NLP 基於kashgari和BERT實現中文命名實體識別（NER）

準備工做，先準備 python 環境，下載 BERT 語言模型

• Python 3.6 環境

須要安裝kashgari

Backend	pypi version desc
TensorFlow 2.x	pip install ‘kashgari>=2.0.0’ coming soon
TensorFlow 1.14+	pip install ‘kashgari>=1.0.0,<2.0.0’ current version
Keras	pip install ‘kashgari<1.0.0’ legacy version

• BERT, Chinese 中文模型
  我選擇的是工大的BERT-wwm-ext模型

在此感謝上述做者

數據集準備

from kashgari.corpus import ChineseDailyNerCorpus

train_x, train_y = ChineseDailyNerCorpus.load_data('train')
valid_x, valid_y = ChineseDailyNerCorpus.load_data('validate')
test_x, test_y  = ChineseDailyNerCorpus.load_data('test')

print(f"train data count: {len(train_x)}")
print(f"validate data count: {len(valid_x)}")
print(f"test data count: {len(test_x)}")
train data count: 20864
validate data count: 2318
test data count: 4636

採用人民日報標註的數據集，格式爲：

海 O
釣 O
比 O
賽 O
地 O
點 O
在 O
廈 B-LOC
門 I-LOC
與 O
金 B-LOC
門 I-LOC
之 O
間 O
的 O
海 O
域 O
。 O

建立 BERT embedding

import kashgari
from kashgari.embeddings import BERTEmbedding

bert_embed = BERTEmbedding('chinese_wwm_ext_L-12_H-768_A-12',
                           task=kashgari.LABELING,
                           sequence_length=100)

建立模型並訓練

from kashgari.tasks.labeling import BiLSTM_CRF_Model

# 還能夠選擇 `CNN_LSTM_Model`, `BiLSTM_Model`, `BiGRU_Model` 或 `BiGRU_CRF_Model`

model = BiLSTM_CRF_Model(bert_embed)
model.fit(train_x,
          train_y,
          x_validate=valid_x,
          y_validate=valid_y,
          epochs=20,
          batch_size=512)
model.save('ner.h5')

模型評估

model.evaluate(test_x, test_y)

BERT + B-LSTM-CRF 模型效果最好。詳細得分以下：

precision	recall	f1-score	support
LOC	0.9208	0.9324	0.9266
ORG	0.8728	0.8882	0.8804
PER	0.9622	0.9633	0.9627
avg / total	0.9169	0.9271	0.9220

模型使用

# -*- coding: utf-8 -*-
import kashgari
import re

loaded_model = kashgari.utils.load_model('per_ner.h5')


def cut_text(text, lenth):
    textArr = re.findall('.{' + str(lenth) + '}', text)
    textArr.append(text[(len(textArr) * lenth):])
    return textArr


def extract_labels(text, ners):
    ner_reg_list = []
    if ners:
        new_ners = []
        for ner in ners:
            new_ners += ner;
        for word, tag in zip([char for char in text], new_ners):
            if tag != 'O':
                ner_reg_list.append((word, tag))

    # 輸出模型的NER識別結果
    labels = {}
    if ner_reg_list:
        for i, item in enumerate(ner_reg_list):
            if item[1].startswith('B'):
                label = ""
                end = i + 1
                while end <= len(ner_reg_list) - 1 and ner_reg_list[end][1].startswith('I'):
                    end += 1

                ner_type = item[1].split('-')[1]
   
                if ner_type not in labels.keys():
                    labels[ner_type] = []
                
                label += ''.join([item[0] for item in ner_reg_list[i:end]])
                labels[ner_type].append(label)    
                
    return labels


while True:
    text_input = input('sentence: ')

    texts = cut_text(text_input, 100)
    ners = loaded_model.predict([[char for char in text] for text in texts])
    print(ners)
    labels = extract_labels(text_input, ners)
    print(labels)

參考文獻

Chinese-BERT-wwm：https://github.com/ymcui/Chinese-BERT-wwm
Kashgari：https://github.com/BrikerMan/Kashgari