基于编码解码器与深度主题特征抽取的多标签文本分类

Multi-label Text Classification Based on Seq2Seq Model and Deep Topic Feature Extraction

本文提出了一种基于编码解码器与深度主题特征的模型,实现了多标签文本分类.针对传统多标签文本分类方法的特征语义缺失的问题,采用一种长短时记忆(long short-term memory,LSTM)网络提取文本的局部特征与主题模型(latent dirichlet allocation,LDA)提取文本的全局特征的深度主题特征提取模型(deep topic feature extraction model,DTFEM),得到具有文本深层语义特征的语义编码向量,并将该编码向量作为解码器网络的输入.解码器网络将多标签文本分类的任务看作序列生成的过程,解决了多标签文本分类的标签相关性的问题,并加入attention机制,计算注意力分布概率,突出关键输入对输出的作用,改进了由于输入过长导致的语义缺失问题,最终实现多标签文本分类.实验结果表明,该模型能够获得比传统的多标签文本分类系统更优的结果.另外,实验证明使用深度主题特征的方法可以提高多标签文本分类的性能.

In this paper,a model based on seq2 seq model and deep topic feature extraction is proposed to realize multilabel text classification. Aiming at the problem of feature semantics loss in traditional multi-label text classification method,a model is proposed to extract the local features of texts by using the Long Short-term Memory( LSTM) network and extract the global features of texts by using topic model( Latent Dirichlet Allocation,LDA) named Deep Topic Feature Extraction Model( DTFEM),and then obtain the semantic coding vector with deep semantic feature,and the vector is used as the input of the decoder network. The decoder network regards the task of multi-label text classification as the process of sequence generation,solves the problem of label correlation of multi-label text classification,and adds the attention mechanism to calculate the probability distribution of attention,highlights the effect of key input on the output,improves the semantic missing problem due to excessive input,and realizes the final multi-label text classification. The experimental results show that the model can obtain better results than the traditional multi-label text classification system. In addition,the experiments have shown that the use of deep topic features can improve the performance of multi-label text classification.