Sentence-level Relation Extraction Method with Document Context Heterogeneous Representation
The goal of relationship extraction is to identify relationships between two entities in a text.Although recent studies have achieved promising results in relationship extraction by grouping data for processing,there is limited interaction between the grouped data,and correlations between them are thus overlooked.Moreover,some existing methods involve many labels,leading to redundant labeling information.To address these issues,this study proposes a sentence-level relation extraction method with document context heterogeneous representation.By employing a document context module based on a heterogeneous graph network,the words and relations within data groups are modeled as nodes in a graph,and intragroup information interaction is achieved through message passing,effectively representing correlations between data within the groups.This method employs a relation information module based on a heterogeneous graph network to capture relation information with shared parameters between the heterogeneous graph network in the document context module,thereby saving computational resources.Additionally,the method employs a fusion labeling strategy that uses a logical virtual label to reduce the number of label categories and minimize redundant labeling information.The experimental results demonstrate that the constructed model achieves F1 values of 93.2％for the NYT dataset and 94.7％for the WebNLG dataset.This method outperforms comparative models in six of the eight subtasks of complex scenarios,validating its effectiveness.
fusion labelingheterogeneous graph networkone-module one-step modelsentence-level relation extractionnatural language processing
国家科技期刊平台
NSTL
万方数据
