融合额外实体信息的层级联合抽取
Hierarchical federated extraction of additional entity information
姚爽 1徐佳美 1连向伟 1马建红1
作者信息
- 1. 河北工业大学人工智能与数据科学学院,天津 300401
- 折叠
摘要
目前,基于长文本的实体关系抽取研究倾向于对实体和关系隐式建模,未能有效利用实体间、实体与关系间的隐含联系,无法获取两个相距较远实体之间的依赖关系.为此,提出一种融合额外实体信息的层级联合抽取模型J-AH,基于注意力建模实体间依赖特征,通过拼接融合实体类型和依赖特征得到实体聚合表示,计算文本聚合表示与关系信息的中间差异因子进行信息层级融合,形成实体间、实体与关系间相互促进、相互增强的正向互动.在英文公开数据集DialogueRE上的实验数据表明,模型的F1值相较于对比模型提升了 1.3个百分点,验证了模型的优势.
Abstract
At present,researches on entity relationship extraction based on long text tend to model entity and relationship imp-licitly,and fail to make effective use of the implicit relationship between entities and entities,and fail to obtain the dependency relationship between two entities that are far away from each other.Therefore,a hierarchical joint extraction model J-AH was proposed to integrate additional entity information.Based on the model of inter-entity dependency characteristics,the entity aggregation representation was obtained by concatenating and merging entity types and dependency characteristics,and the inter-mediate difference factor of text aggregation representation and relational information was calculated for information hierarchical fusion.Positive interactions between entities,entities and relations that promoted and enhanced each other were formed.By means of the test data presented in English public data set DialogueRE,the Fl value of the model is improved by 1.3 percentage points compared with the comparison models,which verifies the advantages of the model.
关键词
联合抽取/编码器-解码器/额外实体信息/依赖信息/注意力机制/层级融合/BERTKey words
joint extraction/encoder-decoder/additional entity information/dependent information/attention mechanism/hiera-rchical convergence/BERT引用本文复制引用
基金项目
科技部创新方法工作专项(2019IM020300)
出版年
2024
NETL
NSTL
万方数据