摘要
[研究目的]在知识海量增长的时代,知识图谱问答面临信息需求与知识图谱加速更新的现实情境,亟需探索关系链接模型在关系类型频繁更新时仍能保持链接效果的方法,实现未见关系类别与用户提问的精准语义匹配。[研究方法]针对模型泛化性不足与灾难性遗忘问题,引入Adapter-Bert迁移学习框架;针对模型对辨别性语义部分的捕获能力不足问题,引入实体特征与问题变换,并将稠密向量与问句抽象意义形式化表示两种不同的语义表示方式进行融合。[研究结论]结果表明,方法在未见关系链接任务上的准确率达到98.80%,相较bert基线模型有显著提高,提升了未见关系链接的效果。
[Research purpose]In the era of explosive knowledge growth,knowledge graph question answering is facing the reality of information demand and knowledge graph accelerated updates.Therefore,It is urgent to explore relation linking methods to maintain linking effectiveness and achieve accurate semantic matching between unseen relations and questions when relations are frequently updated.[Research method]To address the problems of inadequate model generalization and catastrophic forgetting,we adapt Adapter-Bert transfer learning framework.To address the problem of inadequate capture of discriminative semantic parts,we add entity feature and question transformation to model,and combine two different semantic representation methods which are dense vectors and abstract meaning formalized representation.[Research conclusion]The results show that the accuracy of our unseen relation linking method reached 98.80%,which is significantly higher than the Bert baseline model,and improves the effectiveness of unseen relation linking.
作者
徐红霞
Xu Hongxia(School of Information Resource Management,Renmin University of China,Beijing 100080)
出处
《情报杂志》
CSSCI
北大核心
2023年第12期153-158,167,共7页
Journal of Intelligence
基金
北京市自然科学基金项目“计算档案学视角下城市历史遗迹孪生数据管护研究”(编号:9222015)的研究成果。
关键词
自然语言处理
未见关系
迁移学习
细粒度文本特征
抽象意义表示
natural language processing
unseen relation
transfer learning
fine-grained features
abstract meaning representation
