Abstract: Entity-relation joint extraction, as a core part of knowledge graph construction, aims to extract entity-relation triples from unstructured text. Current joint extraction methods often struggle with decoding inefficiencies, resulting in weak interaction modeling between entities and relations, insufficient context understanding, and redundant information. To address these limitations, we propose a model based on parallel decoding and clustering for entity-relation joint extraction. First, the bidirectional encoder representations from transformers (BERT) model is used for text encoding to obtain character vectors rich in semantic information. Next, a non-autoregressive parallel decoder is employed to enhance interactions between entities and relations. To further optimize decoding results, hierarchical agglomerative clustering is combined with a majority voting mechanism, improving contextual information capture and reducing redundancy. Finally, a high-quality set of triples is generated to construct a curriculum knowledge graph. To evaluate the performance of the proposed method, experiments are conducted on the public datasets NYT and WebNLG, as well as a self-constructed C language dataset. The results show that this method outperforms other models in terms of precision and F1 score.

Key words: joint extraction, parallel decoding, hierarchical agglomerative clustering, majority voting mechanism, curriculum knowledge graph