命名实体识别整体逻辑框架（BERT+BiLSTM+CRF）

数据集标注：

标注方式

• BIO
• BMESO
  每个句子以空行区分。

数据预处理

1. 读取所有句子和标签存为两个个list，并判断是否对齐

def load_file(file_path):#读取数据集
    contents = open(file_path, encoding='utf-8').readlines()
    text =[]
    label = []
    texts = []
    labels = []
    for line in contents:
        if line != '\n':
            line = line.strip().split(' ')
            text.append(line[0])
            label.append(line[-1])
        else:
            texts.append(text)
            labels.append(label)
            text = []
            label = []
    return texts, labels
 assert len(texts) == len(labels)#判断是否对齐 

2. 截取最大长度，添加额外的标签

 if len(token) > max_length-2:
            token = token[0:(max_length-2)]
            label = label[0:(max_length-2)]
 tokens_f =['[CLS]'] + token + ['[SEP]']
 label_f = [""] + label + ['']

3. 每个句子字符转id、标签转id，mask
   备注： mask的作用：填充后的字符id全部相同，通过mask中的1和0区分哪些是填充的。

input_ids = [int(vocab[i]) if i in vocab else int(vocab['[UNK]']) for i in tokens_f]
label_ids = [label_dic[i] for i in label_f]
input_mask = [1] * len(input_ids)

4. 小于最大句子长度需要填充padding

        while len(input_ids) < max_length:#填充
            input_ids.append(0)
            input_mask.append(0)
            label_ids.append(label_dic[''])

5. 多种方式存储转换为id的数据集

• 每一个句子都存储在一个类中
• ？？？

6. 封装到pytorch中的Dataloader类中

• 读取每一个句子将id转换为Tensor格式，

train_ids = torch.LongTensor([temp.input_id for temp in train_data])
train_masks = torch.LongTensor([temp.input_mask for temp in train_data])
train_tags = torch.LongTensor([temp.label_id for temp in train_data])

• TensorDataset作用：可以用来对tensor进行,打包封装到DataLoader类中
  

train_dataset = TensorDataset(train_ids, train_masks, train_tags)
train_loader = DataLoader(train_dataset, shuffle=True, batch_size=batch_size)#

7. 封装后的数据读取方式：

    for i, train_batch in enumerate(tqdm(train_loader)):
        sentence, masks, tags = train_batch
        sentence, masks, tags = Variable(sentence), Variable(masks), Variable(tags)

模型

1. 加载模型

#加载模型（初始化一个模型）
model = BERT_LSTM_CRF(bert_model_dir, tagset_size, 768, 200, 1,
                      dropout_ratio=0.4, dropout1=0.4, use_cuda = use_cuda)

2. 是否使用gpu：

if use_cuda:
    model = model.cuda()
    model.load_state_dict(torch.load(model_path))

3.训练设置

model.train()#训练阶段可以更新，eval()不可更新
optimizer = getattr(optim, 'Adam')
optimizer = optimizer(model.parameters(), lr=0.00001, weight_decay=0.00005)

4.迭代epoch训练

for epoch in range(epochs):
    print('epoch: {}，train'.format(epoch))
    for i, train_batch in enumerate(tqdm(train_loader)):
        model.zero_grad()
        sentence, masks, tags = train_batch
        sentence, masks, tags = Variable(sentence), Variable(masks), Variable(tags)
        if use_cuda:
            sentence = sentence.cuda()
            masks = masks.cuda()
            tags = tags.cuda()
        loss = model.neg_log_likelihood_loss(sentence, masks, tags)#计算损失
        loss.backward()
        optimizer.step()

5.计算损失

    def neg_log_likelihood_loss(self, sentence, mask, tags):
        lstm_feats = self.get_output_score(sentence)#得到LSTM输出向量
        loss_value = self.crf.neg_log_likelihood_loss(lstm_feats, mask, tags)
        batch_size = lstm_feats.size(0)
        loss_value /= float(batch_size)
        return loss_value

6. 通过输出向量将其转换为对应的标签，采用viterbi解码

 def _viterbi_decode(self, feats, mask=None):
        """
        Args:
            feats: size=(batch_size, seq_len, self.target_size+2)
            mask: size=(batch_size, seq_len)

        Returns:
            decode_idx: (batch_size, seq_len), viterbi decode结果
            path_score: size=(batch_size, 1), 每个句子的得分
        """
        batch_size = feats.size(0)
        seq_len = feats.size(1)
        tag_size = feats.size(-1)

        length_mask = torch.sum(mask, dim=1).view(batch_size, 1).long()
        mask = mask.transpose(1, 0).contiguous()
        ins_num = seq_len * batch_size
        feats = feats.transpose(1, 0).contiguous().view(
            ins_num, 1, tag_size).expand(ins_num, tag_size, tag_size)

        scores = feats + self.transitions.view(
            1, tag_size, tag_size).expand(ins_num, tag_size, tag_size)
        scores = scores.view(seq_len, batch_size, tag_size, tag_size)

        seq_iter = enumerate(scores)
        # record the position of the best score
        back_points = list()
        partition_history = list()
        mask = (1 - mask.long()).byte()
        try:
            _, inivalues = seq_iter.__next__()
        except:
            _, inivalues = seq_iter.next()
        partition = inivalues[:, self.START_TAG_IDX, :].clone().view(batch_size, tag_size, 1)
        partition_history.append(partition)

        for idx, cur_values in seq_iter:
            cur_values = cur_values + partition.contiguous().view(
                batch_size, tag_size, 1).expand(batch_size, tag_size, tag_size)
            partition, cur_bp = torch.max(cur_values, 1)
            partition_history.append(partition.unsqueeze(-1))

            cur_bp.masked_fill_(mask[idx].view(batch_size, 1).expand(batch_size, tag_size), 0)
            back_points.append(cur_bp)

        partition_history = torch.cat(partition_history).view(
            seq_len, batch_size, -1).transpose(1, 0).contiguous()

        last_position = length_mask.view(batch_size, 1, 1).expand(batch_size, 1, tag_size) - 1
        last_partition = torch.gather(
            partition_history, 1, last_position).view(batch_size, tag_size, 1)

        last_values = last_partition.expand(batch_size, tag_size, tag_size) + \
            self.transitions.view(1, tag_size, tag_size).expand(batch_size, tag_size, tag_size)
        _, last_bp = torch.max(last_values, 1)
        pad_zero = Variable(torch.zeros(batch_size, tag_size)).long()
        if self.use_cuda:
            pad_zero = pad_zero.cuda()
        back_points.append(pad_zero)
        back_points = torch.cat(back_points).view(seq_len, batch_size, tag_size)

        pointer = last_bp[:, self.END_TAG_IDX]
        insert_last = pointer.contiguous().view(batch_size, 1, 1).expand(batch_size, 1, tag_size)
        back_points = back_points.transpose(1, 0).contiguous()

        back_points.scatter_(1, last_position, insert_last)

        back_points = back_points.transpose(1, 0).contiguous()

        decode_idx = Variable(torch.LongTensor(seq_len, batch_size))
        if self.use_cuda:
            decode_idx = decode_idx.cuda()
        decode_idx[-1] = pointer.data
        for idx in range(len(back_points)-2, -1, -1):
            pointer = torch.gather(back_points[idx], 1, pointer.contiguous().view(batch_size, 1))
            decode_idx[idx] = pointer.view(-1).data
        path_score = None
        decode_idx = decode_idx.transpose(1, 0)
        return path_score, decode_idx

7.计算评价指标

######测试函数
def evaluate(medel, dev_loader):
    medel.eval()#不更新参数
    pred = []
    gold = []
    print('evaluate')
    for i, dev_batch in enumerate(dev_loader):#验证集
        sentence, masks, tags = dev_batch
        sentence, masks, tags = Variable(sentence), Variable(masks), Variable(tags)
        if use_cuda:
            sentence = sentence.cuda()
            masks = masks.cuda()
            tags = tags.cuda()
        predict_tags = medel(sentence, masks)
        pred.extend([t for t in predict_tags.tolist()])
        gold.extend([t for t in tags.tolist()])
    pred_label,gold_label = recover_label(pred, gold, l2i_dic,i2l_dic)
    acc, p, r, f = get_ner_fmeasure(gold_label,pred_label)
    print('p: {}，r: {}, f: {}'.format(p, r, f))
    model.train()
    return acc, p, r, f

模型输入输出计算

def forward(self, sentence, masks):
	 	batch_size = sentence.size(0)
        seq_length = sentence.size(1)
        embeds, _ = self.word_embeds(sentence, attention_mask=attention_mask, output_all_encoded_layers=False)#BERT生成词向量

        out = self.idcnn(embeds, seq_length)
        # out = self.linear(out)
        hidden = self.rand_init_hidden(batch_size)
        # if embeds.is_cuda:
        #     hidden = (i.cuda() for i in hidden)
        lstm_out, hidden = self.lstm(out, hidden)#输出和cell，双向维度*2
        lstm_out = lstm_out.contiguous().view(-1, self.hidden_dim * 2)#因为线性变换输入两个维度，不然用自己的想法想，前两个维度是batch——size和length他们可以合并，关系的是最后一个维度做线性变换即(b,l,400)->(b*l,400)
        d_lstm_out = self.dropout1(lstm_out)
        l_out = self.liner(d_lstm_out)
        lstm_feats = l_out.contiguous().view(batch_size, seq_length, -1)#恢复(b*l,class_num)->(b,l,class_num)
        return lstm_feats

• BERT初始化词向量batch_sizelength768
  

模型结构

初始化网络结构

from pytorch_pretrained_bert import BertModel
 def __init__(self, bert_config, tagset_size, embedding_dim, hidden_dim, rnn_layers, dropout_ratio, dropout1, use_cuda):
        super(BERT_LSTM_CRF, self).__init__()
        self.embedding_dim = embedding_dim
        self.hidden_dim = hidden_dim
        self.word_embeds = BertModel.from_pretrained(bert_config)#加载BERT预训练模型

        self.idcnn = IDCNN(input_size=embedding_dim, filters=64)
        self.linear = nn.Linear(64, 256)

        self.lstm = nn.LSTM(64, hidden_dim,
                            num_layers=rnn_layers, bidirectional=True, dropout=dropout_ratio, batch_first=True)
        self.rnn_layers = rnn_layers
        self.dropout1 = nn.Dropout(p=dropout1)
        self.crf = CRF(target_size=tagset_size, average_batch=True, use_cuda=use_cuda)
        self.liner = nn.Linear(hidden_dim*2, tagset_size+2)
        self.tagset_size = tagset_size
        self.use_cuda =  use_cuda

CRF网络结构函数解析[参考]

1. 初始化函数

 def __init__(self, **kwargs):
        """
        Args:
            target_size: int, target size
            use_cuda: bool, 是否使用gpu, default is True
            average_batch: bool, loss是否作平均, default is True
        """
        super(CRF, self).__init__()
        for k in kwargs:
            self.__setattr__(k, kwargs[k])
        self.START_TAG_IDX, self.END_TAG_IDX = -2, -1
        init_transitions = torch.zeros(self.target_size+2, self.target_size+2)
        init_transitions[:, self.START_TAG_IDX] = -1000.
        init_transitions[self.END_TAG_IDX, :] = -1000.
        if self.use_cuda:
            init_transitions = init_transitions.cuda()
        self.transitions = nn.Parameter(init_transitions)#转移分数

neg_log_likelihood_loss = forward_score - gold_score
2. 计算所有路径的分数（forward_score ）参考参考网站

• 输入和之前的一样： scores, mask, tags ，
  其中scores是上一步的输出，feats + transition matrix
• 定义 new_tags
  为了在 展平之后的 score ：scores.view(seq_len, batch_size, -1) 中找到真实路径对应的索引位置

    def _forward_alg(self, feats, mask=None):### 求CRF中的分母"Z", 用于loss
        """
        Do the forward algorithm to compute the partition function (batched).

        Args:
            feats: size=(batch_size, seq_len, self.target_size+2)
            mask: size=(batch_size, seq_len)

        Returns:
            xxx
        """
        batch_size = feats.size(0)
        seq_len = feats.size(1)
        tag_size = feats.size(-1)

        # 1. mask 转置 后 shape 为: (seq_len, batch), 
        #    feats 原先 shape=(batch_size, seq_len, tag_size) 
        #          先转置:    (seq_len, batch_size, tag_size)
        #          view:  (seq_len*batch_size, 1, tag_size)
        #          然后在 -2 维度复制： (seq_len*batch_size, [tag_size], tag_size)
        mask = mask.transpose(1, 0).contiguous()
        ins_num = batch_size * seq_len
        feats = feats.transpose(1, 0).contiguous().view(
            ins_num, 1, tag_size).expand(ins_num, tag_size, tag_size)#tagsize,tagsize目的是和转移矩阵维度一样

        # 2. scores： LSTM所有时间步的输出 feats 先加上 转移分数
        scores = feats + self.transitions.view(
            1, tag_size, tag_size).expand(ins_num, tag_size, tag_size)
        scores = scores.view(seq_len, batch_size, tag_size, tag_size)
        seq_iter = enumerate(scores) 
        # seq_iter： t=0 开始的LSTM所有时间步迭代输出
        # inivalues: t=1 开始的LSTM所有时间步迭代输出
        try:
            _, inivalues = seq_iter.__next__()
        except:
            _, inivalues = seq_iter.next()

        # 2. 计算 a 在 t=0 时刻的初始值
        partition = inivalues[:, self.START_TAG_IDX, :].clone().view(batch_size, tag_size, 1)
        # 3. 迭代计算 a （即partition ） 在 t=1，2，。。。更新的值
        for idx, cur_values in seq_iter: # fro idx = 1,2,3...， cur_values是LSTM输出＋转移分数的值，循环的是每一个字符对应的标签分数计算，len=句子长度
            cur_values = cur_values + partition.contiguous().view(
                batch_size, tag_size, 1).expand(batch_size, tag_size, tag_size)
            cur_partition = log_sum_exp(cur_values, tag_size)
            mask_idx = mask[idx, :].view(batch_size, 1).expand(batch_size, tag_size)
            masked_cur_partition = cur_partition.masked_select(mask_idx.byte())
            if masked_cur_partition.dim() != 0:
                # 将mask_idx中值为1元素对应的masked_cur_partition中位置的元素复制到本partition中。
                # mask应该有和partition相同数目的元素。
                # 即 mask 部分的 partition值不再更新
                mask_idx = mask_idx.contiguous().view(batch_size, tag_size, 1)
                partition.masked_scatter_(mask_idx.byte(), masked_cur_partition)
        
        cur_values = self.transitions.view(1, tag_size, tag_size).expand(
                batch_size, tag_size, tag_size) + partition.contiguous().view(
                batch_size, tag_size, 1).expand(batch_size, tag_size, tag_size)
        cur_partition = log_sum_exp(cur_values, tag_size)
        final_partition = cur_partition[:, self.END_TAG_IDX]
        return final_partition.sum(), scores

def log_sum_exp(vec, m_size):## 模型中经常用到的一种路径运算的实现
    """
    结果和右式相同：torch.log(torch.sum(torch.exp(vec),1))
    直接计算可能会出现 exp(999)=INF 上溢问题
    所以 考虑 torch.max(vec, 1)这部分， 以避免 上溢问题

    Args:
        vec: size=(batch_ size, vanishing_dim, hidden_dim)
        m_size: hidden_dim
    Returns:
        size=(batch_size, hidden_dim)

    """
    _, idx = torch.max(vec, 1)  # B * 1 * M ,为了防止 log（过大值max），所有值减去每列最大值
    max_score = torch.gather(vec, 1, idx.view(-1, 1, m_size)).view(-1, 1, m_size)  # B * M
    return max_score.view(-1, m_size) + torch.log(torch.sum(
        torch.exp(vec - max_score.expand_as(vec)), 1)).view(-1, m_size)

##############################################################

• 为了计算真实径的分数（gold_score）
  
  lstm输出

    def _score_sentence(self, scores, mask, tags):### 求路径pair: frames->tags 的分值
        """
        Args:
            scores: size=(seq_len, batch_size, tag_size, tag_size)
            注：这个score是 _forward_alg中返回的第二项值
              （即LSTM输出的feats 和transition matrix相加之后的结果）
            mask: size=(batch_size, seq_len)
            tags: size=(batch_size, seq_len)

        Returns:
            score:
        """
        batch_size = scores.size(1)
        seq_len = scores.size(0)
        tag_size = scores.size(-1)

        new_tags = Variable(torch.LongTensor(batch_size, seq_len))
        if self.use_cuda:
            new_tags = new_tags.cuda()
        for idx in range(seq_len):
            if idx == 0:
                new_tags[:, 0] = (tag_size - 2) * tag_size + tags[:, 0]
            else:
                new_tags[:, idx] = tags[:, idx-1] * tag_size + tags[:, idx]

        end_transition = self.transitions[:, self.END_TAG_IDX].contiguous().view(
            1, tag_size).expand(batch_size, tag_size)
        length_mask = torch.sum(mask, dim=1).view(batch_size, 1).long()
        end_ids = torch.gather(tags, 1, length_mask-1)

        end_energy = torch.gather(end_transition, 1, end_ids)

        new_tags = new_tags.transpose(1, 0).contiguous().view(seq_len, batch_size, 1)
        tg_energy = torch.gather(scores.view(seq_len, batch_size, -1), 2, new_tags).view(
            seq_len, batch_size)
        tg_energy = tg_energy.masked_select(mask.transpose(1, 0))

        gold_score = tg_energy.sum() + end_energy.sum()

        return gold_score

3. 解码部分
   _, last_bp = torch.max(last_values, 1)#取预测标签的id

• 采用viterbi解码，输出预测标签

def evaluate(medel, dev_loader):
    medel.eval()
    pred = []
    gold = []
    print('evaluate')
    for i, dev_batch in enumerate(dev_loader):
        sentence, masks, tags = dev_batch
        sentence, masks, tags = Variable(sentence), Variable(masks), Variable(tags)
        if use_cuda:
            sentence = sentence.cuda()
            masks = masks.cuda()
            tags = tags.cuda()
        predict_tags = medel(sentence, masks)#解码的是viterbi算法。
        pred.extend([t for t in predict_tags.tolist()])
        gold.extend([t for t in tags.tolist()])
    pred_label,gold_label = recover_label(pred, gold, l2i_dic,i2l_dic)

• 将预测id转换为标签

 pred_label,gold_label = recover_label(pred, gold, l2i_dic,i2l_dic)#得到标签序列
 def recover_label(pred_var, gold_var, l2i_dic, i2l_dic):
    assert len(pred_var) == len(gold_var)
    pred_variable = []
    gold_variable = []
    for i in range(len(gold_var)):
     
        start_index = gold_var[i].index(l2i_dic[''])#寻找一个list中该值所在的下标位置
        end_index = gold_var[i].index(l2i_dic[''])
        pred_variable.append(pred_var[i][start_index:end_index])#只取eos，去掉填充部分
        gold_variable.append(gold_var[i][start_index:end_index])

    pred_label = []
    gold_label = []
    for j in range(len(gold_variable)):#转换为预测标签和真实标签
        pred_label.append([ i2l_dic[t] for t in pred_variable[j] ])
        gold_label.append([ i2l_dic[t] for t in gold_variable[j] ])

    return pred_label, gold_label

• 计算p/r/f1值

acc, p, r, f = get_ner_fmeasure(gold_label,pred_label)
def get_ner_fmeasure(golden_lists, predict_lists, label_type="BMES"):
    sent_num = len(golden_lists)
    golden_full = []
    predict_full = []
    right_full = []
    right_tag = 0
    all_tag = 0
    for idx in range(0,sent_num):
        # word_list = sentence_lists[idx]
        golden_list = golden_lists[idx]
        predict_list = predict_lists[idx]
        for idy in range(len(golden_list)):
            if golden_list[idy] == predict_list[idy]:
                right_tag += 1#预测正确标签数
        all_tag += len(golden_list)#所有真正标签数
        if label_type == "BMES":
            gold_matrix = get_ner_BMES(golden_list)#句子中的存在实体[（start,end）类型]
            pred_matrix = get_ner_BMES(predict_list)
        else:
            gold_matrix = get_ner_BIO(golden_list)
            pred_matrix = get_ner_BIO(predict_list)
        right_ner = list(set(gold_matrix).intersection(set(pred_matrix)))#真实标签和预测标签位置是否正确
        golden_full += gold_matrix
        predict_full += pred_matrix
        right_full += right_ner
    right_num = len(right_full)
    golden_num = len(golden_full)
    predict_num = len(predict_full)
    if predict_num == 0:
        precision = -1
    else:
        precision =  (right_num+0.0)/predict_num#准确率：正确预测实体数/所有预测实体数
    if golden_num == 0:
        recall = -1
    else:
        recall = (right_num+0.0)/golden_num#召回率：正确预测实体数/真实标签正确实体数量
    if (precision == -1) or (recall == -1) or (precision+recall) <= 0.:
        f_measure = -1
    else:
        f_measure = 2*precision*recall/(precision+recall)
    accuracy = (right_tag+0.0)/all_tag#准确率：正确标签/所有标签
    # print "Accuracy: ", right_tag,"/",all_tag,"=",accuracy
    print ("gold_num = ", golden_num, " pred_num = ", predict_num, " right_num = ", right_num)
    return accuracy, precision, recall, f_measure

def _viterbi_decode(self, feats, mask=None):## 求最优路径分值 和 最优路径
    """
    Args:
        feats: size=(batch_size, seq_len, self.target_size+2)
        mask: size=(batch_size, seq_len)

    Returns:
        decode_idx: (batch_size, seq_len), viterbi decode结果
        path_score: size=(batch_size, 1), 每个句子的得分
    """
    batch_size = feats.size(0)
    seq_len = feats.size(1)
    tag_size = feats.size(-1)

    length_mask = torch.sum(mask, dim=1).view(batch_size, 1).long()
    mask = mask.transpose(1, 0).contiguous()
    ins_num = seq_len * batch_size
    feats = feats.transpose(1, 0).contiguous().view(
        ins_num, 1, tag_size).expand(ins_num, tag_size, tag_size)

    scores = feats + self.transitions.view(
        1, tag_size, tag_size).expand(ins_num, tag_size, tag_size)
    scores = scores.view(seq_len, batch_size, tag_size, tag_size)

    seq_iter = enumerate(scores)
    # record the position of the best score
    back_points = list()
    partition_history = list()
    mask = (1 - mask.long()).byte()
    try:
        _, inivalues = seq_iter.__next__()
    except:
        _, inivalues = seq_iter.next()
    partition = inivalues[:, self.START_TAG_IDX, :].clone().view(batch_size, tag_size, 1)
    partition_history.append(partition)

    for idx, cur_values in seq_iter:
        cur_values = cur_values + partition.contiguous().view(
            batch_size, tag_size, 1).expand(batch_size, tag_size, tag_size)
        partition, cur_bp = torch.max(cur_values, 1)
        partition_history.append(partition.unsqueeze(-1))

        cur_bp.masked_fill_(mask[idx].view(batch_size, 1).expand(batch_size, tag_size), 0)
        back_points.append(cur_bp)

    partition_history = torch.cat(partition_history).view(
        seq_len, batch_size, -1).transpose(1, 0).contiguous()

    last_position = length_mask.view(batch_size, 1, 1).expand(batch_size, 1, tag_size) - 1
    last_partition = torch.gather(
        partition_history, 1, last_position).view(batch_size, tag_size, 1)

    last_values = last_partition.expand(batch_size, tag_size, tag_size) + \
        self.transitions.view(1, tag_size, tag_size).expand(batch_size, tag_size, tag_size)
    _, last_bp = torch.max(last_values, 1)#取预测标签的id
    pad_zero = Variable(torch.zeros(batch_size, tag_size)).long()
    if self.use_cuda:
        pad_zero = pad_zero.cuda()
    back_points.append(pad_zero)
    back_points = torch.cat(back_points).view(seq_len, batch_size, tag_size)

    pointer = last_bp[:, self.END_TAG_IDX]
    insert_last = pointer.contiguous().view(batch_size, 1, 1).expand(batch_size, 1, tag_size)
    back_points = back_points.transpose(1, 0).contiguous()

    back_points.scatter_(1, last_position, insert_last)

    back_points = back_points.transpose(1, 0).contiguous()

    decode_idx = Variable(torch.LongTensor(seq_len, batch_size))
    if self.use_cuda:
        decode_idx = decode_idx.cuda()
    decode_idx[-1] = pointer.data
    for idx in range(len(back_points)-2, -1, -1):
        pointer = torch.gather(back_points[idx], 1, pointer.contiguous().view(batch_size, 1))
        decode_idx[idx] = pointer.view(-1).data
    path_score = None
    decode_idx = decode_idx.transpose(1, 0)
    return path_score, decode_idx

4. 计算损失
   

# 求一对  在当前参数下的负对数似然，作为loss
 def neg_log_likelihood_loss(self, feats, mask, tags):
        """
        Args:
            feats: size=(batch_size, seq_len, tag_size)
            mask: size=(batch_size, seq_len)
            tags: size=(batch_size, seq_len)
        """
        batch_size = feats.size(0)
        mask = mask.byte()
        forward_score, scores = self._forward_alg(feats, mask)# 所有路径的分数和
        gold_score = self._score_sentence(scores, mask.byte().bool(), tags)## 正确路径的分数
        if self.average_batch:
            return (forward_score - gold_score) / batch_size
        return forward_score - gold_score