from fastNLP.io.loader import CSVLoader
bundle = CSVLoader(headers=['raw_words1', 'raw_words2', 'target'],
sep='\t').load(sys.argv[1])
#####test
import jieba
from fastNLP.core import Vocabulary
bundle.apply(lambda line: jieba.lcut(line['raw_words1']) + ['[SEP]'] + jieba.
lcut(line['raw_words2']),
new_field_name='words')
bundle.apply(lambda line: len(line['words']), new_field_name='seq_len')
bundle.apply(lambda line: 1, new_field_name='target')
vocab = Vocabulary()
vocab.from_dataset(bundle.get_dataset("train"),
field_name='words',
no_create_entry_dataset=[
bundle.get_dataset("test"),
bundle.get_dataset("dev")
])
vocab.index_dataset(bundle.get_dataset("train"), field_name='words')
vocab.index_dataset(bundle.get_dataset("test"), field_name='words')
vocab.index_dataset(bundle.get_dataset("dev"), field_name='words')
# establish the model
from fastNLP import Const
import torch
from fastNLP.models import BertForSentenceMatching
from fastNLP.embeddings.bert_embedding import BertEmbedding
embed = BertEmbedding(vocab, model_dir_or_name='cn-base', requires_grad=False)
#pad the input array
bundle.set_pad_val("words", 0)
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
number_normalized=False,
lattice_min_freq=1,
only_train_min_freq=0):
from fastNLP.core import Vocabulary
def normalize_char(inp):
result = []
for c in inp:
if c.isdigit():
result.append('0')
else:
result.append(c)
return result
def normalize_bigram(inp):
result = []
for bi in inp:
tmp = bi
if tmp[0].isdigit():
tmp = '0' + tmp[:1]
if tmp[1].isdigit():
tmp = tmp[0] + '0'
result.append(tmp)
return result
if number_normalized == 3:
for k, v in datasets.items():
v.apply_field(normalize_char, 'chars', 'chars')
vocabs['char'] = Vocabulary()
vocabs['char'].from_dataset(
datasets['train'],
field_name='chars',
no_create_entry_dataset=[datasets['dev'], datasets['test']])
for k, v in datasets.items():
v.apply_field(normalize_bigram, 'bigrams', 'bigrams')
vocabs['bigram'] = Vocabulary()
vocabs['bigram'].from_dataset(
datasets['train'],
field_name='bigrams',
no_create_entry_dataset=[datasets['dev'], datasets['test']])
if only_lexicon_in_train:
print('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
from V0.utils_ import Trie
from functools import partial
from fastNLP.core import Vocabulary
from fastNLP_module import StaticEmbedding
from fastNLP import DataSet
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
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')
if number_normalized == 1:
for k, v in datasets.items():
v.apply_field(normalize_char, 'chars', 'chars')
vocabs['char'] = Vocabulary()
vocabs['char'].from_dataset(
datasets['train'],
field_name='chars',
no_create_entry_dataset=[datasets['dev'], datasets['test']])
if number_normalized == 2:
for k, v in datasets.items():
v.apply_field(normalize_char, 'chars', 'chars')
vocabs['char'] = Vocabulary()
vocabs['char'].from_dataset(
datasets['train'],
field_name='chars',
no_create_entry_dataset=[datasets['dev'], datasets['test']])
for k, v in datasets.items():
v.apply_field(normalize_bigram, 'bigrams', 'bigrams')
vocabs['bigram'] = Vocabulary()
vocabs['bigram'].from_dataset(
datasets['train'],
field_name='bigrams',
no_create_entry_dataset=[datasets['dev'], datasets['test']])
def concat(ins):
chars = ins['chars']
lexicons = ins['lexicons']
result = chars + list(map(lambda x: x[2], lexicons))
# print('lexicons:{}'.format(lexicons))
# print('lex_only:{}'.format(list(filter(lambda x:x[2],lexicons))))
# print('result:{}'.format(result))
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
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')
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
"""
1.word_embedding_path 这个参数到底是用做什么的？
我将其设置成了 None。但是如果为None，那么embedding['word']没有了还可以吗？
2.StaticEmbedding：
给定预训练embedding的名称或路径，根据vocab从embedding中抽取相应的数据(只会将出现在vocab中的词抽取出来， 如果没有找到，则会随机初始化一个值(但如果该word是被标记为no_create_entry的话，则不会单独创建一个值，而是会被指向unk的index))
"""
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
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
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,
number_normalized=False,
lattice_min_freq=1,
only_train_min_freq=0):
from fastNLP.core import Vocabulary
def normalize_char(inp):
result = []
for c in inp:
if c.isdigit():
result.append('0')
else:
result.append(c)
return result
def normalize_bigram(inp):
result = []
for bi in inp:
tmp = bi
if tmp[0].isdigit():
tmp = '0' + tmp[:1]
if tmp[1].isdigit():
tmp = tmp[0] + '0'
result.append(tmp)
return result
if number_normalized == 3:
for k, v in datasets.items():
v.apply_field(normalize_char, 'chars', 'chars')
vocabs['char'] = Vocabulary()
vocabs['char'].from_dataset(
datasets['train'],
field_name='chars',
no_create_entry_dataset=[datasets['dev'], datasets['test']])
for k, v in datasets.items():
v.apply_field(normalize_bigram, 'bigrams', 'bigrams')
vocabs['bigram'] = Vocabulary()
vocabs['bigram'].from_dataset(
datasets['train'],
field_name='bigrams',
no_create_entry_dataset=[datasets['dev'], datasets['test']])
if only_lexicon_in_train:
print('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
from V0.utils_ import Trie
from functools import partial
from fastNLP.core import Vocabulary
# from fastNLP.embeddings import StaticEmbedding
from fastNLP_module import StaticEmbedding
from fastNLP import DataSet
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
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')
if number_normalized == 1:
for k, v in datasets.items():
v.apply_field(normalize_char, 'chars', 'chars')
vocabs['char'] = Vocabulary()
vocabs['char'].from_dataset(
datasets['train'],
field_name='chars',
no_create_entry_dataset=[datasets['dev'], datasets['test']])
if number_normalized == 2:
for k, v in datasets.items():
v.apply_field(normalize_char, 'chars', 'chars')
vocabs['char'] = Vocabulary()
vocabs['char'].from_dataset(
datasets['train'],
field_name='chars',
no_create_entry_dataset=[datasets['dev'], datasets['test']])
for k, v in datasets.items():
v.apply_field(normalize_bigram, 'bigrams', 'bigrams')
vocabs['bigram'] = Vocabulary()
vocabs['bigram'].from_dataset(
datasets['train'],
field_name='bigrams',
no_create_entry_dataset=[datasets['dev'], datasets['test']])
def concat(ins):
chars = ins['chars']
lexicons = ins['lexicons']
result = chars + list(map(lambda x: x[2], lexicons))
# print('lexicons:{}'.format(lexicons))
# print('lex_only:{}'.format(list(filter(lambda x:x[2],lexicons))))
# print('result:{}'.format(result))
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
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')
# print(list(datasets['train'][:10]['lexicons']))
# print(list(datasets['train'][:10]['lattice']))
# print(list(datasets['train'][:10]['lex_s']))
# print(list(datasets['train'][:10]['lex_e']))
# print(list(datasets['train'][:10]['pos_s']))
# print(list(datasets['train'][:10]['pos_e']))
# exit(1208)
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
# for k,v in datasets.items():
# v.apply_field(lambda x:[ list(map(lambda x:x[0],p)) for p in x],'skips_l2r','skips_l2r_source')
# v.apply_field(lambda x:[ list(map(lambda x:x[1],p)) for p in x], 'skips_l2r', 'skips_l2r_word')
#
# for k,v in datasets.items():
# v.apply_field(lambda x:[ list(map(lambda x:x[0],p)) for p in x],'skips_r2l','skips_r2l_source')
# v.apply_field(lambda x:[ list(map(lambda x:x[1],p)) for p in x], 'skips_r2l', 'skips_r2l_word')
# for k,v in datasets.items():
# v.apply_field(lambda x:list(map(len,x)), 'skips_l2r_word', 'lexicon_count')
# v.apply_field(lambda x:
# list(map(lambda y:
# list(map(lambda z:word_vocab.to_index(z),y)),x)),
# 'skips_l2r_word',new_field_name='skips_l2r_word')
#
# v.apply_field(lambda x:list(map(len,x)), 'skips_r2l_word', 'lexicon_count_back')
#
# v.apply_field(lambda x:
# list(map(lambda y:
# list(map(lambda z:word_vocab.to_index(z),y)),x)),
# 'skips_r2l_word',new_field_name='skips_r2l_word')
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
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')
vocabs['span_label'].index_dataset(*(datasets.values()),
field_name='span_label',
new_field_name='span_label')
vocabs['attr_label'].index_dataset(*(datasets.values()),
field_name='attr_start_label',
new_field_name='attr_start_label')
vocabs['attr_label'].index_dataset(*(datasets.values()),
field_name='attr_end_label',
new_field_name='attr_end_label')
return datasets, vocabs, embeddings
class CustomizedNER(object):
def __init__(self, modelFile, vocabFile, addTarget2Vocab=False):
# CHAR_INPUT="chars", 并且会转化为word_index
self._vocabFile = vocabFile
self._addTarget2Vocab = addTarget2Vocab
self._CONST_CHAR = Const.CHAR_INPUT
self._CONST_WORDS = Const.INPUT
self._CONST_TARGET = Const.TARGET
self._input_fields = [self._CONST_WORDS, Const.INPUT_LEN]
self._word_counter, self._word_vocab, self._target_counter, \
self._target_vocab, self._target = self._get_vocabs()
self._vocab4word = Vocabulary()
self._update_word()
if self._addTarget2Vocab:
self._vocab4target = Vocabulary(unknown=None, padding=None)
self._input_fields.append(self._CONST_TARGET)
self._update_target()
self._model = Predictor(ModelLoader().load_pytorch_model(modelFile))
def _target_token(self, word_token, cont, number="", word=""):
ret = dict()
sign = True
lastIdx = len(word_token) - 1
for num, token in zip(enumerate(word_token), cont):
if num[1] in self._target:
if sign:
number += str(num[1])
word += token
if num[0] < lastIdx and not word_token[num[0] + 1]:
sign = False
else:
ret.setdefault(number, set())
ret[number].add(word)
number = ""
word = token
sign = True
if number:
ret.setdefault(number, set())
ret[number].add(word)
return ret
def _extract_ner(self, tokenNum, token, weighted=False):
if not weighted:
cls = self._target.get(int(max(tokenNum, key=tokenNum.count)), "")
if cls.endswith("LOC"):
return {"LOC": [x for x in token]}
elif cls.endswith("PER"):
return {"PER": [x for x in token]}
elif cls.endswith("ORG"):
return {"ORG": [x for x in token]}
def _get_ner(self, tokenNumber, tokenWord):
nerDict = self._target_token(tokenNumber, tokenWord)
ret = dict()
for num, token in nerDict.items():
if len(num) == 1:
continue
for k, v in self._extract_ner(num, token).items():
ret.setdefault(k, list())
ret[k].extend(v)
return ret
def _read_vocab(self):
with open(self._vocabFile, "r", encoding="utf-8") as vocabIn:
return eval(vocabIn.read())
def _reverse_dict(self, dic):
ret = dict()
for key, value in dic.items():
ret.setdefault(value, key)
return ret
def _tartget_label(self, dic):
ret = self._reverse_dict(dic)
del ret[0]
return ret
def _get_vocabs(self):
vocabs = self._read_vocab()
word_count = vocabs.get("wordsWc", dict())
wordsVocab = vocabs.get("wordsVocab", dict())
target_count = vocabs.get("targetWc", dict())
targetVocab = vocabs.get("targetVocab", dict())
reverseTargetVocab = self._tartget_label(targetVocab)
return Counter(word_count), wordsVocab, Counter(
target_count), targetVocab, reverseTargetVocab
def _update_word(self):
self._vocab4word.update(self._word_vocab)
self._vocab4word.word_count = self._word_counter
def _update_target(self):
self._vocab4target.update(self._target_vocab)
self._vocab4target.word_count = self._target_counter
@property
def model(self):
if not self._model:
raise
return self._model
def formatRowString(self, msg):
msg = msg.strip()
tokenized_char = [x for x in msg]
self._dataset = DataSet()
if self._addTarget2Vocab:
ins = Instance(chars=tokenized_char,
raw_chars=tokenized_char,
target=list(dict(self._target_vocab).keys()))
else:
ins = Instance(chars=tokenized_char, raw_chars=tokenized_char)
self._dataset.append(ins)
@property
def dataset(self):
# if input as dict format:
# data = DataSet({"raw_chars":[msg], "words":[[x for x in msg]], "seq_len":[len(word_list)]})
# 从该dataset中的chars列建立词表
self._vocab4word.from_dataset(self._dataset,
field_name=self._CONST_CHAR)
# 使用vocabulary将chars列转换为index
self._vocab4word.index_dataset(self._dataset,
field_name=self._CONST_CHAR,
new_field_name=self._CONST_WORDS)
if self._addTarget2Vocab:
self._vocab4target.from_dataset(self._dataset,
field_name=self._CONST_TARGET)
self._vocab4target.index_dataset(self._dataset,
field_name=self._CONST_TARGET)
self._dataset.add_seq_len(self._CONST_CHAR)
self._dataset.set_input(*self._input_fields)
return self._dataset
def _content(self):
for line in self._dataset["raw_chars"].content:
yield "".join(line)
def result(self, dataset):
# 打印数据集中的预测结果
ret = self.model.predict(dataset)["pred"]
for line, cont in zip(ret, self._content()):
yield self._get_ner(line[0].tolist(), cont)