意图识别及槽填充联合模型cnn-seq2seq
作者:互联网
此分类模型是来自序列模型Convolutional Sequence to Sequence Learning,整体构架如上图所示。
原论文是用来做语言翻译,这里我将稍微修改用来做问答中的slot filling和intent detection联合建模。
本项目中的图片和原始代码是改自https://github.com/bentrevett/pytorch-seq2seq 在此非常感谢作者实现了这么通俗易懂的代码架构,可以让其它人在上面进行修改。
Encoder:
1.句子token和其对应的position经过embedding后,逐元素加和作为source embedding。
2.source embedding经过: 线性层 -> 卷积块后得到的特征 -> 线性层。
3.以上的输出和source embedding进行残差连接。
4.以上的输出,我这里加了一个平均池化后进入线性层,预测输出intent概率。(这时是用来做intent detection,即意图识别)
5.原模型的encoder的输出包含两部分,一个是卷积输出;一个是卷积输出 + source embedding -> 这两个输出将用于deocder中的卷积块中计算相应attention context。
Encoder-conv(encoder中的卷积块):
1.卷积块的初始输入是 source embedding加一个线性层,padding后输入卷积。
2.卷积后经过glu激活函数
3.激活后的输出和padding后的输入进行残差连接,进入下一个卷积块。
4.最终输出卷积特征。
In [ ]:
Decoder:
1.target标签的token和其对应的position经过embedding后,逐元素加和作为target embedding。
2.target embedding经过线性层的输出和target embedding -> 卷积块后得到的特征 -> 线性层。
3.再一次经过线性层输出预测slot标签概率。
注:可以到deocder的卷积块的输入还包含还来encoder的两个输出conved,combined
Decoder-conv(decoder中的卷积块):
1.卷积块的初始输入包含4个部分,分别是:target embedding; 经过一个线性层的target embedding; encoder的卷积块输出conved; encoder联合了source embedding和卷积块的输出conved的联合输出combined。
2.与encoder的卷积块类似,卷积后经过glu激活函数
3.激活后的输出和target embedding; encoder conved; encoder combined一起计算attention。
4.经过以上计算的输出,和padding后的输入进行残差连接。
5.以上的输出进入下一个卷积块。
程序(完整项目见:https://github.com/jiangnanboy/intent_detection_and_slot_filling/blob/master/model4/train.ipynb):
''' 编码器Encoder的实现 ''' class Encoder(nn.Module): def __init__(self, input_dim, emb_dim, intent_dim, hid_dim, n_layers, kernel_size, dropout, max_length=50): super(Encoder, self).__init__() assert kernel_size % 2 == 1,'kernel size must be odd!' # 卷积核size为奇数,方便序列两边pad处理 self.scale = torch.sqrt(torch.FloatTensor([0.5])).to(device) # 确保整个网络的方差不会发生显著变化 self.tok_embedding = nn.Embedding(input_dim, emb_dim) # token编码 self.pos_embedding = nn.Embedding(max_length, emb_dim) # token的位置编码 self.emb2hid = nn.Linear(emb_dim, hid_dim) # 线性层,从emb_dim转为hid_dim self.hid2emb = nn.Linear(hid_dim, emb_dim) # 线性层,从hid_dim转为emb_dim # 卷积块 self.convs = nn.ModuleList([nn.Conv1d(in_channels=hid_dim, out_channels=2*hid_dim, # 卷积后输出的维度,这里2*hid_dim是为了后面的glu激活函数 kernel_size=kernel_size, padding=(kernel_size - 1)//2) # 序列两边补0个数,保持维度不变 for _ in range(n_layers)]) self.dropout = nn.Dropout(dropout) # intent detection 意图识别 self.intent_output = nn.Linear(emb_dim, intent_dim) def forward(self, src): # src: [batch_size, src_len] batch_size = src.shape[0] src_len = src.shape[1] # 创建token位置信息 pos = torch.arange(src_len).unsqueeze(0).repeat(batch_size, 1).to(device) # [batch_size, src_len] # 对token与其位置进行编码 tok_embedded = self.tok_embedding(src) # [batch_size, src_len, emb_dim] pos_embedded = self.pos_embedding(pos.long()) # [batch_size, src_len, emb_dim] # 对token embedded和pos_embedded逐元素加和 embedded = self.dropout(tok_embedded + pos_embedded) # [batch_size, src_len, emb_dim] # embedded经过一线性层,将emb_dim转为hid_dim,作为卷积块的输入 conv_input = self.emb2hid(embedded) # [batch_size, src_len, hid_dim] # 转变维度,卷积在输入数据的最后一维进行 conv_input = conv_input.permute(0, 2, 1) # [batch_size, hid_dim, src_len] # 以下进行卷积块 for i, conv in enumerate(self.convs): # 进行卷积 conved = conv(self.dropout(conv_input)) # [batch_size, 2*hid_dim, src_len] # 进行激活glu conved = F.glu(conved, dim=1) # [batch_size, hid_dim, src_len] # 进行残差连接 conved = (conved + conv_input) * self.scale # [batch_size, hid_dim, src_len] # 作为下一个卷积块的输入 conv_input = conved # 经过一线性层,将hid_dim转为emb_dim,作为enocder的卷积输出的特征 conved = self.hid2emb(conved.permute(0, 2, 1)) # [batch_size, src_len, emb_dim] # 又是一个残差连接,逐元素加和输出,作为encoder的联合输出特征 combined = (conved + embedded) * self.scale # [batch_size, src_len, emb_dim] # 意图识别,加一个平均池化,池化后的维度是:[batch_size, emb_dim] intent_output = self.intent_output(F.avg_pool1d(combined.permute(0, 2, 1), combined.shape[1]).squeeze()) # [batch_size, intent_dim] return conved, combined, intent_output ''' 解码器Decoder实现 ''' class Decoder(nn.Module): def __init__(self, output_dim, emb_dim, hid_dim, n_layers,kernel_size, dropout, trg_pad_idx, max_length=50): super(Decoder, self).__init__() self.kernel_size = kernel_size self.trg_pad_idx = trg_pad_idx self.scale = torch.sqrt(torch.FloatTensor([0.5])).to(device) self.tok_embedding = nn.Embedding(output_dim, emb_dim) self.pos_embedding = nn.Embedding(max_length, emb_dim) self.emb2hid = nn.Linear(emb_dim, hid_dim) self.hid2emb = nn.Linear(hid_dim, emb_dim) self.attn_hid2emb = nn.Linear(hid_dim, emb_dim) self.attn_emb2hid = nn.Linear(emb_dim, hid_dim) # slot filling,槽填充 self.slot_out = nn.Linear(emb_dim, output_dim) self.convs = nn.ModuleList([nn.Conv1d(in_channels=hid_dim, out_channels=2*hid_dim, kernel_size=kernel_size) for _ in range(n_layers)]) self.dropout = nn.Dropout(dropout) def calculate_attention(self, embedded, conved, encoder_conved, encoder_combined): ''' embedded:[batch_size, trg_Len, emb_dim] conved:[batch_size, hid_dim, trg_len] encoder_conved:[batch_size, src_len, emb_dim] encoder_combined:[batch_size, src_len, emb_dim] ''' # 经过一线性层,将hid_dim转为emb_dim,作为deocder的卷积输出的特征 conved_emb = self.attn_hid2emb(conved.permute(0, 2, 1)) # [batch_size, trg_len, emb_dim] # 一个残差连接,逐元素加和输出,作为decoder的联合输出特征 combined = (conved_emb + embedded) * self.scale # [batch_size, trg_len, emb_dim] # decoder的联合特征combined与encoder的卷积输出进行矩阵相乘 energy = torch.matmul(combined, encoder_conved.permute(0, 2, 1)) # [batch_size, trg_len, src_len] attention = F.softmax(energy, dim=2) # [batch_size, trg_len, src_len] attention_encoding = torch.matmul(attention, encoder_combined) # [batch_size, trg_len, emb_dim] # 经过一线性层,将emb_dim转为hid_dim attended_encoding = self.attn_emb2hid(attention_encoding) # [batch_size, trg_len, hid_dim] # 一个残差连接,逐元素加和输出 attended_combined = (conved + attended_encoding.permute(0, 2, 1)) * self.scale # [batch_size, hid_dim, trg_len] return attention, attended_combined def forward(self, trg, encoder_conved, encoder_combined): ''' trg:[batch_size, trg_len] encoder_conved:[batch_size, src_len, emb_dim] encoder_combined:[batch_size, src_len, emb_dim] ''' batch_size = trg.shape[0] trg_len = trg.shape[1] # 位置编码 pos = torch.arange(trg_len).unsqueeze(0).repeat(batch_size, 1).to(device) # [batch_size, trg_len] # 对token和pos进行embedding tok_embedded = self.tok_embedding(trg) # [batch_size, trg_len, emb_dim] pos_embedded = self.pos_embedding(pos.long()) # [batch_size, trg_len, emb_dim] # 对token embedded和pos_embedded逐元素加和 embedded = self.dropout(tok_embedded + pos_embedded) # [batch_size, trg_len, emb_dim] # 经过一线性层,将emb_dim转为hid_dim,作为卷积的输入 conv_input = self.emb2hid(embedded) # [batch_size, trg_len, hid_dim] # 转变维度,卷积在输入数据的最后一维进行 conv_input = conv_input.permute(0, 2, 1) # [batch_size, hid_dim, trg_len] batch_size = conv_input.shape[0] hid_dim = conv_input.shape[1] # 卷积块 for i, conv in enumerate(self.convs): conv_input = self.dropout(conv_input) # 在序列的一端进行pad padding = torch.zeros(batch_size, hid_dim, self.kernel_size - 1).fill_(self.trg_pad_idx).to(device) padded_conv_input = torch.cat((padding, conv_input), dim=2) # [batch_size, hid_dim, trg_len + kernel_size - 1] # 进行卷积 conved = conv(padded_conv_input) # [batch_size, 2 * hid_dim, trg_len] # 经过glu激活,会将原hidden_dim分成两部分 conved = F.glu(conved, dim=1) # [batch_size, hid_dim, trg_len] # 计算attention attention, conved = self.calculate_attention(embedded, conved, encoder_conved, encoder_combined) # [batch_size, trg_len, src_len], [batch_size, hid_dim, trg_len] # 残差连接 conved = (conved + conv_input) * self.scale # [batch_size, hid_dim, trg_len] # 作为下一层卷积的输入 conv_input = conved conved = self.hid2emb(conved.permute(0, 2, 1)) # [batch_size, trg_len, emb_dim] # 预测输出 output = self.slot_out(self.dropout(conved)) # [batch_size, trg_len, output_dim] return output, attention # 包装Encoder与Decoer class Seq2Seq(nn.Module): def __init__(self, encoder, decoder): super(Seq2Seq, self).__init__() # 编码器 self.encoder = encoder # 解码器用于slot槽识别 self.decoder = decoder def forward(self, src, trg): ''' src:[batch_size, src_len] trg:[batch_size, trg_Len-1] # decoder的输入去除了<eos> encoder_conved是encoder中最后一个卷积层的输出 encoder_combined是encoder_conved + (src_embedding + postional_embedding) ''' encoder_conved, encoder_combined, intent_output = self.encoder(src) # [batch_size, src_len, emb_dim]; [batch_size, src_len, emb_dim] # decoder是对一批数据进行预测输出 slot_output, attention = self.decoder(trg, encoder_conved, encoder_combined) # [batch_size, trg_len-1, output_dim]; [batch_size, trg_len-1, src_len] return intent_output, slot_output, attention
标签:dim,填充,emb,self,seq2seq,len,batch,cnn,size 来源: https://www.cnblogs.com/little-horse/p/14462023.html