https://reader.elsevier.com/reader/sd/pii/S0950705120300447

Heterogeneous graph neural networks for noisy few-shot relation classification

未完待续

1. 简介

1.1 摘要

Relation classification is an essential and fundamental task in natural language processing. Distant supervised methods have achieved great success on relation classification, which improve the per-formance of the task through automatically extending the dataset. However, the distant supervisedmethods also bring the problem of wrong labeling. Inspired by people learning new knowledge fromonly a few samples, we focus on predicting formerly unseen classes with a few labeled data. In thispaper, we propose a heterogeneous graph neural network for few-shot relation classification, whichcontains sentence nodes and entity nodes. We build the heterogeneous graph based on the messagepassing between entity nodes and sentence nodes in the graph, which can capture rich neighborhoodinformation of the graph. Besides, we introduce adversarial learning for training a robust model andevaluate our heterogeneous graph neural networks under the scene of introducing different rates ofnoise data. Experimental results have demonstrated that our model outperforms the state-of-the-artbaseline models on the FewRel dataset.

关系分类是NLP中一项非常重要且基础的任务。远程监督方法通过自动扩展数据集来提高模型表现，在关系分类任务中取得了很好地表现。但是远程监督方法也存在一些问题——错误判别标签。受人类只需从少量样本中学习到知识启发，我们希望模型能够只通过少量有标签样本就能够识别之前未见过的类别样本。在这篇文章中我们提出了一种用于小样本关系分类的异构图神经网络，其包含句子节点和实体节点。我们基于图中实体节点和句子节点之间的消息传播构建了这个异构图，可以捕捉到图中丰富的邻居信息。另外，我们引入了对抗学习来提高模型稳定性，并且在引入不同比例噪声数据的场景下对我们的模型进行了评估。实验结果表明我们的模型在一些真实数据集上的表现要优于当前最好的方法。