基于RoBERTa和加权图卷积网络的中文地质实体关系抽取  

Chinese Geological Entity Relation Extraction Based on RoBERTa and Weighted Graph Convolutional Networks

作　　者：张鲁 段友祥[1] 刘娟[1] 陆誉翕 ZHANG Lu;DUAN Youxiang;LIU Juan;LU Yuxi(College of Computer Science and Technology,China University of Petroleum(East China),Qingdao,Shandong 266580,China)

机构地区：[1]中国石油大学(华东)计算机科学与技术学院,山东青岛266580

出　　处：《计算机科学》2024年第8期297-303,共7页Computer Science

基　　金：中央高校基本科研业务费专项资金(20CX05017A);中石油重大科技项目(ZD2019-183-006)。

摘　　要：知识是大数据和人工智能的基石,知识图谱的可解释性和可扩展性等优势使其成为智能系统的重要技术。智能决策在各个领域都有迫切的应用需求,为知识图谱提供基于数据分析和推理的决策支持和应用场景,但领域场景复杂、数据多源、知识维度广,因此知识图谱的构建和应用都面临着很多挑战。针对地质领域知识图谱构建过程中领域知识模式完备性差的问题,以及现有实体关系抽取方法在处理非欧氏数据时存在的不足,提出了一种基于图结构的实体关系抽取模型RoGCN-ATT。该模型使用RoBERTa-wwm-ext-large中文预训练模型作为序列编码器,结合BiLSTM获取更丰富的语义信息,使用加权图卷积网络结合注意力机制获取结构依赖信息,以增强模型对关系三元组的抽取性能。在地质数据集上F1值达78.56%,与其他模型的对比实验表明,RoGCN-ATT有效提升了实体关系抽取性能,为地质知识图谱的构建和应用提供了有力的支持。Knowledge is the cornerstone of big data and artificial intelligence.Knowledge graphs offer interpretability and sca-lability advantages,making them crucial in intelligent systems.Intelligent decision has urgent application demand in various fields,providing decision support and application scenarios for knowledge graphs based on data analysis and reasoning.However,constructing and applying knowledge graphs face challenges due to complex domain scenarios,multi-source data,and extensive knowledge dimensions.To address the problem of incomplete domain knowledge patterns during geological domain knowledge graph construction and the limitations of existing entity relationship extraction methods in dealing with non-Euclidean data,a graph structure-based entity relationship extraction model RoGCN-ATT is proposed.This model utilizes RoBERTa-wwm-ext-large,a Chinese pre-trained model,as the sequence encoder combined with BiLSTM to capture richer semantic information.It also employs weighted graph convolutional networks along with attention mechanisms to capture structural dependency information and enhance the extraction performance of relation triplets.Experimental results show that the F1 value reaches 78.56%on the geological dataset.Compared with other models,RoGCN-ATT effectively improves the entity-relationship extraction performance and provides strong support for the construction and application of geological knowledge maps.

关 键 词：实体关系抽取 图卷积网络 依存句法分析 注意力机制 地质领域 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]