An Entity-Relation Extraction Method Based on the Mixture-of-Experts Model and Dependency Parsing
Yuanxi Li, Haiyan Wang, Dong Zhang · Applied Sciences · 2025
Entity-relation extraction (ERE) aims to identify entity types and the relationships between them from unstructured texts and is one of the key technologies for constructing knowledge graphs. However, ERE tasks face challenges such as insufficient semantic representations and the complexity of relationship types, which lead to the difficulty of triplet extraction. To address these issues, we propose an entity-relation extraction model that incorporates dependency parsing and a mixture-of-experts architecture. Specifically, we use BERT as a character encoder, while integrating dependency syntax information as a separate encoding path. We apply additive attention to fuse the two pathways of encoding, assigning different weights to each vector in the encoding layer output through a learned weighting process. This enables the model to flexibly adjust the attention given to different features, allowing for a more accurate identification and utilization of syntactic dependencies within a sentence. In the relation classification layer, we employ a mixture-of-experts architecture, allowing each expert to focus on learning different relationship labels, thereby enhancing the model’s ability to accurately identify and capture specific entity relationships. The proposed model achieves superior results to the baseline models on two public ERE datasets, providing a novel and effective solution for entity-relation extraction tasks.