W2NER详解

论文：https://arxiv.org/pdf/2112.10070.pdf

代码：https://github.com/ljynlp/W2NER

W2NER

介绍

W2NER模型，将NER任务转化预测word-word（备注，中文是字-字），它能够统一处理扁平实体、重叠实体和非连续实体三种NER任务。

假定摄入的句子 X 由 N 个tokne或word组成， X = { x 1 , x 2 , . . . , x N } X = \{x_1,x_2,...,x_N\} X={x1​,x2​,...,xN​}，模型对每个word pair( x i , x j x_i,x_j xi​,xj​)中的两个word关系类别R进行预测，其中 R ∈ { N o n e , N N W , T H W − ∗ } R\in\{None,NNW,THW-^*\} R∈{None,NNW,THW−∗}

• None：两个word之间没有关系，不属于同一实体
• NNW：即Next-Neighboring-Word，表示这两个word在同一个实体中相邻的位置
• THW-*：即Tail-Head-Word-*，表示这两个word在同一个实体中，且分别是实体的结尾和开始。用来判断实体的类别和边界，其中*是实体类型

举一个具体的例子（蓝色箭头为NNW、红色箭头为THW-*）：

上面的句子中由两个症状（symptom）实体，“aching in legs” 和 “aching in shoulders”，分别记作 e 1 , e 2 e_1,e_2 e1​,e2​；针对这两个实体，可以得到（b）中的word-word之间的关系，将句子按word维度构建二维矩阵为：

模型架构

W2NER模型主要是用来预测word pair中两个word之间的关系，也就是最右边的这个图。

接下来，让我们来看下数据流转：

1. 输入的sentence经过EncoderLayer（BERT + BiLSTM）得到word_reps

word_reps = {batch_size，cur_batch_max_sentence_length，lstm_hidden_size}

1. 将word_reps经过CLN（Conditional Layer Normalization）层，得到cln

cln = {batch_size，cur_batch_max_sentence_length，cur_batch_max_sentence_length，lstm_hidden_size}

1. 将word pair的distance_embedding和 三角区域的region_embedding 和 word_reps按最后一个维度拼接，得到conv_inputs

conv_inputs = {batch_size, cur_batch_max_sentence_length, cur_batch_max_sentence_length, dist_emb_size + type_emb_size + lstm_hidden_size}

1. 将conv_inputs经过卷积层（核为1*1的常规二维卷积 + 核为3*3的多层空洞卷积），得到conv_outputs

conv_outputs = {batch_size, output_height = cur_batch_max_sentence_length, output_width = cur_batch_max_sentence_length, conv_hidden_size * 3}

1. 将conv_outputs经过CoPredictor（由Biaffine + MLP组成），得到output

output = {batch_size, cur_batch_max_sentence_length, cur_batch_max_sentence_length, label_num}

此时对output对最后一个维度取softmax，可得到word-word pair，再进行关系解码

解码

情况a（扁平实体）

	（B,A）的关系为THW，则表示B是实体的结尾，A是实体的开始；又（A,B）的关系为NNW，表示A和B是在同一个实体中的相邻位置，所以得到扁平实体“AB”
	同理可得扁平实体“DE”

情况b（重叠实体）

	（C,A）的关系为THW，则C是实体的结尾，A是实体的开始；又（A,B）和（B,C）的关系均为NNW，表示A和B是在同一个实体中的相邻位置，B和C是在同一个实体中的相邻位置，所以得到扁平实体“ABC”
	同理得到扁平实体“BC”

情况c（扁平实体 + 非连续实体）

	得到扁平实体“ABC”、“ABD”

情况d（扁平实体 + 非连续实体）

	得到非连续实体“ACD”、“BCE”

源码介绍

数据输入格式

B指batch_size，L指当前句子的长度

• bert_inputs：bert模型的输入token_ids，也就是input_ids包含[CLS]和[SEP] 维度[B，L + 2]
• grid_labels：标注数据实体构建的THW和NHW关系二维矩阵 维度[B,L,L]
• grid_mask2d：网格mask信息，有效信息True，padding为False，维度[B,L,L]
• dist_inputs：网格字符的相对位置信息，维度[B,L,L]
• pieces2word：维度[B,L,L+2]
• entity_text：用来标明实体信息，包括位置，类别。最后用来做评估使用

假设有句子：常建良，男

实体为：常建良（Name类型）

则pieces2word、pieces2word、grid_mask2d、grid_labels如下

id2index为

dis2idx = np.zeros((1000), dtype='int64')
dis2idx[1] = 1
dis2idx[2:] = 2
dis2idx[4:] = 3
dis2idx[8:] = 4
dis2idx[16:] = 5
dis2idx[32:] = 6
dis2idx[64:] = 7
dis2idx[128:] = 8
dis2idx[256:] = 9

模型代码

模型主类Model

class Model(BaseModel):
    def __init__(self, use_bert_last_4_layers=False):
        super().__init__()
        self.use_bert_last_4_layers = use_bert_last_4_layers
        self.bert = build_transformer_model(config_path=config_path, checkpoint_path=checkpoint_path, # segment_vocab_size=0, 
                                            output_all_encoded_layers = True if use_bert_last_4_layers else False)
        lstm_input_size = self.bert.configs['hidden_size']

        self.dis_embs = nn.Embedding(20, dist_emb_size)
        self.reg_embs = nn.Embedding(3, type_emb_size)

        self.encoder = nn.LSTM(lstm_input_size, lstm_hid_size // 2, num_layers=1, batch_first=True,
                            bidirectional=True)

        conv_input_size = lstm_hid_size + dist_emb_size + type_emb_size

        self.convLayer = ConvolutionLayer(conv_input_size, conv_hid_size, dilation, conv_dropout)
        self.dropout = nn.Dropout(emb_dropout)
        self.predictor = CoPredictor(label_num, lstm_hid_size, biaffine_size,
                                     conv_hid_size * len(dilation), ffnn_hid_size, out_dropout)

        self.cln = LayerNorm(lstm_hid_size, conditional_size=lstm_hid_size)

    def forward(self, token_ids, pieces2word, dist_inputs, sent_length, grid_mask2d):
        bert_embs = self.bert([token_ids, torch.zeros_like(token_ids)])
        if self.use_bert_last_4_layers:
            bert_embs = torch.stack(bert_embs[-4:], dim=-1).mean(-1) # 取最后四层的均值

        length = pieces2word.size(1)
        min_value = torch.min(bert_embs).item()

        # 最大池化
        _bert_embs = bert_embs.unsqueeze(1).expand(-1, length, -1, -1)
        _bert_embs = torch.masked_fill(_bert_embs, pieces2word.eq(0).unsqueeze(-1), min_value)
        word_reps, _ = torch.max(_bert_embs, dim=2)

        # LSTM
        word_reps = self.dropout(word_reps)
        packed_embs = pack_padded_sequence(word_reps, sent_length.cpu(), batch_first=True, enforce_sorted=False)
        packed_outs, (hidden, _) = self.encoder(packed_embs)
        word_reps, _ = pad_packed_sequence(packed_outs, batch_first=True, total_length=sent_length.max())

        # 条件LayerNorm
        cln = self.cln(word_reps.unsqueeze(2), word_reps)

        # concat
        dis_emb = self.dis_embs(dist_inputs)
        tril_mask = torch.tril(grid_mask2d.clone().long())
        reg_inputs = tril_mask + grid_mask2d.clone().long()
        reg_emb = self.reg_embs(reg_inputs)
        conv_inputs = torch.cat([dis_emb, reg_emb, cln], dim=-1)
        
        # 卷积层
        conv_inputs = torch.masked_fill(conv_inputs, grid_mask2d.eq(0).unsqueeze(-1), 0.0)
        conv_outputs = self.convLayer(conv_inputs)
        conv_outputs = torch.masked_fill(conv_outputs, grid_mask2d.eq(0).unsqueeze(-1), 0.0)

        # 输出层
        outputs = self.predictor(word_reps, word_reps, conv_outputs)
        return outputs

ConvolutionLayer类

   class ConvolutionLayer(nn.Module):
    '''卷积层
    '''
    def __init__(self, input_size, channels, dilation, dropout=0.1):
        super(ConvolutionLayer, self).__init__()
        self.base = nn.Sequential(
            nn.Dropout2d(dropout),
            nn.Conv2d(input_size, channels, kernel_size=1),
            nn.GELU(),
        )

        self.convs = nn.ModuleList(
            [nn.Conv2d(channels, channels, kernel_size=3, groups=channels, dilation=d, padding=d) for d in dilation])

    def forward(self, x):
        x = x.permute(0, 3, 1, 2).contiguous()
        x = self.base(x)

        outputs = []
        for conv in self.convs:
            x = conv(x)
            x = F.gelu(x)
            outputs.append(x)
        outputs = torch.cat(outputs, dim=1)
        outputs = outputs.permute(0, 2, 3, 1).contiguous()
        return outputs

CoPredictor类

class CoPredictor(nn.Module):
    def __init__(self, cls_num, hid_size, biaffine_size, channels, ffnn_hid_size, dropout=0):
        super().__init__()
        self.mlp1 = MLP(n_in=hid_size, n_out=biaffine_size, dropout=dropout)
        self.mlp2 = MLP(n_in=hid_size, n_out=biaffine_size, dropout=dropout)
        self.biaffine = Biaffine(n_in=biaffine_size, n_out=cls_num, bias_x=True, bias_y=True)
        self.mlp_rel = MLP(channels, ffnn_hid_size, dropout=dropout)
        self.linear = nn.Linear(ffnn_hid_size, cls_num)
        self.dropout = nn.Dropout(dropout)

    def forward(self, x, y, z):
        h = self.dropout(self.mlp1(x))
        t = self.dropout(self.mlp2(y))
        o1 = self.biaffine(h, t)

        z = self.dropout(self.mlp_rel(z))
        o2 = self.linear(z)
        return o1 + o2

MLP类

class MLP(nn.Module):
    '''MLP全连接
    '''
    def __init__(self, n_in, n_out, dropout=0):
        super().__init__()

        self.linear = nn.Linear(n_in, n_out)
        self.activation = nn.GELU()
        self.dropout = nn.Dropout(dropout)

    def forward(self, x):
        x = self.dropout(x)
        x = self.linear(x)
        x = self.activation(x)
        return x

Biaffine类

class Biaffine(nn.Module):
    '''仿射变换
    '''
    def __init__(self, n_in, n_out=1, bias_x=True, bias_y=True):
        super(Biaffine, self).__init__()

        self.n_in = n_in
        self.n_out = n_out
        self.bias_x = bias_x
        self.bias_y = bias_y
        weight = torch.zeros((n_out, n_in + int(bias_x), n_in + int(bias_y)))
        nn.init.xavier_normal_(weight)
        self.weight = nn.Parameter(weight, requires_grad=True)

    def extra_repr(self):
        s = f"n_in={self.n_in}, n_out={self.n_out}"
        if self.bias_x:
            s += f", bias_x={self.bias_x}"
        if self.bias_y:
            s += f", bias_y={self.bias_y}"

        return s

    def forward(self, x, y):
        if self.bias_x:
            x = torch.cat((x, torch.ones_like(x[..., :1])), -1)
        if self.bias_y:
            y = torch.cat((y, torch.ones_like(y[..., :1])), -1)
        # [batch_size, n_out, seq_len, seq_len]
        s = torch.einsum('bxi,oij,byj->boxy', x, self.weight, y)
        # remove dim 1 if n_out == 1
        s = s.permute(0, 2, 3, 1)

        return s

参考资料

https://blog.csdn.net/HUSTHY/article/details/123870372

https://zhuanlan.zhihu.com/p/546602235

参照代码：

https://github.com/Tongjilibo/bert4torch/blob/master/examples/sequence_labeling/task_sequence_labeling_ner_W2NER.py