计算机工程与应用 ›› 2021, Vol. 57 ›› Issue (6): 101-107.DOI: 10.3778/j.issn.1002-8331.1912-0196

• 模式识别与人工智能 • 上一篇    下一篇

基于位置感知能力胶囊网络的实体关系提取

刘博闻,范春晓   

  1. 北京邮电大学 电子工程学院,北京 100876
  • 出版日期:2021-03-15 发布日期:2021-03-12

Relation Extraction Based on CapsuleNet via Position Perception

LIU Bowen, FAN Chunxiao   

  1. School of Electronic Engineering, Beijing University of Posts and Telecommunications, Beijing 100876, China
  • Online:2021-03-15 Published:2021-03-12

摘要:

目前实体关系提取大都使用卷积神经网络(CNN)和循环神经网络(RNN)。但CNN和RNN均以标量形式进行特征表达,对位置信息的敏感度不够理想。同时,CNN和RNN的最大池化(max-pooling)导致特征信息丢失。针对这两个问题,引入胶囊网络(CapsuleNet),构建了具备位置感知能力的Position Perception CapsuleNet(PPCNet)。胶囊(Capsule)是一组神经元,特征表达基于向量形式。PPCNet将词间的位置关系转化为位置向量(position embedding)融入Capsule以获得位置感知能力。此外,PPCNet使用动态路由(Dynamic routing)替代池化,以减少特征损失,在SemEval-2010task8数据集上得到82.84%的F1值。

关键词: 自然语言处理, 实体关系提取, 胶囊网络(CapsuleNet), 位置感知能力, PPCNet

Abstract:

Most of the current relation extraction methods apply Convolutional Neural Network(CNN) and Recurrent Neural Network(RNN) methods. However, CNN and RNN both express their features in scalar form, which leads to their poor sensitivity to location information. Besides, the maximum pooling of CNN and RNN results in the loss of feature information. Aiming at these two problems, Position Perception CapsuleNet(PPCNet) has been structured. Capsule is a group of neurons whose feature expression is based on vector form. PPCNet transforms the position relation ship between words into position embedding and integrates it into capsule to obtain position perception ability. Furthermore, PPCNet uses dynamic routing instead of pooling to reduce feature loss and gets 82.84% F1 value on SemEval-2010task8 dataset.

Key words: Natural Language Processing(NLP), relation extraction, CapsuleNet, position perception, Position Perception CapsuleNet(PPCNet)