Joint Extraction of Nested Entities and Relations Based on Multi-task Learning

Nested named entity recognition and relation extraction are two crucial tasks in information extraction. Traditional systems often treat them as separate, sequential tasks, which can lead to error propagation. To mitigate this issue, we present a joint extraction model for nested named entities and relations based on a two-level structure, which facilitates joint learning of these subtasks through parameter sharing. Initially, we employ a hierarchical network to identify nested entities. Then, to extract relationships between the nested entities identified at different layers, we introduce multiple rounds of hierarchical relation extraction, creating a dual-dynamic hierarchical network structure. Moreover, as there is a current lack of suitable tagging schemes, we propose a novel tagging scheme grounded in a hierarchical structure. Utilizing this approach, we relabel three datasets: Genia, KBP, and NYT. Experimental results indicate that our proposed joint extraction model significantly outperforms traditional methods in both tasks.