BiLSTM+CRF（Keras）

数据集为玻森命名实体数据。

目前代码流程跑通了，后续再进行优化。

项目地址：https://github.com/cyandn/practice/tree/master/NER

步骤：

数据预处理：

def data_process():
    zh_punctuation = ['，', '。', '？', '；', '！', '……']

    with open('data/BosonNLP_NER_6C_process.txt', 'w', encoding='utf8') as fw:
        with open('data/BosonNLP_NER_6C.txt', encoding='utf8') as fr:
            for line in fr.readlines():
                line = ''.join(line.split()).replace('\\n', '')  # 去除文本中的空字符

                i = 0
                while i < len(line):
                    word = line[i]

                    if word in zh_punctuation:
                        fw.write(word + '/O')
                        fw.write('\n')
                        i += 1
                        continue

                    if word == '{':
                        i += 2
                        temp = ''
                        while line[i] != '}':
                            temp += line[i]
                            i += 1
                        i += 2

                        type_ne = temp.split(':')
                        etype = type_ne[0]
                        entity = type_ne[1]
                        fw.write(entity[0] + '/B_' + etype + ' ')
                        for item in entity[1:]:
                            fw.write(item + '/I_' + etype + ' ')
                    else:
                        fw.write(word + '/O ')
                        i += 1

加载数据：

def load_data(self):
    maxlen = 0

    with open('data/BosonNLP_NER_6C_process.txt', encoding='utf8') as f:
        for line in f.readlines():
            word_list = line.strip().split()
            one_sample, one_label = zip(
                *[word.rsplit('/', 1) for word in word_list])
            one_sample_len = len(one_sample)
            if one_sample_len > maxlen:
                maxlen = one_sample_len
            one_sample = ' '.join(one_sample)
            one_label = [config.classes[label] for label in one_label]
            self.total_sample.append(one_sample)
            self.total_label.append(one_label)

    tok = Tokenizer()
    tok.fit_on_texts(self.total_sample)
    self.vocabulary = len(tok.word_index) + 1
    self.total_sample = tok.texts_to_sequences(self.total_sample)

    self.total_sample = np.array(pad_sequences(
        self.total_sample, maxlen=maxlen, padding='post', truncating='post'))
    self.total_label = np.array(pad_sequences(
        self.total_label, maxlen=maxlen, padding='post', truncating='post'))[:, :, None]

    print('total_sample shape:', self.total_sample.shape)
    print('total_label shape:', self.total_label.shape)

    X_train, self.X_test, y_train, self.y_test = train_test_split(
        self.total_sample, self.total_label, test_size=config.proportion['test'], random_state=666)
    self.X_train, self.X_val, self.y_train, self.y_val = train_test_split(
        X_train, y_train, test_size=config.proportion['val'], random_state=666)

    print('X_train shape:', self.X_train.shape)
    print('y_train shape:', self.y_train.shape)
    print('X_val shape:', self.X_val.shape)
    print('y_val shape:', self.y_val.shape)
    print('X_test shape:', self.X_test.shape)
    print('y_test shape:', self.y_test.shape)

    del self.total_sample
    del self.total_label

构建模型：

def build_model(self):
    model = Sequential()

    model.add(Embedding(self.vocabulary, 100, mask_zero=True))
    model.add(Bidirectional(LSTM(64, return_sequences=True)))
    model.add(CRF(len(config.classes), sparse_target=True))
    model.summary()

    opt = Adam(lr=config.hyperparameter['learning_rate'])
    model.compile(opt, loss=crf_loss, metrics=[crf_viterbi_accuracy])

    self.model = model

训练：

def train(self):
    save_dir = os.path.join(os.getcwd(), 'saved_models')
    model_name = '{epoch:03d}_{val_crf_viterbi_accuracy:.4f}.h5'
    if not os.path.isdir(save_dir):
        os.makedirs(save_dir)

    tensorboard = TensorBoard()
    checkpoint = ModelCheckpoint(os.path.join(save_dir, model_name),
                                    monitor='val_crf_viterbi_accuracy',
                                    save_best_only=True)
    lr_reduce = ReduceLROnPlateau(
        monitor='val_crf_viterbi_accuracy', factor=0.2, patience=10)

    self.model.fit(self.X_train, self.y_train,
                    batch_size=config.hyperparameter['batch_size'],
                    epochs=config.hyperparameter['epochs'],
                    callbacks=[tensorboard, checkpoint, lr_reduce],
                    validation_data=[self.X_val, self.y_val])

预测：

def evaluate(self):
    best_model_name = sorted(os.listdir('saved_models')).pop()
    self.best_model = load_model(os.path.join('saved_models', best_model_name),
                                    custom_objects={'CRF': CRF,
                                                    'crf_loss': crf_loss,
                                                    'crf_viterbi_accuracy': crf_viterbi_accuracy})
    scores = self.best_model.evaluate(self.X_test, self.y_test)
    print('test loss:', scores[0])
    print('test accuracy:', scores[1])

参考：

https://zhuanlan.zhihu.com/p/44042528

https://blog.csdn.net/buppt/article/details/81180361

https://github.com/stephen-v/zh-NER-keras

http://www.voidcn.com/article/p-pykfinyn-bro.html