融合液态神经网络与多层级图卷积的关系抽取方法

打开文本图片集

关键词：关系抽取；液态神经网络；图卷积网络；预训练模型；注意力门控；多层感知机中图分类号：TP391.1 文献标志码：A 文章编号：1001-3695（2026）01-008-0069-07doi：10.19734/j. issn. 1001-3695.2025.06.0183

Relation extraction method integrating liquid neural networks and hierarchical graph convolution

Li Ziliang，Li Xingchun† （SchoolofElectronics&InformationEnginering，Wuyi University，JiangmenGuangdong529O2O，China）

Abstract：This paper proposed arelationextractionmodel named BLGAM to adesslimitations inmodeling long-distance dependencies andunderstandingcomplex semantics.Themodel fistlyapplied BERTto encode contextualsemanticsandobtain initialtextrepresentations.Itusedaliquid neuralnetwork basedonaclosed-formcontinuous-timesolutiontocapturedyamic temporal featuresand modelong-distance dependencies.The model thenused dependencysyntax and entitystructures toconstructamulti-levelgraphconvolutionalnetworkforextractinglocalandglobalstructuralfeatures.Anatention-gatingmechanismfused temporal and structuralfeatures，andamulti-layer perceptron enhancedrelationrecognitionaccuracyandrobustness.Experiments on NYT and WebNLG datasets achieve F1 scores of 92.6% and 92.1% ，respectively， surpassing mainstreambaselines.Resultsdemonstratethesuperiorityofliquid neural networksinlong-distancedependencymodelinganddynamicinformationcapturing，andthecomplementaryroleofmulti-levelgraphconvolutionalnetworksinuncoveringimplicitentityrelationships.The method provides an eficient solution for relation extraction in complex semantic scenarios.

Key words：relation extraction；liquid neuralnetwork；graphconvolution network；pre-trained model;atentiongating；MLP

0 引言

关系抽取是自然语言处理的核心任务之一，旨在从非结构化文本中识别给定实体对间的语义关系，生成结构化的“〈主体，关系，客体>”三元组，为知识图谱构建、智能问答、推荐系统及网络空间测绘等应用提供关键支持[1]。（剩余18159字）