A structure-aware semantic-topic interactive framework for generative joint entity and relation extraction

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Generative paradigms have emerged as a powerful approach for Joint Entity and Relation Extraction (JERE), offering unparalleled flexibility and task unification. However, existing generative models suffer from two fundamental limitations: (1) Inadequate exploitation of topic-aware global semantic cues; (2) spurious cross-triple dependencies arising from linearized autoregressive decoding. When these issues coexist, they induce a cascading failure in multi-triple scenarios, causing the model to lose both semantic focus and structural alignment and resulting in redundant, missing, or inconsistent extractions. To address these issues, we propose the Structure-aware Semantic–Topic Interactive Framework (SSTIF) , which integrates topic-level semantic guidance with structure-aware decoding for robust triple generation. Specifically, SSTIF introduces an online topic modeling module with a bi-level topic control mechanism, which injects dynamic topic information into attention computation and vocabulary prediction to enhance global semantic consistency. Moreover, a role encoding mechanism is developed to explicitly model entity and relation roles, reducing structural confusion and mitigating cross-triple interference during generation. Extensive experiments on CoNLL04, ADE, and SciERC demonstrate the effectiveness of SSTIF, with superior performance on ADE and SciERC and competitive results on CoNLL04.