[总结]领域用户内容挖掘

本质上来说就是一个分类模型，然后喂不同的数据，尝试不同的模型，看最终的效果。

评论数据

1.尝试了取前1024个字符过textcnn
2.对用户的评论提取关键短语，取topk个短语过textcnn/textrnn
3.transformer + fc
4.Document bert

textcnn

最基础的模型了，如下图所示：

部分代码如下所示：

class TextCNN(nn.Module):
    def __init__(self,pretrain_embed,vocab,config):
        super(TextCNN,self).__init__()
        self.config=config
        self.vocab=vocab
        # if pretrain_embed is not None:
        #     self.embedding=nn.Embedding.from_pretrained(pretrain_embed,padding_idx=len(vocab)-1,freeze=True)
        # else:
        self.embedding=nn.Embedding(len(vocab),config.embed_size,padding_idx=len(vocab)-1)
        
        self.convs=nn.ModuleList([nn.Conv1d(config.embed_size,config.num_filters,kernel_size) for kernel_size in config.filter_sizes])
        self.dropout=nn.Dropout(config.dropout)

        self.fc = nn.Linear(config.num_filters*len(config.filter_sizes),len(config.labels))

    def conv_and_pool(self,conv,x):
        x=F.relu(conv(x)) #[batch_size,embed_size,seq_len] -> [batch_size,num_filters,?]
        x=F.max_pool1d(x,x.size(2)).squeeze(2) #[batch_size,num_filters]
        return x

    # x:[batch_size]
    def forward(self,x):
        x,mask=x
        x=self.embedding(x) # [batch_size,seq_len,embed_size]
        x=x.permute(0,2,1)
        x=torch.cat([self.conv_and_pool(conv,x) for conv in self.convs],1)
        x=self.dropout(x)
        x=self.fc(F.relu(x)) # 
        return x

LSTM

class LSTMModel(nn.Module):
    def __init__(self,pretrain_embed,vocab,config):
        super(LSTMModel,self).__init__()
        self.config=config
        self.vocab=vocab
        if pretrain_embed is not None:
            self.embedding=nn.Embedding.from_pretrained(pretrain_embed,padding_idx=len(vocab)-1,freeze=False)
        else:
            self.embedding=nn.Embedding(len(vocab),config.embed_size,padding_idx=len(vocab)-1)
        self.dropout=nn.Dropout(config.dropout)
        self.lstm = nn.LSTM(config.embed_size,config.hidden_size,config.num_layers,batch_first=True,dropout=config.dropout)
        self.fc = nn.Linear(config.hidden_size,len(config.labels))

    # x:[batch_size,seq_len]
    def forward(self,x):
        x,_=x
        x=self.embedding(x) #[batch_size,seq_len,embed_size]
        x=self.dropout(x)
        x,_=self.lstm(x) # [batch_size,seq_len,hidden_size]
        x=self.fc(x[:,-1,:]) # last output
        return x

transformer + fc

transformer的attention部分如下：