A Span-Constrained Decoding Method for Relation Extraction

Abstract

Driven by the accelerating digitalization of education, artificial intelligence technologies have been progressively embedded into critical application scenarios, including experimental pedagogy and the development of educational resources. The automated extraction of entities and relations from unstructured experimental texts constitutes a foundational prerequisite for knowledge modeling and semantic comprehension, with its performance directly governing the usability and robustness of downstream systems. To tackle the challenges confronting existing relation extraction methods in this domain—specifically, the difficulty of precisely identifying entity boundaries and the compounding of errors stemming from erroneous candidate propagation—this paper proposes SCOPE-CasRel, a span-constrained decoding approach for relation extraction specifically tailored to secondary school experimental texts. Grounded in the CasRel framework, the proposed method introduces a span-constrained pairing mechanism that jointly models the start and end boundaries of entities, complemented by a length-constrained strategy to refine entity span selection, thereby enhancing boundary consistency. Additionally, a candidate filtering strategy is devised for the decoding phase, wherein low-confidence candidates are pruned via a dynamic threshold coupled with a Top-K mechanism, effectively mitigating the accumulation of false positives. Empirical validation is performed on the publicly available NYT and WebNLG datasets. The experimental results reveal that the proposed method surpasses several mainstream baselines across precision, recall, and F1-score, achieving substantial gains in extraction accuracy and overall performance while preserving strong recall capability.