我需要使用基于规则方法的BIO方案注释文本语料库(我有一个预定义的标记及其标签列表)。我正在为此任务使用spaCy的EntityRuler
类。我的问题是,是否有一种简洁有效的方法来使用spaCy实施BIO标签?另外,我也在努力实现多令牌BIO标记:
'He used sodium-bicarb 5 gr' ->
['O', 'O', 'B-DRUG', 'I-DRUG', 'I-DRUG', 'B-STRENGTH', 'I-STRENGTH')
我有一个简单的(基于规则)脚本来标记我感兴趣的实体:
import numpy as np
import pandas as pd
import spacy
from spacy.pipeline import EntityRuler
nlp = spacy.load('en')
ruler = EntityRuler(nlp).from_disk('drug_patterns.jsonl')
nlp.add_pipe(ruler, before='ner')
text = 'He has been prescribed ipratropium-albuterol a small dose of 20mg, denzapine and amil-co'
doc = nlp(text)
for ent in doc.ents:
print(ent.text, ent.start_char, ent.end_char, ent.label_)
输出:
ipratropium 23 34 DRUG
20mg 61 65 STRENGTH
denzapine 67 76 DRUG
amil-co 81 88 DRUG
因此,我不确定如何将'amil-co'
分成三个标签'B-DRUG, I-DRUG and I-DRUG'
。
理想情况下,我想添加以下注释:
token BIO
0 He O
1 has O
2 been O
3 prescribed O
4 ipratropium B-DRUG
5 - I-DRUG
6 albuterol I-DRUG
7 a O
8 small O
9 dose O
10 of O
11 20 B-STRENGTH
12 mg I-STRENGTH
13 , O
14 denzapine B-DRUG
15 and O
16 amil B-DRUG
17 - I-DRUG
18 co I-DRUG
19 . O
也,在我的词汇表drug_patterns.json
中,相同的长标记可能会出现一次,而不是出现一次:
{"label": "DRUG", "pattern": [{"lower": "ipratropium"}]}
{"label": "DRUG", "pattern": [{"lower": "ipratropium"}, {"lower": "bromide"}]}
{"label": "DRUG", "pattern": [{"lower": "ipratropium"}, {"lower": "-"}, {"lower": "albuterol"}]}
,而不是整个ipratropium-albuterol
只会选择第一个(最短的)令牌ipratropium
(如输出所示)。是否有一种简单的方法告诉spaCy选择最长的令牌?
任何想法都会受到高度赞赏。
答案 0 :(得分:1)
嗯,这是一个令人尴尬的简单解决方案,但希望它对其他人可能很有趣。只需使用令牌的.ent_iob_
和.ent_type_
属性即可。即:
pd.DataFrame([(e.text, e.ent_iob_, e.ent_type_) for e in doc])
0 1 2
0 He O
1 has O
2 been O
3 prescribed O
4 ipratropium B DRUG
5 - O
6 albuterol O
7 a O
8 small O
9 dose O
10 of O
11 20 B STRENGTH
12 mg I STRENGTH
13 , O
14 denzapine B DRUG
15 and O
16 amil B DRUG
17 - I DRUG
18 co I DRUG
19 . O
,然后可以轻松地将后两列以适当的格式与连字符合并。 SpaCy很棒!