作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2022, Vol. 48 ›› Issue (10): 110-115,122. doi: 10.19678/j.issn.1000-3428.0062891

• 人工智能与模式识别 • 上一篇    下一篇

结合残差BiLSTM与句袋注意力的远程监督关系抽取

江旭, 钱雪忠, 宋威   

  1. 江南大学 人工智能与计算机学院, 江苏 无锡 214112
  • 收稿日期:2021-10-08 修回日期:2021-12-13 发布日期:2022-10-09
  • 作者简介:江旭(1995—),男,硕士研究生,主研方向为深度学习、数据挖掘;钱雪忠,副教授、硕士;宋威,教授、博士。
  • 基金资助:
    国家自然科学基金(62076110);江苏省自然科学基金(BK20181341)。

Distantly Supervised Relationship Extraction Combined with Residual BiLSTM and Sentence Bag Attention

JIANG Xu, QIAN Xuezhong, SONG Wei   

  1. School of Artificial Intelligence and Computer Science, Jiangnan University, Wuxi, Jiangsu 214122, China
  • Received:2021-10-08 Revised:2021-12-13 Published:2022-10-09

摘要: 实体关系抽取是从非结构化和程序化的海量文本中识别出实体之间的语义关系,为本体构建、智能检索等任务提供数据支持,然而现有远程监督关系抽取方法普遍存在需要大量人工标注语料库、提取特征含有噪声且忽略了实体与句子之间关联关系等问题。提出一种基于残差双向长短时记忆网络(BiLSTM)与句袋内和句袋间注意力机制的关系抽取模型,在将词向量和位置向量作为模型输入的基础上,通过残差BiLSTM网络提取语句与实体词中的长距离文本信息,利用句袋内和句袋间注意力机制对提取到的特征信息进行处理,使模型在远程监督过程中减少实体之间的特征提取噪声,并提高模型识别准确性。在NYT数据集上的实验结果表明,该模型能够充分利用实体与关系特征,平均精确率达到86.2%,相比于将卷积神经网络和分段卷积神经网络作为句子编码器的同类模型具有更好的远程监督关系抽取性能。

关键词: 自然语言处理, 关系抽取, 残差连接, 双向长短时记忆网络, 句袋注意力机制

Abstract: Entity relationship extraction entails identifying the semantic relationship between entities from unstructured and programmed massive texts and can provide data support for ontology construction, intelligent retrieval, and other tasks.However, the existing distantly supervised relationship extraction methods generally suffer from three problems:the need for numerous manual annotation corpora, noise contained in the extraction features, and the relationship between entities and sentences being ignored.This paper proposes a relationship extraction model based on a residual Bi-directional Long Short-Term Memory(BiLSTM) network and intra-sentence and inter-sentence bag attention mechanisms.Based on using word vectors and position vectors as model input, the model extracts long-distance text information in sentences and entity words through a residual BiLSTM network, and it uses the intra-sentence and inter-sentence bag attention mechanisms to process the extracted feature information.This enables the model to reduce the feature extraction noise between entities in the distantly supervision process and improve its recognition accuracy.The experimental results on the New York Times(NYT) dataset demonstrate that the model can make full use of the entity and relationship features, and the average precision reaches 86.2%.Compared with similar models using Convolutional Neural Network(CNN) and Piecewise Convolutional Neural Network(PCNN) as sentence encoders, the proposed model has better remote supervised relationship extraction performance.

Key words: natural language processing, relation extraction, residual connection, Bi-directional Long Short-Term Memory(BiLSTM) network, sentence bag attention mechanism

中图分类号: