航空事故领域的知识抽取方法研究与实现

doi:10.12068/j.issn.1005-3026.2026.20240234

摘要/Abstract

摘要：

随着航空运输业与信息技术的快速发展，航空应急管理给海量、异构的航空安全数据高效利用带来了挑战.本文针对航空事故知识图谱的知识抽取问题，即命名实体识别与关系抽取问题，提出以下方法：1）提出基于BERT（bidirectional encoder representations from Transformers）的改进BiGRU-IDCNN-CRF模型，实现94.69%的命名实体识别精确率；2）构建基于强化学习的聚类远程监督关系抽取模型，结合改进K均值聚类与远程监督标注降低数据噪声，并通过强化学习优化去噪过程，最终结合分段卷积神经网络（PCNN）与注意力机制，实现84.16%的关系抽取精确率.实验结果表明，本文方法有效提升了航空事故知识图谱的信息提取质量，为航空安全管理提供了精准的信息支撑.

关键词: 航空事故, 知识抽取, 命名实体识别, 关系抽取, 远程监督, 强化学习

Abstract:

In light of the rapid development of air transportation and information technology， the efficient utilization of massive and heterogeneous aviation safety data in aviation emergency management faces challenges. The problem of knowledge extraction for an aviation accident knowledge graph was studied， specifically named entity recognition and relation extraction， and the following methods were proposed： 1） An improved BiGRU-IDCNN-CRF model based on bidirectional encoder representations from Transformers （BERT） was presented， achieving a named entity recognition accuracy of 94.69%； 2） A reinforcement learning-based clustering distant supervision relation extraction model was constructed， in which data noise was reduced by integrating improved K-means clustering with distant supervision labeling， and the denoising process was optimized via reinforcement learning； a combination of piecewise convolutional neural network （PCNN） and an attention mechanism was applied to achieve a relation extraction accuracy of 84.16%. Experimental results indicate that the quality of information extraction for the aviation accident knowledge graph is effectively improved， providing accurate information support for aviation safety management.

Key words: aviation accident, knowledge extraction, named entity recognition, relation extraction, distant supervision, reinforcement learning

中图分类号:

TP 391

刘军, 曹悦, 刘向军, 王宏艳. 航空事故领域的知识抽取方法研究与实现[J]. 东北大学学报（自然科学版）, 2026, 47(1): 89-98.

Jun LIU, Yue CAO, Xiang-jun LIU, Hong-yan WANG. Research and Implementation of Knowledge Extraction in Aviation Accident Domain[J]. Journal of Northeastern University(Natural Science), 2026, 47(1): 89-98.

图/表 9

图1 基于BERT嵌入的改进BiGRU-IDCNN-CRF航空事故识别模型

Fig.1 Improved BiGRU-IDCNN-CRF aviation accident recognition model based on BERT embedding

图2 基于强化学习的聚类远程监督关系抽取

Fig.2 Reinforcement learning-based clustering distant supervision relation extraction

表1 聚类远程监督方法实现步骤

Table 1 Implementation steps of clustering distant supervision method

输入：句子样本集 $M = {m 1, m 2, m 3, …}$ ，K是聚类簇数，N是最大迭代次数
输出:聚类划分后的句子集 $C = {C 1, C 2, C 3, …, C K}$
开始:
1.从数据集D中随机选取K个句子样本作为初始聚类中心: ${μ 1, μ 2, μ 3, …, μ K}$
2.初始簇划分C为 $C t = ∅, t = 1,2, …, K$ 3.对于样本集 M 中的每一个样本 $m n$ 执行 4.计算句子向量 $m n$ 到各聚类中心 $μ i$ 的距离: $d i s (m n, μ i)$
5.将样本分配给具有最小距离 $d i s (m n, μ i)$ 的中心点 $μ λ$ ，并更新簇划分 $C λ = C λ ⋃ {m n}$
6.为每个簇重新计算质心 $μ$
7.若所有K个聚类中心的位置均未发生变化，或目标函数已收敛，或迭代次数已达到设定上限N，则跳转至步骤8
8.输出簇划分 $C = {C 1, C 2, C 3, …, C K}$

表1 聚类远程监督方法实现步骤

Table 1 Implementation steps of clustering distant supervision method

输入：句子样本集 $M = {m 1, m 2, m 3, …}$ ，K是聚类簇数，N是最大迭代次数
输出:聚类划分后的句子集 $C = {C 1, C 2, C 3, …, C K}$
开始:
1.从数据集D中随机选取K个句子样本作为初始聚类中心: ${μ 1, μ 2, μ 3, …, μ K}$
2.初始簇划分C为 $C t = ∅, t = 1,2, …, K$ 3.对于样本集 M 中的每一个样本 $m n$ 执行 4.计算句子向量 $m n$ 到各聚类中心 $μ i$ 的距离: $d i s (m n, μ i)$
5.将样本分配给具有最小距离 $d i s (m n, μ i)$ 的中心点 $μ λ$ ，并更新簇划分 $C λ = C λ ⋃ {m n}$
6.为每个簇重新计算质心 $μ$
7.若所有K个聚类中心的位置均未发生变化，或目标函数已收敛，或迭代次数已达到设定上限N，则跳转至步骤8
8.输出簇划分 $C = {C 1, C 2, C 3, …, C K}$

图3 APCNNA模型图

Fig.3 APCNNA model diagram

表2 不同命名实体识别模型的对比结果 (recognition models %)

Table 2 Comparison results of different named entity

模型	P	R	F1
CRF	83.68	84.76	84.22
GRU	85.48	86.62	86.05
IDCNN	83.23	79.56	81.35
BiGRU	86.56	87.23	86.89
BiGRU-CRF	88.80	87.16	87.97
IDCNN-CRF	87.39	81.64	84.41
BERT-BiGRU-CRF	92.20	91.00	91.59
BERT-IDCNN-CRF	91.80	88.10	89.91
BERT-BiGRU-IDCNN-CRF	93.15	90.39	91.74
BERT-改进BiGRU-IDCNN-CRF	94.69	91.72	93.18

图4 稳定性实验结果对比

Fig.4 Comparison of stability experimental results

表3 有无注意力机制模型的不同表现对比

Table 3 Comparison of different performance of models with and without attention mechanism %

模型	Micro_P	Micro_R	Micro_F1
PCNN	75.27	70.18	72.48
ATT_Z+PCNN	76.27	81.03	78.50
PCNN+ATT_C	80.12	78.86	79.48
APCNNA	81.59	83.26	82.42

表4 聚类远程监督方法效果

Table 4 Effect of clustering distant supervision method

方法	传统远程监督				聚类远程监督
方法	Micro_P/%	Micro_R/%	Micro_F1/%	t/h	Micro_P/%	Micro_R/%	Micro_F1/%	t/h
Mintz	39.80	35.20	37.36	2	40.15	38.23	39.17	2.2
MultiR	54.20	40.15	46.13	1	50.16	48.74	49.44	1.2
MIMLRE	42.60	36.90	39.54	4	45.20	40.10	42.50	4.2

表5 不同模型的对比结果 (%)

Table 5 Comparison results of different models

模型	Micro_P	Micro_R	Micro_F1
MultiR	50.16	48.74	49.44
MIMLRE	45.20	40.10	42.50
BGWA	69.46	71.24	70.34
PCNN+ONE	72.56	74.92	73.72
PCNN	75.27	70.18	72.48
APCNNA	81.59	83.26	82.42
APCNNA+RL	84.16	83.41	83.96

参考文献 23

[1]	Pujara J， Miao H， Getoor L， et al. Knowledge graph identification［C］//The Semantic Web-ISWC 2013. Berlin， Heidelberg： Springer， 2013： 542-557.
[2]	张汝佳，代璐，王邦，等. 基于深度学习的中文命名实体识别最新研究进展综述［J］. 中文信息学报， 2022， 36（6）： 20-35.
	Zhang Ru-jia， Dai Lu， Wang Bang， et al. Recent advances of Chinese named entity recognition based on deep learning［J］. Journal of Chinese Information Processing， 2022， 36（6）： 20-35.
[3]	Goller C， Kuchler A. Learning task-dependent distributed representations by backpropagation through structure［C］//Proceedings of International Conference on Neural Networks （ICNN’96）. Washington DC， 1996： 347-352.
[4]	Gers F A， Schmidhuber J， Cummins F. Learning to forget： continual prediction with LSTM［J］. Neural Computation， 2000， 12（10）： 2451-2471.
[5]	Qin Q L， Zhao S， Liu C M. A BERT-BiGRU-CRF model for entity recognition of Chinese electronic medical records［J］. Complexity， 2021， 2021： 6631837.
[6]	Li Z R， Hu C M， Guo X H， et al. An unsupervised multiple-task and multiple-teacher model for cross-lingual named entity recognition［C］//Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics. Dublin， 2022： 170-179.
[7]	Boudjellal N， Zhang H P， Khan A， et al. ABioNER： a BERT-based model for Arabic biomedical named-entity recognition［J］. Complexity， 2021， 2021： 6633213.
[8]	Zhou R， Li X， Bing L D， et al. Improving self-training for cross-lingual named entity recognition with contrastive and prototype learning［C］//Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics. Toronto， 2023： 4018-4031.
[9]	Ma J， Ballesteros M， Doss S， et al. Label semantics for few shot named entity recognition［C］//Findings of the Association for Computational Linguistics： ACL 2022. Dublin，2022： 1956-1971.
[10]	鄂海红，张文静，肖思琪，等．深度学习实体关系抽取研究综述［J］. 软件学报， 2019， 30（6）： 1793-1818.
	Hai-hong E， Zhang Wen-jing， Xiao Si-qi， et al. Survey of entity relationship extraction based on deep learning ［J］. Journal of Software，2019，30（6）：1793-1818.
[11]	He H， Ganjam K， Jain N， et al. An insight extraction system on BioMedical literature with deep neural networks［C］//Proceedings of the 2017 Conference on Empirical Methods in NaturalLanguage Processing. Copenhagen， 2017： 2691-2701.
[12]	Zhang M S， Zhang Y， Fu G H. End-to-end neural relation extraction with global optimization［C］//Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing. Copenhagen， 2017： 1730-1740.
[13]	Yang S Z， Liu Y X， Zhang K W， et al. Overview of remote supervision relationship extraction［J］.Journal of Computer Science， 2021， 44（8）：1636-1660.
[14]	许鑫冉，王腾宇，鲁才. 图神经网络在知识图谱构建与应用中的研究进展［J］. 计算机科学与探索， 2023， 17（10）： 2278-2299.
	Xu Xin-ran， Wang Teng-yu， Lu Cai. Research progress of graph neural network in knowledge graph construction and application［J］. Journal of Frontiers of Computer Science and Technology， 2023， 17（10）： 2278-2299.
[15]	郑志蕴，徐亚媚，李伦，等.融合位置特征注意力与关系增强机制的远程监督关系抽取［J］. 小型微型计算机系统， 2023， 44（12）：2678-2684.
	Zheng Zhi-yun， Xu Ya-mei， Li Lun， et al. Distantly supervised relation extraction with position feature attention and relation enhancement［J］.Journal of Chinese Computer Systems， 2023， 44（12）：2678-2684.
[16]	Wang G Y， Zhang W， Wang R X， et al. Label-free distant supervision for relation extraction via knowledge graph embedding［C］//Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. Brussels， 2018： 2246-2255.
[17]	Ji G L， Liu K， He S Z， et al. Distant supervision for relation extraction with sentence-level attention and entity descriptions［C］// Proceedings of the Thirty-First AAAI Conference on Artificial Intelligence. San Francisco， 2017： 3060-3066.
[18]	Chen T， Shi H Z， Tang S L， et al. CIL： contrastive instance learning framework for distantly supervised relation extraction［C］//Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics. Online， 2021： 6191-6200.
[19]	Luo B F， Feng Y S， Wang Z， et al. Learning with noise： enhance distantly supervised relation extraction with dynamic transition matrix［C］//Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics. Vancouver，2017： 430-439.
[20]	Zhou Y R， Pan L M， Bai C Y， et al. Self-selective attention using correlation between instances for distant supervision relation extraction［J］. Neural Networks， 2021， 142： 213-220.
[21]	Mintz M， Bills S， Snow R， et al. Distant supervision for relation extraction without labeled data［C］//Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP. Singapore， 2009： 1003-1011.
[22]	Hoffmann R， Zhang C L， Ling X， et al. Knowledge-based weak supervision for information extraction of overlapping relations［C］//Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics： Human Language Technologies. Portland， 2011： 541-550.
[23]	Surdeanu M， Tibshirani J， Nallapati R， et al. Multi-instance multi-label learning for relation extraction［C］//Conference on Empirical Methods in Natural Language Processing. Jeju Island， 2012：455-465.