RaSE-KGC: A Relation-Aware Segment Encoding Approach for Knowledge Graph Completion

Pre-trained language models (PLMs) have shown strong potential for knowledge graph completion (KGC) by leveraging the rich semantic information contained in entity descriptions. However, their performance is often limited by fixed input length constraints, which lead to the truncation of important textual cues. Existing solutions, such as context compression and multi-segment encoding, can process longer texts but either lose fine-grained semantic details or introduce substantial computational overhead. Moreover, they lack the ability to adaptively focus on the segment that is most relevant to a given query. To address these limitations, we propose RaSE-KGC, a relation-aware segment encoding approach for KGC that employs hierarchical cross-attention to dynamically capture the most informative segments of entity descriptions with respect to the query relation. RaSE-KGC organizes the information of incomplete triples into three hierarchical levels—sentence, description, and entity—and selectively aggregates their representations through relation-guided attention. To enhance discriminative representation learning while ensuring training efficiency, RaSE-KGC incorporates (1) a recursive encoder that progressively processes only the most salient tokens through shared parameters, and (2) a dynamic corpus refinement strategy that discards sentences and neighbors receiving low attention from the query relation. Extensive experiments on benchmark datasets of varying scales show that RaSE-KGC achieves state-of-theart performance while preserving high computational efficiency, offering a new direction for adaptive and scalable PLM-based KGC.