松耦合图卷积文本分类网络

Loosely Coupled Graph Convolutional Neural Network for Text Classification

文本分类任务是自然语言处理领域内一个重要的研究问题.近年来,因处理复杂网络结构的出色能力,图神经网络模型(Graph Neural Network,GNN)受到广泛关注并被引入到文本分类任务中.在之前的研究中,基于图卷积网络(Graph Convolutional Neural Netw ork,GCN)的分类模型使用紧耦合方式将语料库中的文本和单词组织到同一张网络中,然而这种紧耦合处理方法存在消耗内存过大、对新样本不友好等问题.为解决上述问题,本文设计了一个松耦合图卷积文本分类网络模型(Loosely Coupled Graph Convolutional Neural Netw ork,LCGCN).模型将分类过程分解为核心提取和一般计算两部分,从而完成对紧耦合模型的解耦合操作.该模型能够在保持分类性能的基础上,有效地降低模型内存需求并动态地处理新来测试样本.另外,模型还将标签信息引入到图卷积网络中,进一步提升分类能力.实验表明,相比于其他文本分类网络模型,我们的模型在多个公开文本分类数据集上取得了最优的表现.

Text classification is a fundamental and important task in natural language processing.Recently,Graph Neural Network(GNN) has attracted lots of attention and been applied to natural language processing for its outstanding ability in handling complex graph structure.However,previous GNN-based models organize documents and words in a tightly coupled way,which consumes high memory and is unfriendly to new-coming instances.In order to solve the problems,we propose a Loosely Coupled Graph Convolutional Neural Network(LCGCN) for text classification task in this paper.By decoupling the tightly coupled graph convolutional neural network text classification model,the proposed model can efficiently reduce memory consumption and deal with the new-coming instances dynamically while maintaining the classification performance.Furthermore,the category information is integrated into graphs to improve the classification performance.Practical experiments show that our model achieves the best performance on multiple open text classification data sets.