Основные понятия
Dependency and constituency syntax fusion enhances document-level relation extraction effectiveness.
Аннотация
本文提出了一种新颖的模型,用于处理文档级关系抽取任务。该模型利用两种额外的句法信息,即依存和从属句法融合,以增强文档级关系抽取的效果。通过在依存图中增加文档节点和使用从属树来增强依存图,我们可以改善依存图的表达能力,并更好地捕捉长距离相关性。实验结果表明,我们的模型在三个公共DocRE数据集上优于现有方法。
Статистика
132,275 entities in DocRED dataset.
96 relation classes in DocRED dataset.
1,500 PubMed abstracts in CDR dataset.
30,192 abstracts in GDA dataset.
Цитаты
"Document-level Relation Extraction (DocRE) aims to identify relation labels between entities within a single document."
"Our model can achieve leading performance in DocRE data in the general domain."
"Our model outperforms the existing method on three public DocRE datasets."