基于特征分组与蒸馏网络的关系抽取方法

Relationship Extraction Method Based on Feature Grouping and Distillation Network

关系抽取的主要目的从非结构化的文本中识别出预定义的实体关系,是信息抽取领域的核心分支之一。目前主流的关系抽取方法直接将神经网络每个卷积核提取到的特征杂糅,没有考虑特征之间的异同。为此,论文提出了一种基于特征分组与蒸馏网络的关系抽取方法。此方法将神经网络的特征通道分成若干子特征组,从不同的角度来得到多样化的语义语法信息,获得更加细粒度的特征;此外,基于蒸馏网络的思想,将全局信息的预测输出作为学习目标,引导各个子特征组信息去学习软标签分布。在远程监督关系数据集NYT上的实验结果表明,论文所提方法优于目前主流的关系抽取方法。

The main purpose of relation extraction is to identify predefined entity relationships from unstructured text.It is one of the core branches in the field of information extraction.At present,the mainstream relationship extraction method directly mixes the features extracted by each convolution kernel of neural network,without considering the similarities and differences between features.Therefore,this paper proposes a relationship extraction method based on feature grouping and distillation network.This method divides the feature channel of neural network into several sub-feature groups,obtains diversified sub feature semantic and grammatical information from different perspectives,obtains more fine-grained features.In addition,based on the ideology of distillation network,the predicted output of global information is taken as the learning goal to guide each sub-feature group information in the network to learn the soft label distribution.The experimental results on the remotely supervised relational dataset NYT show that the proposed method is superior to the current mainstream relational extraction methods.