中文文本层次多标签分类算法研究

A Deep Learning Based Method for Hierarchical Multi-Label Text Classification

标签具有层次结构的多标签文本分类任务是具有挑战的与应用价值的任务。常见的层次多标签分类方法往往忽略了标签结构与文本之间的联系,标签之间的关系也难以建模。本文使用图神经网络对标签的层次结构进行建模形成图标签表示,然后与预训练文本表示通过注意力机制进行融合,构建多标签分类模型。通过收集真实数据构建一个中文新闻数据集,在构建的模型在该数据集上进行测试,实验结果说明该模型的有效性。

The task of multi-label text classification with a hierarchical label structure is a challenging yet practical task.Common hierarchical multilabel classification methods often ignore the correlation between label structure and text,and the relationship between labels is also diffi⁃cult to model.In this article,a graph neural network is used to model the label hierarchy to form a graph label representation,which is then fused with the text representation through the attention mechanism to build a multi-label classification model.A Chinese news data set was constructed by collecting real data,and the constructed model was tested on this data set.The experimental results showed the effective⁃ness of the model.