This project is aim to provide a super easy way to train crf model and extract entities from text.
pip install simcrf
crf usually use iob tagging (https://en.wikipedia.org/wiki/Inside_Outside_Beginning)
input data can be:
- word, pos_tag, iob_tag
- word, iob_tag
iob_tag:
- I: inside of entity
- O: outside of entity
- B: beginning of entity
Example:
打印机 n O
采购 v O
品目 n O
采购 v O
单位 n O
曲周县 nr B
职业 n I
技术 n I
教育 vn I
中心 n I
行政区域 n O
曲周县 nr O
公告 n O
时间 n O
技术 n I
教育 vn I
中心 n I
采购 v O
单位地址 n O
曲周县 nr B
职业 n I
技术 n I
教育 vn I
中心 n I
采购 v O
单位 n O
联系方式 l O
18932708288 m O
中心 n I
采购 v O
人 n O
地址 n O
： x O
曲周县 nr B
职业 n I
技术 n I
教育 vn I
中心 n I
采购 v O
人 n O
联系方式 l O
： x O
from simcrf import SimCRF
ner = SimCRF()
# note: also support only tokens
X_train = [
[
("打印机", "n"), ("采购", "v"), ("品目", "n"), ("采购", "v"), ("单位", "n"), ("曲周县", "nr"), ("职业", "n"), ("技术", "n"), ("教育", "vn"), ("中心", "n"), ("行政区域", "n"), ("曲周县", "nr"), ("公告", "n"), ("时间", "n")
],
[
("打印机", "n"), ("采购", "v"), ("品目", "n"), ("采购", "v"), ("单位", "n"), ("曲周县", "nr"), ("职业", "n"), ("技术", "n"), ("教育", "vn"), ("中心", "n"), ("行政区域", "n"), ("曲周县", "nr"), ("公告", "n"), ("时间", "n")
]
]
y_train = [
["O","O","O","O","O","B","I","I","I","I","O","O","O","O"],
["O","O","O","O","O","B","I","I","I","I","O","O","O","O"]
]
X_features = ner.transform(X_train)
ner.fit(X_features, y_train)
print(ner.extract_taggedtokens(X_train[0]))ner.save("~/crf_test.pkl")
ner = SimCRF.load("~/crf_test.pkl")
To support different tokenizer, you need tokenize your text first and feed to crf model.
import jieba.posseg as pseg
ner = SimCRF.load("xxxx.pkl")
text = """ 哈尔滨工业大学招标与采购管理中心受总务处的委托，就哈尔滨工业大学部分住宅小区供热入网项目（项目编号：GC2017DX035）组织采购，评标工作已经结束，中标结果如下：
一、项目信息
项目编号：GC2017DX035
项目名称：哈尔滨工业大学部分住宅小区供热入网
项目联系人：李占奎 王 吉
联系方式：电话： 0451-86417953 13936645563
二、采购单位信息
采购单位名称：总务处
采购单位地址：哈尔滨市南岗区西大直街92号
采购单位联系方式：孔繁武 0451-86417975
三、项目用途、简要技术要求及合同履行日期：
见结果公示
四、采购代理机构信息
采购代理机构全称：哈尔滨工业大学招标与采购管理中心
采购代理机构地址：哈尔滨市南岗区西大直街92号哈尔滨工业大学行政办公楼203房间
采购代理机构联系方式：李占奎 王 吉 电话： 0451-86417953 13936645563
"""
sent = [tuple(pair) for pair in pseg.cut(text)]
ret = ner.extract_taggedtokens(sent)
print(ret)SimCrf aim to provide a simple and easy way to train and extract entities. It take off the feature trasfroming and trainning apart from you.So to customize crf model, you need train a sklearn-crfsuite model. You would change trainning parameter and generation of features yourself, and pass the model to SimCRF:
from simcrf import SimCRF
import sklearn_crfsuite
crf_model = sklearn_crfsuite.CRF(
algorithm="lbfgs",
c1=0.1,
c2=0.1,
max_iterations=100,
all_possible_transitions=True
)
crf_model.fit(X_train, y_train)
ner = SimCRF(crf_model)
ret = ner.extract(sent)sklearn-crfsuite docs: https://sklearn-crfsuite.readthedocs.io/
crfsuite docs: http://www.chokkan.org/software/crfsuite/manual.html