
# @Time : 2021/9/26 11:01 上午
# @File : AddLattice.py
import torch
from fastNLP.embeddings import StaticEmbedding
from LoadData import load_weibo_ner
from utils import Trie, extract_word_list
from functools import partial
from fastNLP.core import Vocabulary
from fastNLP import DataSet
def equip_chinese_ner_with_lexicon(datasets,
vocabs,
embeddings,
w_list,
word_embedding_path=None,
only_lexicon_in_train=False,
word_char_mix_embedding_path=None, # 字和词的embedding信息
lattice_min_freq=1,
only_train_min_freq=0):
"""
datasets {dict: 3}: {'train':{'chars':, 'target':, 'bigrams':,'seq_len'} , 'test': , 'dev':}
vocab {dict: 3}: {'char':由datasets中的 'chars'构建, 'label':,'bigram':由datasets中的'bigrams'构建}
embeddings {dict 1} : {'char': embedding}
"""
if only_lexicon_in_train:
print(f'已支持只加载在trian中出现过的词汇')
def get_skip_path(chars, w_trie):
sentence = ''.join(chars)
result = w_trie.get_lexicon(sentence)
# print(result)
return result
a = DataSet()
w_trie = Trie()
for w in w_list:
w_trie.insert(w)
if only_lexicon_in_train:
lexicon_in_train = set()
for s in datasets['train']['chars']:
lexicon_in_s = w_trie.get_lexicon(s)
for s, e, lexicon in lexicon_in_s:
lexicon_in_train.add(''.join(lexicon))
print('lexicon in train:{}'.format(len(lexicon_in_train)))
print('i.e.: {}'.format(list(lexicon_in_train)[:10]))
w_trie = Trie()
for w in lexicon_in_train:
w_trie.insert(w)
import copy
# lexicons format ；[]
for k, v in datasets.items():
v.apply_field(partial(get_skip_path, w_trie=w_trie), 'chars', 'lexicons')
v.apply_field(copy.copy, 'chars', 'raw_chars')
v.add_seq_len('lexicons', 'lex_num')
v.apply_field(lambda x: list(map(lambda y: y[0], x)), 'lexicons', 'lex_s')
v.apply_field(lambda x: list(map(lambda y: y[1], x)), 'lexicons', 'lex_e')
def concat(ins):
chars = ins['chars']
lexicons = ins['lexicons']
result = chars + list(map(lambda x: x[2], lexicons))
"""
result = ['我','爱','南','京','南京']
"""
return result
def get_pos_s(ins):
lex_s = ins['lex_s']
seq_len = ins['seq_len']
pos_s = list(range(seq_len)) + lex_s
return pos_s
def get_pos_e(ins):
lex_e = ins['lex_e']
seq_len = ins['seq_len']
pos_e = list(range(seq_len)) + lex_e
return pos_e
# v.lattice 为 [['我','爱','南','京','南京'],[...]]
for k, v in datasets.items():
v.apply(concat, new_field_name='lattice')
v.set_input('lattice')
v.apply(get_pos_s, new_field_name='pos_s')
v.apply(get_pos_e, new_field_name='pos_e')
v.set_input('pos_s', 'pos_e')
# vocabs {dict: 3} {'char':..., 'label':..., 'bigram':...}
word_vocab = Vocabulary()
word_vocab.add_word_lst(w_list)
vocabs['word'] = word_vocab
lattice_vocab = Vocabulary()
lattice_vocab.from_dataset(datasets['train'], field_name='lattice',
no_create_entry_dataset=[v for k, v in datasets.items() if k != 'train'])
vocabs['lattice'] = lattice_vocab
if word_embedding_path is not None:
word_embedding = StaticEmbedding(word_vocab, word_embedding_path, word_dropout=0)
embeddings['word'] = word_embedding
if word_char_mix_embedding_path is not None:
lattice_embedding = StaticEmbedding(lattice_vocab, word_char_mix_embedding_path, word_dropout=0.01,
min_freq=lattice_min_freq, only_train_min_freq=only_train_min_freq)
embeddings['lattice'] = lattice_embedding
else:
embeddings['lattice'] = StaticEmbedding(lattice_vocab, embedding_dim=50)
vocabs['char'].index_dataset(*(datasets.values()),
field_name='chars', new_field_name='chars')
vocabs['bigram'].index_dataset(*(datasets.values()),
field_name='bigrams', new_field_name='bigrams')
vocabs['label'].index_dataset(*(datasets.values()),
field_name='target', new_field_name='target')
vocabs['lattice'].index_dataset(*(datasets.values()),
field_name='lattice', new_field_name='lattice')
return datasets, vocabs, embeddings
if __name__ == '__main__':
# 这里 默认的 数据集路径 data/train.demo data/test.demo data/dev.demo
ds, vb, ed = load_weibo_ner('./data', unigram_embedding_path='./data/gigaword_chn.all.a2b.uni.ite50.vec',
index_token=False)
w_list = extract_word_list('/Users/orange/Desktop/Github/Flat-ner/V1/data/wordsListdemo.txt')
datasets, vocabs, embeddings = equip_chinese_ner_with_lexicon(ds, vb, ed, w_list, word_embedding_path=None)
from fastNLP.embeddings import BertEmbedding
# model path : /Users/orange/.fastNLP/embedding/bert-chinese-wwm
bert_embedding = BertEmbedding(vocabs['lattice'], model_dir_or_name='cn-wwm', requires_grad=False,
word_dropout=0.01)
inputs = [[vocabs['lattice'][word] for word in ['我', '爱', '南', '京', "南京"]]]
inputs = torch.tensor(inputs, dtype=torch.long)
print(inputs)
# inputs 输入 bert 时,会被添加 'CLS'和'SEP',即 ['[CLS]','我','爱','南','京',"南京","[SEP]"]
# "南京" 在 BertEmbedding 内部被处理时会被映射为 "南" 和"##京" 而不是 "南" 和 "京"
x = bert_embedding(inputs)
# torch.Size([1, 5, 768])
print(x)
"""
bert_embedding = BertEmbedding(vocabs['lattice'], model_dir_or_name='cn-wwm', requires_grad=False,
word_dropout=0.01)
inputs = [[vocabs['lattice'][word] for word in ['我','爱','南','京',"南京"]]]
inputs = torch.tensor(inputs, dtype=torch.long)
x = bert_embedding(inputs)
BertEmbedding会根据vocabs['lattice']构建词典, 注意 vocabs['lattice'].word2idx
BertEmbedding会仿照vocabs['lattice'].idx2word 构建自己的映射 BertEmbedding._BertWordModel.word_to_wordpieces
vocabs['lattice'].word2idx :
{'<pad>': 0,
'<unk>': 1,
'我': 2,
'爱': 3,
'市': 4,
'长': 5,
'江': 6,
'大': 7,
'桥': 8,
'市长': 9,
'长江大桥': 10,
'大桥': 11,
'重': 12,
'庆': 13,
'人': 14,
'和': 15,
'药': 16,
'店': 17,
'重庆': 18,
'人和药店': 19,
'药店': 20,
'南': 21,
'京': 22,
'南京': 23,
'南京市': 24,
'阳': 25,
'南阳': 26,
'南阳市': 27,
'武': 28,
'汉': 29,
'武汉': 30}
BertEmbedding._BertWordModel.word_to_wordpieces :
array([list([0]), list([100]), list([2769]), list([4263]), list([2356]),
list([7270]), list([3736]), list([1920]), list([3441]),
list([2356, 20327]), list([7270, 16793, 14977, 16498]),
list([1920, 16498]), list([7028]), list([2412]), list([782]),
list([1469]), list([5790]), list([2421]), list([7028, 15469]),
list([782, 14526, 18847, 15478]), list([5790, 15478]),
list([1298]), list([776]), list([1298, 13833]),
list([1298, 13833, 15413]), list([7345]), list([1298, 20402]),
list([1298, 20402, 15413]), list([3636]), list([3727]),
list([3636, 16784])], dtype=object)
故 对于 一个 输入 token 如"南"处理顺序为 先在 vocabs['lattice'].word2idx 找到其index 为 21
而 BertEmbedding._BertWordModel.word_to_wordpieces中第21个元素为 list([1298])
1298 即为 "南"在bert词表中的位置
inputs 输入 bert_embedding 时,会被添加 '[CLS]'和'[SEP]', 即 ['index([CLS])','index(我)','index(爱)','index(南)','index(京)',"index(南京)","index([SEP])"]
得到input的表征：([1, 8, 768])
tensor([[[ 0.0839, 0.2205, 0.2262, ..., 0.8225, 0.2395, -0.2320],
[ 1.0193, 0.4427, -0.5075, ..., -0.2693, 0.3188, -0.4572],
[ 1.1579, -0.1583, -0.6160, ..., 0.0205, 0.4773, -0.4224],
...,
[-0.6216, -0.0428, 0.7047, ..., -0.6202, 0.6006, -0.6023],
[-0.3761, 0.0325, -0.3896, ..., 0.9522, 0.0729, -0.2888],
[ 0.0975, 0.4757, 0.6848, ..., -0.1029, 0.1026, 0.1724]]])
去掉"[CLS]"和"[SEP]"的表征 : torch.Size([1, 6, 768])
tensor([[[ 1.0193, 0.4427, -0.5075, ..., -0.2693, 0.3188, -0.4572],
[ 1.1579, -0.1583, -0.6160, ..., 0.0205, 0.4773, -0.4224],
[ 0.5190, 0.0286, 0.1981, ..., -0.4006, 0.6613, -0.3119],
[ 0.3893, 0.6789, 0.4636, ..., 0.1320, 0.5017, -0.2525],
[-0.6216, -0.0428, 0.7047, ..., -0.6202, 0.6006, -0.6023],
[-0.3761, 0.0325, -0.3896, ..., 0.9522, 0.0729, -0.2888]]])
上面 最后两行分别对应 "南"和"##京", 所以要恢复到原来的size 即 [1, 5, 768]
这里恢复的策略有好几种 , bert_embedding 内部提供了几种 "first"(默认), "last", "max"等。
这里以 "first"为例
x：torch.Size([1, 1, 5, 768])
tensor([[[[ 1.0193, 0.4427, -0.5075, ..., -0.2693, 0.3188, -0.4572],
[ 1.1579, -0.1583, -0.6160, ..., 0.0205, 0.4773, -0.4224],
[ 0.5190, 0.0286, 0.1981, ..., -0.4006, 0.6613, -0.3119],
[ 0.3893, 0.6789, 0.4636, ..., 0.1320, 0.5017, -0.2525],
[-0.6216, -0.0428, 0.7047, ..., -0.6202, 0.6006, -0.6023]]]])
可以看到所谓的 "first"其实就是只保留每个word切割为subword下第一个token的对应的 embedding
放到这个例子就是 "南"和"##京"作为一个word切割后我们只保留第一个token 也就是 "南"对应的embedding
在我们这个例子中
['我','爱','南','京',"南京"] 在 vocabs['lattice']内 部索引表示为tensor([[ 2, 3, 21, 22, 23]])
['我','爱','南','京',"南京"] 在 bert_embedding 内部索引表示(进入bertModel前,加入了"[CLS]"和"[SEP]") tensor([[ 101, 2769, 4263, 1298, 776, 1298, 13833, 102]])
"""