计算机工程

• •    

基于异构图卷积网络的小样本短文本分类方法

  

  • 发布日期:2020-12-09

Few-shot Short Text Classification Method Based On Heterogeneous Graph Convolutional Network

  • Published:2020-12-09

摘要: 为了解决小样本短文本分类中语义稀疏与过拟合问题,提出了一种基于异构图卷积网络的小样本短文本分类模型。 首先利用 BTM 主题模型在短文本数据集上提取主题信息,并构建一个集成实体和主题信息的短文本异构信息网络,用于解决 短文本语义稀疏问题;其次,建立基于随机去邻法和双重注意力机制的异构图卷积网络,用于提取短文本异构信息网络中的 语义信息,其中,随机去邻法用于数据增强,缓解过拟合问题,双重注意力机制可以学习不同相邻节点的重要性和不同节点 类型对当前节点的重要性。在三个短文本数据集上的实验结果表明,与 LSTM、Text GCN、HGAT 等基准模型相比,所构建 的模型在每个类别只有十个标记样本时仍能达到最优性能。

Abstract: In order to solve the problem of semantic sparseness and overfitting in few-shot short text classification, this paper proposed a few-shot short text classification method based on heterogeneous graph convolutional network. Firstly, use the BTM to extract topic information from the short text datasets, and construct a short-text heterogeneous information network that can integrate entity and topic information to solve the problem of short text semantic sparseness. Secondly, construct a novel Heterogeneous Graph Convolutional Networks based on dual-level attention mechanism and the method of randomly reducing neighbors to extract semantic information from short text heterogeneous information network. The method of randomly reducing neighbors is used for data enhancement to alleviate the problem of over-fitting. The dual-level attention mechanism can learn the importance of different neighboring nodes and the importance of different node types to the current node. The experimental results on three short text datasets show that compared with the benchmark model such as LSTM, Text GCN, HGAT and so on, the model still achieves state-of-the-art performance when there are only ten labeled samples per class.