A joint-embedding framework fusing multi-feature information for cultural relic entity alignment in knowledge graphs

Abstract

Entity alignment across multi-source knowledge bases is frequently hindered by the problem of ’heterogeneous names but equivalent semantics,’ a challenge that becomes particularly acute in domains with specialized nomenclature and rare characters, such as cultural relics. To address this universal issue, this paper proposes a joint-embedding framework, MulF-ELMo-BERT, suitable for entity alignment in cultural relic knowledge graphs. This framework integrates multi-dimensional features such as entity names, attributes, summaries, and full-text texts, achieving comprehensive extraction of entity features from four levels: characters, words, sentences, and paragraphs. It effectively filters out weakly relevant entities and breaks through the limitations of semantic representation relying solely on a single feature. Given the dense presence of rare characters and specialized terms in cultural relic entity names, the ELMo context-aware word embedding model is introduced. It can dynamically adjust the vector representation of rare words based on their context, significantly enhancing semantic adaptability. Meanwhile, a Chinese BERT model with an integrated whole-word masking strategy is adopted to avoid local co-occurrence interference