Generation-Then-Discrimination: Hierarchical Entity-Relation Decoder for Relational Triplet Extraction

Xiaonan Jing, Gongqing Wu, Xingrui Zhuo

Abstract: Relational Triplet Extraction (RTE) is the key to organizing a clear knowledge graph from unstructured text. Benefiting from the representational power of pre-trained language models (PLMs), RTE methods have achieved impressive performance. Existing PLM-based approaches can be broadly categorized into two frameworks: generative and discriminative. Generative models leverage PLMs to capture the complete semantics of entities and directly generate relational triplets. While discriminative models predict entity positions under relation-specific constraints, ensuring relational uniqueness. However, both paradigms suffer from inherent limitations: generative models are susceptible to hallucination when generating relations. In contrast, discriminative models are sensitive to the prediction errors of entity positions. In this work, we demonstrate that the respective weaknesses of the two frameworks can be effectively mitigated by leveraging the strengths that complement one another. Consequently, we propose a Hierarchical entity-relation Decoder for Relational Triplet Extraction (HD-RTE), which synergistically integrates the advantages of generative and discriminative modeling within a two-stage hierarchical framework. In the first stage, to alleviate the error in entity position prediction, a generative PLM is employed to extract semantically complete entities by leveraging its implicit knowledge and global context understanding ability. In the second stage, to avoid hallucination in relation generation, a discriminative relation decoder operates on the extracted entity pairs. By modeling fine-grained contextual interactions through the attention mechanism, the decoder can precisely determine the relations between entities. Experimental results on widely adopted benchmark datasets, including NYT and WebNLG, demonstrate that HD-RTE delivers state-of-the-art performance, particularly in scenarios involving complex entity-relation networks.

External IDs:doi:10.1109/taslpro.2026.3671604