APP下载

基于半监督学习与CRF的应急预案命名实体识别

2020-05-28刘彤魏静倪维健陈思源

软件导刊 2020年3期

刘彤 魏静 倪维健 陈思源

摘 要:传统基于统计的命名实体识别方法存在需要大量人工标注的缺陷,导致识别准确率较低。为了提升识别效果,提出一种基于条件随机场的半监督学习方法(S-CRF)对命名实体进行识别。该方法将实体识别看作序列标注问题,对少量数据进行人工标注并构建实体集,通过K-means聚类算法选取有代表性的未标注数据文本进行自动标注,采用条件随机场对语料进行训练测试。选取中文应急预案文档进行实验,该方法在各个标签上的识别效果分别达到93.52%、93.04%、95.81%。实验结果表明,该方法优于传统规则方法,能有效提高应急预案命名实体的识别效果。

关键词:应急预案;命名实体识别;条件随机场;半监督学习

DOI:10. 11907/rjdk. 192096

中图分类号:TP301   文献标识码:A                文章编号:1672-7800(2020)003-0035-04

Entity Identification Based on Semi-supervised Learning and CRF

LIU Tong, WEI Jing, NI Wei-jian, CHEN Si-yuan

(College of Computer Science and Engineering,Shandong University of Science and Technology,Qingdao 266590,China)

Abstract:The traditional statistical-based named entity recognition method requires large number of manual labeling defects, resulting in low recognition accuracy. In order to improve the recognition effect, we propose a method of conditional random field semi-supervised learning method (S-CRF) to identify and extract named entities. This method regards the entity recognition as the sequence labeling problem, manually label small amounts of data and constructed entity set. The K-means clustering algorithm is used to select representative unlabeled data texts for automatic labeling, and the conditional random field is used to sequence the corpus. The Chinese emergency plan document was selected for experiment. The accuracy of the B, M, and O labels reached 93.52%, 93.04% and 95.81%, respectively. The experimental results show that the method is superior to the traditional rules method and can effectively improve the identification effect of named entity of the contingency plan.

Key Words: emergency plan; named entity identification; conditional random field; semi-supervised learning

0 引言

現代社会中出现的各种突发事件通常具有规模大、复杂度高等特征,各级政府和单位需要设置合理的机构应对突发事件。因此,各省都制定了多种类型且精准的应急预案体系。自然语言作为一个应急预案文档的主要组成成分,其包含命名实体和非命名实体,其中命名实体是指人名、地名、机构名等专有名词及其它名称实体(时间、日期等)[1]。命名实体识别是指对这些词进行识别、整理归类与分析。命名实体识别是NLP处理领域的问题,在机器翻译和信息检索抽取等方面应用广泛,能否正确识别命名实体决定着信息抽取效率[1]。目前,国内外针对命名实体识别的研究方法主要有传统识别方法和神经网络方法,对诸多领域的命名实体识别研究已较为成熟,并且取得了较好效果,但对中文应急预案文本方面的研究较少。……

登录APP查看全文