Dependency-augmented graph aggregation networks for document-level relation extraction with cross-sentence semantics

Finding semantic relationships between words in several sentences is the goal of document-level relation extraction (DocRE), a crucial problem in natural language processing. Current research is unable to accurately characterize long-range interdependence and cross-sentence interactions, which restricts their capacity to capture document-level semantics. In order to alleviate this issue, we introduce the Dependency-Augmented Graph Aggregation Network(DAGA), which is a novel DocRE model. In particular, we design a Dependency Graph Aggregation Module (DGAM) that integrates Sentence Related Graph and Dependency Structure Graph to explore both local and global relational patterns. To explicitly capture document-level sentence-related dependencies and semantic interactions, we propose Dependency-Augmented Attention Mechanism (DAAM). Results from experiments show that our suggested approach improves the F1 score by 1.39 and the Ign F1 score by 1.52 on publicly available benchmark datasets. In summary, DAGA demonstrates higher performance in dealing with complicated semantic relationships at the document level.