基于动态权重的多模型集成水产动物疾病防治事件抽取方法  

Multi-model integrated event extraction for aquatic animal disease prevention and control based on dynamic weight

作　　者：沙明洋 张思佳 傅庆财 于红 李枳錡 喻文甫 刘珈宁 SHA Mingyang;ZHANG Sijia;FU Qingcai;YU Hong;LI Zhiqi;YU Wenfu;LIU Jianing(College of Information Engineering/Liaoning Provincial Key Laboratory of Marine Information Technology,Dalian Ocean University,Dalian 116023,China;Key Laboratory of Environment Controlled Aquaculture(Dalian Ocean University),Ministry of Education,Dalian 116023,China)

机构地区：[1]大连海洋大学信息工程学院/辽宁省海洋信息技术重点实验室,大连116023 [2]设施渔业教育部重点实验室(大连海洋大学),大连116023

出　　处：《华中农业大学学报》2023年第3期80-87,共8页Journal of Huazhong Agricultural University

基　　金：设施渔业教育部重点实验室开放课题(2021MOEKLECA-KF-05);计算机体系结构国家重点实验室开放课题(CARCH201921);辽宁省教育厅高等学校基本科研项目面上项目(20220056);辽宁省教育科学“十四五”规划课题(JG21DB076)。

摘　　要：为提高水产动物疾病防治事件抽取的准确性,有效解决抽取过程中出现的专有名词边界模糊和事件实体过长等问题,本研究将动态权重思想引入多模型集成的事件抽取方法中。改进后的方法利用百度自然语言理解开放平台(enhanced representation through knowledge integration,ERNIE)和澎湃BERT(MLM as correction BERT,MacBERT)2个预训练模型来学习文本语义信息;采用动态权重的gate模块融合特征;将学习到的语义信息传入双向长短时记忆网络(bi-directional long shortterm memory,BiLSTM)中,并通过条件随机场(conditional random field,CRF)对输出标签序列进行约束。选取ERNIE⊕MacBERT-CRF模型和ERNIE⊕MacBERT-BiLSTM-CRF模型(⊕代表简单相加求平均的融合方法)作为对照模型对提出的方法进行融合性能对比试验验证,结果显示,该方法 F1值达74.15%,比经典模型BiLSTM-CRF提高了20.02个百分点。结果表明,该方法用于水产动物疾病防治事件抽取具有更好的效果。In order to enhance the accuracy of event extraction for aquatic animal disease prevention and control,and effectively address issues such as ambiguous boundaries of proprietary terms and excessively lengthy event entities during the extraction process,the research introduces the idea of dynamic weight into the event extraction method of multi-model integration.Two pre-training models,ERNIE(enhanced representation through knowledge integration)and MacBERT(MLM as correction BERT),are used to learn the text semantic information.A gate module with dynamic weights is used to fuse features to enhance the semantic information of the original text.Pass the learned semantic information into BiLSTM(bi-directional long shortterm memory),and constrain the output label sequence through CRF(conditional random field).Select the ERNIE ⊕ MacBERT-CRF model and the ERNIE ⊕ MacBERT-BiLSTM-CRF model (⊕ represents the fusion method of simple addition and averaging) as the control model to conduct a comparative test of the fusion performance of the proposed method.The results show that the F1-score of this method reaches 74.15%,which is 20.02 percentage points higher than the classic model BiLSTM-CRF.The results show that this method has a better effect in the extraction of aquatic animal disease prevention and control events.

关 键 词：水产动物疾病 事件抽取 ERNIE MacBERT 动态权重 健康养殖 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]