首先，最近出来了很多优秀的预训练语言模型的工作。谷歌的ELECTRA还有fastNLP。国内的一些开源组织，也进行了中文预训练任务的复现。

所以比较秀的是，我们可以根据更多的预训练给出一个任务在不同任务下表现的结果。然后将结果融合到一起。从而提高我们在比赛中的成绩。其实代码变动也不是很大。也就是下面的build_transformer_model换了一下。

#! -*- coding:utf-8 -*-
# 三元组抽取任务，基于“半指针-半标注”结构
# 文章介绍：https://kexue.fm/archives/7161
# 数据集：http://ai.baidu.com/broad/download?dataset=sked
# 最优f1=0.82198
# 换用RoBERTa Large可以达到f1=0.829+

import json

import numpy as np
from bert4keras.backend import keras, K, batch_gather
from bert4keras.layers import LayerNormalization
from bert4keras.models import build_transformer_model
from bert4keras.optimizers import Adam, ExponentialMovingAverage
from bert4keras.snippets import open
from bert4keras.snippets import sequence_padding, DataGenerator
from bert4keras.tokenizers import Tokenizer
from keras.callbacks import TensorBoard
from keras.layers import Input, Dense, Lambda, Reshape
from keras.models import Model
from tqdm import tqdm

config_path = '../electra_tiny/bert_config_tiny.json'
checkpoint_path = '../electra_tiny/model.ckpt-1000000'
dict_path = '../electra_tiny/vocab.txt'

maxlen = 128
batch_size = 64


def load_data(filename):
    D = []
    with open(filename, encoding='utf-8') as f:
        for l in f:
            l = json.loads(l)
            D.append({
                'text': l['text'],
                'spo_list': [
                    (spo['subject'], spo['predicate'], spo['object'])
                    for spo in l['spo_list']
                ]
            })
    return D


# 加载数据集
train_data = load_data('../data/baidu_relation_extraction/train_data.json')
valid_data = load_data('../data/baidu_relation_extraction/dev_data.json')
predicate2id, id2predicate = {}, {}

with open('../data/baidu_relation_extraction/all_50_schemas') as f:
    for l in f:
        l = json.loads(l)
        if l['predicate'] not in predicate2id:
            id2predicate[len(predicate2id)] = l['predicate']
            predicate2id[l['predicate']] = len(predicate2id)

# 建立分词器
tokenizer = Tokenizer(dict_path, do_lower_case=True)


def search(pattern, sequence):
    """从sequence中寻找子串pattern
    如果找到，返回第一个下标；否则返回-1。
    """
    n = len(pattern)
    for i in range(len(sequence)):
        if sequence[i:i + n] == pattern:
            return i
    return -1


class data_generator(DataGenerator):
    """数据生成器
    """

    def __iter__(self, random=False):
        batch_token_ids, batch_segment_ids = [], []
        batch_subject_labels, batch_subject_ids, batch_object_labels = [], [], []
        for is_end, d in self.sample(random):
            token_ids, segment_ids = tokenizer.encode(d['text'], max_length=maxlen)
            # 整理三元组 {s: [(o, p)]}
            spoes = {}
            for s, p, o in d['spo_list']:
                s = tokenizer.encode(s)[0][1:-1]
                p = predicate2id[p]
                o = tokenizer.encode(o)[0][1:-1]
                s_idx = search(s, token_ids)
                o_idx = search(o, token_ids)
                if s_idx != -1 and o_idx != -1:
                    s = (s_idx, s_idx + len(s) - 1)
                    o = (o_idx, o_idx + len(o) - 1, p)
                    if s not in spoes:
                        spoes[s] = []
                    spoes[s].append(o)
            if spoes:
                # subject标签
                subject_labels = np.zeros((len(token_ids), 2))
                for s in spoes:
                    subject_labels[s[0], 0] = 1
                    subject_labels[s[1], 1] = 1
                # 随机选一个subject
                start, end = np.array(list(spoes.keys())).T
                start = np.random.choice(start)
                end = np.random.choice(end[end >= start])
                subject_ids = (start, end)
                # 对应的object标签
                object_labels = np.zeros((len(token_ids), len(predicate2id), 2))
                for o in spoes.get(subject_ids, []):
                    object_labels[o[0], o[2], 0] = 1
                    object_labels[o[1], o[2], 1] = 1
                # 构建batch
                batch_token_ids.append(token_ids)
                batch_segment_ids.append(segment_ids)
                batch_subject_labels.append(subject_labels)
                batch_subject_ids.append(subject_ids)
                batch_object_labels.append(object_labels)
                if len(batch_token_ids) == self.batch_size or is_end:
                    batch_token_ids = sequence_padding(batch_token_ids)
                    batch_segment_ids = sequence_padding(batch_segment_ids)
                    batch_subject_labels = sequence_padding(batch_subject_labels, padding=np.zeros(2))
                    batch_subject_ids = np.array(batch_subject_ids)
                    batch_object_labels = sequence_padding(batch_object_labels,
                                                           padding=np.zeros((len(predicate2id), 2)))
                    yield [
                              batch_token_ids, batch_segment_ids,
                              batch_subject_labels, batch_subject_ids, batch_object_labels
                          ], None
                    batch_token_ids, batch_segment_ids = [], []
                    batch_subject_labels, batch_subject_ids, batch_object_labels = [], [], []


def extrac_subject(inputs):
    """根据subject_ids从output中取出subject的向量表征
    """
    output, subject_ids = inputs
    subject_ids = K.cast(subject_ids, 'int32')
    start = batch_gather(output, subject_ids[:, :1])
    end = batch_gather(output, subject_ids[:, 1:])
    subject = K.concatenate([start, end], 2)
    return subject[:, 0]


# 补充输入
subject_labels = Input(shape=(None, 2), name='Subject-Labels')
subject_ids = Input(shape=(2,), name='Subject-Ids')
object_labels = Input(shape=(None, len(predicate2id), 2), name='Object-Labels')

# 加载预训练模型
bert = build_transformer_model(
    config_path=config_path,
    checkpoint_path=checkpoint_path,
    model="electra",
    return_keras_model=False,
)

# 预测subject
output = Dense(units=2,
               activation='sigmoid',
               kernel_initializer=bert.initializer)(bert.model.output)
subject_preds = Lambda(lambda x: x ** 2)(output)

subject_model = Model(bert.model.inputs, subject_preds)

# 传入subject，预测object
# 通过Conditional Layer Normalization将subject融入到object的预测中
output = bert.model.layers[-2].get_output_at(-1)
subject = Lambda(extrac_subject)([output, subject_ids])
output = LayerNormalization(conditional=True)([output, subject])
output = Dense(units=len(predicate2id) * 2,
               activation='sigmoid',
               kernel_initializer=bert.initializer)(output)
output = Lambda(lambda x: x ** 4)(output)
object_preds = Reshape((-1, len(predicate2id), 2))(output)

object_model = Model(bert.model.inputs + [subject_ids], object_preds)

# 训练模型
train_model = Model(bert.model.inputs + [subject_labels, subject_ids, object_labels],
                    [subject_preds, object_preds])

mask = bert.model.get_layer('Embedding-Token').output_mask
mask = K.cast(mask, K.floatx())

subject_loss = K.binary_crossentropy(subject_labels, subject_preds)
subject_loss = K.mean(subject_loss, 2)
subject_loss = K.sum(subject_loss * mask) / K.sum(mask)

object_loss = K.binary_crossentropy(object_labels, object_preds)
object_loss = K.sum(K.mean(object_loss, 3), 2)
object_loss = K.sum(object_loss * mask) / K.sum(mask)

train_model.add_loss(subject_loss + object_loss)
train_model.compile(optimizer=Adam(1e-5))


def extract_spoes(text):
    """抽取输入text所包含的三元组
    """
    tokens = tokenizer.tokenize(text, max_length=maxlen)
    token_ids, segment_ids = tokenizer.encode(text, max_length=maxlen)
    # 抽取subject
    subject_preds = subject_model.predict([[token_ids], [segment_ids]])
    start = np.where(subject_preds[0, :, 0] > 0.6)[0]
    end = np.where(subject_preds[0, :, 1] > 0.5)[0]
    subjects = []
    for i in start:
        j = end[end >= i]
        if len(j) > 0:
            j = j[0]
            subjects.append((i, j))
    if subjects:
        spoes = []
        token_ids = np.repeat([token_ids], len(subjects), 0)
        segment_ids = np.repeat([segment_ids], len(subjects), 0)
        subjects = np.array(subjects)
        # 传入subject，抽取object和predicate
        object_preds = object_model.predict([token_ids, segment_ids, subjects])
        for subject, object_pred in zip(subjects, object_preds):
            start = np.where(object_pred[:, :, 0] > 0.6)
            end = np.where(object_pred[:, :, 1] > 0.5)
            for _start, predicate1 in zip(*start):
                for _end, predicate2 in zip(*end):
                    if _start <= _end and predicate1 == predicate2:
                        spoes.append((subject, predicate1, (_start, _end)))
                        break
        return [
            (
                tokenizer.decode(token_ids[0, s[0]:s[1] + 1], tokens[s[0]:s[1] + 1]),
                id2predicate[p],
                tokenizer.decode(token_ids[0, o[0]:o[1] + 1], tokens[o[0]:o[1] + 1])
            ) for s, p, o in spoes
        ]
    else:
        return []


class SPO(tuple):
    """用来存三元组的类
    表现跟tuple基本一致，只是重写了 __hash__ 和 __eq__ 方法，
    使得在判断两个三元组是否等价时容错性更好。
    """

    def __init__(self, spo):
        self.spox = (
            tuple(tokenizer.tokenize(spo[0])),
            spo[1],
            tuple(tokenizer.tokenize(spo[2])),
        )

    def __hash__(self):
        return self.spox.__hash__()

    def __eq__(self, spo):
        return self.spox == spo.spox


def evaluate(data):
    """评估函数，计算f1、precision、recall
    """
    X, Y, Z = 1e-10, 1e-10, 1e-10
    f1, precision, recall = 0.0, 0.0, 0.0
    f = open('dev_pred.json', 'w', encoding='utf-8')
    pbar = tqdm()
    for d in data:
        R = set([SPO(spo) for spo in extract_spoes(d['text'])])
        T = set([SPO(spo) for spo in d['spo_list']])
        X += len(R & T)
        Y += len(R)
        Z += len(T)
        f1, precision, recall = 2 * X / (Y + Z), X / Y, X / Z
        pbar.update()
        pbar.set_description('f1: %.5f, precision: %.5f, recall: %.5f' %
                             (f1, precision, recall))
        s = json.dumps(
            {
                'text': d['text'],
                'spo_list': list(T),
                'spo_list_pred': list(R),
                'new': list(R - T),
                'lack': list(T - R),
            },
            ensure_ascii=False,
            indent=4)
        f.write(s + '\n')
    pbar.close()
    f.close()
    return f1, precision, recall


class Evaluator(keras.callbacks.Callback):
    """评估和保存模型
    """

    def __init__(self):
        self.best_val_f1 = 0.

    def on_epoch_end(self, epoch, logs=None):
        EMAer.apply_ema_weights()
        f1, precision, recall = evaluate(valid_data)
        if f1 >= self.best_val_f1:
            self.best_val_f1 = f1
            train_model.save_weights('best_model.weights')
        EMAer.reset_old_weights()
        print('f1: %.5f, precision: %.5f, recall: %.5f, best f1: %.5f\n' %
              (f1, precision, recall, self.best_val_f1))


if __name__ == '__main__':

    train_generator = data_generator(train_data, batch_size)
    evaluator = Evaluator()
    EMAer = ExponentialMovingAverage(0.999)
    tf_board_callback = TensorBoard(log_dir='../ner_tf_dir', update_freq=10)

    train_model.fit_generator(train_generator.forfit(),
                              steps_per_epoch=len(train_generator),
                              epochs=20,
                              callbacks=[evaluator, EMAer,tf_board_callback])

else:

    train_model.load_weights('best_model.weights')

然后下面是我在我本地机器上面训练的过程部分的截图。感觉还是不错的。

2705/2705 [==============================] - 221s 82ms/step - loss: 0.4439
f1: 0.00000, precision: 1.00000, recall: 0.00000: : 21639it [00:53, 406.05it/s]
f1: 0.00000, precision: 1.00000, recall: 0.00000, best f1: 0.00000

Epoch 2/20
2705/2705 [==============================] - 221s 82ms/step - loss: 0.1126
f1: 0.57437, precision: 0.81871, recall: 0.44235: : 21639it [01:39, 217.19it/s]
f1: 0.57437, precision: 0.81871, recall: 0.44235, best f1: 0.57437

Epoch 3/20
2705/2705 [==============================] - 220s 81ms/step - loss: 0.0785
f1: 0.67632, precision: 0.77211, recall: 0.60168: : 21639it [01:43, 209.86it/s]
f1: 0.67632, precision: 0.77211, recall: 0.60168, best f1: 0.67632

Epoch 4/20
2705/2705 [==============================] - 220s 81ms/step - loss: 0.0671
f1: 0.70757, precision: 0.77068, recall: 0.65402: : 21639it [01:43, 208.13it/s]
f1: 0.70757, precision: 0.77068, recall: 0.65402, best f1: 0.70757

#! -*- coding:utf-8 -*-
# 三元组抽取任务，基于“半指针-半标注”结构
# 文章介绍：https://kexue.fm/archives/7161
# 数据集：http://ai.baidu.com/broad/download?dataset=sked
# 最优f1=0.82198
# 换用RoBERTa Large可以达到f1=0.829+
# 欢迎大家加我的微信15246115202

import json

import numpy as np
from bert4keras.backend import keras, K, batch_gather
from bert4keras.layers import LayerNormalization
from bert4keras.models import build_transformer_model
from bert4keras.optimizers import Adam, ExponentialMovingAverage
from bert4keras.snippets import open
from bert4keras.snippets import sequence_padding, DataGenerator
from bert4keras.tokenizers import Tokenizer
from keras.callbacks import TensorBoard
from keras.layers import Input, Dense, Lambda, Reshape
from keras.models import Model
from tqdm import tqdm

config_path = '../electra_tiny/bert_config_tiny.json'
checkpoint_path = '../electra_tiny/model.ckpt-1000000'
dict_path = '../electra_tiny/vocab.txt'

maxlen = 128
batch_size = 64


def load_data(filename):
    D = []
    with open(filename, encoding='utf-8') as f:
        for l in f:
            l = json.loads(l)
            D.append({
                'text': l['text'],
                'spo_list': [
                    (spo['subject'], spo['predicate'], spo['object'])
                    for spo in l['spo_list']
                ]
            })
    return D


# 加载数据集
train_data = load_data('../data/baidu_relation_extraction/train_data.json')
valid_data = load_data('../data/baidu_relation_extraction/dev_data.json')
predicate2id, id2predicate = {}, {}

with open('../data/baidu_relation_extraction/all_50_schemas') as f:
    for l in f:
        l = json.loads(l)
        if l['predicate'] not in predicate2id:
            id2predicate[len(predicate2id)] = l['predicate']
            predicate2id[l['predicate']] = len(predicate2id)

# 建立分词器
tokenizer = Tokenizer(dict_path, do_lower_case=True)


def search(pattern, sequence):
    """从sequence中寻找子串pattern
    如果找到，返回第一个下标；否则返回-1。
    """
    n = len(pattern)
    for i in range(len(sequence)):
        if sequence[i:i + n] == pattern:
            return i
    return -1


class data_generator(DataGenerator):
    """数据生成器
    """

    def __iter__(self, random=False):
        batch_token_ids, batch_segment_ids = [], []
        batch_subject_labels, batch_subject_ids, batch_object_labels = [], [], []
        for is_end, d in self.sample(random):
            token_ids, segment_ids = tokenizer.encode(d['text'], max_length=maxlen)
            # 整理三元组 {s: [(o, p)]}
            spoes = {}
            for s, p, o in d['spo_list']:
                s = tokenizer.encode(s)[0][1:-1]
                p = predicate2id[p]
                o = tokenizer.encode(o)[0][1:-1]
                s_idx = search(s, token_ids)
                o_idx = search(o, token_ids)
                if s_idx != -1 and o_idx != -1:
                    s = (s_idx, s_idx + len(s) - 1)
                    o = (o_idx, o_idx + len(o) - 1, p)
                    if s not in spoes:
                        spoes[s] = []
                    spoes[s].append(o)
            if spoes:
                # subject标签
                subject_labels = np.zeros((len(token_ids), 2))
                for s in spoes:
                    subject_labels[s[0], 0] = 1
                    subject_labels[s[1], 1] = 1
                # 随机选一个subject
                start, end = np.array(list(spoes.keys())).T
                start = np.random.choice(start)
                end = np.random.choice(end[end >= start])
                subject_ids = (start, end)
                # 对应的object标签
                object_labels = np.zeros((len(token_ids), len(predicate2id), 2))
                for o in spoes.get(subject_ids, []):
                    object_labels[o[0], o[2], 0] = 1
                    object_labels[o[1], o[2], 1] = 1
                # 构建batch
                batch_token_ids.append(token_ids)
                batch_segment_ids.append(segment_ids)
                batch_subject_labels.append(subject_labels)
                batch_subject_ids.append(subject_ids)
                batch_object_labels.append(object_labels)
                if len(batch_token_ids) == self.batch_size or is_end:
                    batch_token_ids = sequence_padding(batch_token_ids)
                    batch_segment_ids = sequence_padding(batch_segment_ids)
                    batch_subject_labels = sequence_padding(batch_subject_labels, padding=np.zeros(2))
                    batch_subject_ids = np.array(batch_subject_ids)
                    batch_object_labels = sequence_padding(batch_object_labels,
                                                           padding=np.zeros((len(predicate2id), 2)))
                    yield [
                              batch_token_ids, batch_segment_ids,
                              batch_subject_labels, batch_subject_ids, batch_object_labels
                          ], None
                    batch_token_ids, batch_segment_ids = [], []
                    batch_subject_labels, batch_subject_ids, batch_object_labels = [], [], []


def extrac_subject(inputs):
    """根据subject_ids从output中取出subject的向量表征
    """
    output, subject_ids = inputs
    subject_ids = K.cast(subject_ids, 'int32')
    start = batch_gather(output, subject_ids[:, :1])
    end = batch_gather(output, subject_ids[:, 1:])
    subject = K.concatenate([start, end], 2)
    return subject[:, 0]


# 补充输入
subject_labels = Input(shape=(None, 2), name='Subject-Labels')
subject_ids = Input(shape=(2,), name='Subject-Ids')
object_labels = Input(shape=(None, len(predicate2id), 2), name='Object-Labels')

# 加载预训练模型
bert = build_transformer_model(
    config_path=config_path,
    checkpoint_path=checkpoint_path,
    model="electra",
    return_keras_model=False,
)

# 预测subject
output = Dense(units=2,
               activation='sigmoid',
               kernel_initializer=bert.initializer)(bert.model.output)
subject_preds = Lambda(lambda x: x ** 2)(output)

subject_model = Model(bert.model.inputs, subject_preds)

# 传入subject，预测object
# 通过Conditional Layer Normalization将subject融入到object的预测中
output = bert.model.layers[-2].get_output_at(-1)
subject = Lambda(extrac_subject)([output, subject_ids])
output = LayerNormalization(conditional=True)([output, subject])
output = Dense(units=len(predicate2id) * 2,
               activation='sigmoid',
               kernel_initializer=bert.initializer)(output)
output = Lambda(lambda x: x ** 4)(output)
object_preds = Reshape((-1, len(predicate2id), 2))(output)

object_model = Model(bert.model.inputs + [subject_ids], object_preds)

# 训练模型
train_model = Model(bert.model.inputs + [subject_labels, subject_ids, object_labels],
                    [subject_preds, object_preds])

mask = bert.model.get_layer('Embedding-Token').output_mask
mask = K.cast(mask, K.floatx())

subject_loss = K.binary_crossentropy(subject_labels, subject_preds)
subject_loss = K.mean(subject_loss, 2)
subject_loss = K.sum(subject_loss * mask) / K.sum(mask)

object_loss = K.binary_crossentropy(object_labels, object_preds)
object_loss = K.sum(K.mean(object_loss, 3), 2)
object_loss = K.sum(object_loss * mask) / K.sum(mask)

train_model.add_loss(subject_loss + object_loss)
train_model.compile(optimizer=Adam(1e-5))


def extract_spoes(text):
    """抽取输入text所包含的三元组
    """
    tokens = tokenizer.tokenize(text, max_length=maxlen)
    token_ids, segment_ids = tokenizer.encode(text, max_length=maxlen)
    # 抽取subject
    subject_preds = subject_model.predict([[token_ids], [segment_ids]])
    start = np.where(subject_preds[0, :, 0] > 0.6)[0]
    end = np.where(subject_preds[0, :, 1] > 0.5)[0]
    subjects = []
    for i in start:
        j = end[end >= i]
        if len(j) > 0:
            j = j[0]
            subjects.append((i, j))
    if subjects:
        spoes = []
        token_ids = np.repeat([token_ids], len(subjects), 0)
        segment_ids = np.repeat([segment_ids], len(subjects), 0)
        subjects = np.array(subjects)
        # 传入subject，抽取object和predicate
        object_preds = object_model.predict([token_ids, segment_ids, subjects])
        for subject, object_pred in zip(subjects, object_preds):
            start = np.where(object_pred[:, :, 0] > 0.6)
            end = np.where(object_pred[:, :, 1] > 0.5)
            for _start, predicate1 in zip(*start):
                for _end, predicate2 in zip(*end):
                    if _start <= _end and predicate1 == predicate2:
                        spoes.append((subject, predicate1, (_start, _end)))
                        break
        return [
            (
                tokenizer.decode(token_ids[0, s[0]:s[1] + 1], tokens[s[0]:s[1] + 1]),
                id2predicate[p],
                tokenizer.decode(token_ids[0, o[0]:o[1] + 1], tokens[o[0]:o[1] + 1])
            ) for s, p, o in spoes
        ]
    else:
        return []


class SPO(tuple):
    """用来存三元组的类
    表现跟tuple基本一致，只是重写了 __hash__ 和 __eq__ 方法，
    使得在判断两个三元组是否等价时容错性更好。
    """

    def __init__(self, spo):
        self.spox = (
            tuple(tokenizer.tokenize(spo[0])),
            spo[1],
            tuple(tokenizer.tokenize(spo[2])),
        )

    def __hash__(self):
        return self.spox.__hash__()

    def __eq__(self, spo):
        return self.spox == spo.spox


def evaluate(data):
    """评估函数，计算f1、precision、recall
    """
    X, Y, Z = 1e-10, 1e-10, 1e-10
    f1, precision, recall = 0.0, 0.0, 0.0
    f = open('dev_pred.json', 'w', encoding='utf-8')
    pbar = tqdm()
    for d in data:
        R = set([SPO(spo) for spo in extract_spoes(d['text'])])
        T = set([SPO(spo) for spo in d['spo_list']])
        X += len(R & T)
        Y += len(R)
        Z += len(T)
        f1, precision, recall = 2 * X / (Y + Z), X / Y, X / Z
        pbar.update()
        pbar.set_description('f1: %.5f, precision: %.5f, recall: %.5f' %
                             (f1, precision, recall))
        s = json.dumps(
            {
                'text': d['text'],
                'spo_list': list(T),
                'spo_list_pred': list(R),
                'new': list(R - T),
                'lack': list(T - R),
            },
            ensure_ascii=False,
            indent=4)
        f.write(s + '\n')
    pbar.close()
    f.close()
    return f1, precision, recall


class Evaluator(keras.callbacks.Callback):
    """评估和保存模型
    """

    def __init__(self):
        self.best_val_f1 = 0.

    def on_epoch_end(self, epoch, logs=None):
        EMAer.apply_ema_weights()
        f1, precision, recall = evaluate(valid_data)
        if f1 >= self.best_val_f1:
            self.best_val_f1 = f1
            train_model.save_weights('best_model.weights')
        EMAer.reset_old_weights()
        print('f1: %.5f, precision: %.5f, recall: %.5f, best f1: %.5f\n' %
              (f1, precision, recall, self.best_val_f1))


if __name__ == '__main__':

    train_generator = data_generator(train_data, batch_size)
    evaluator = Evaluator()
    EMAer = ExponentialMovingAverage(0.999)
    tf_board_callback = TensorBoard(log_dir='../ner_tf_dir', update_freq=10)

    train_model.fit_generator(train_generator.forfit(),
                              steps_per_epoch=len(train_generator),
                              epochs=20,
                              callbacks=[evaluator, EMAer,tf_board_callback])

else:

    train_model.load_weights('best_model.weights')