KIMV: Enhancing Entity Disambiguation with a Knowledge-Integrated Multi-View Encoder

Entity Disambiguation (ED) is a critical task in Natural Language Processing that links ambiguous entity mentions in text to their canonical entries in a knowledge base. While modern deep learning models have advanced the state of the art, they face persistent challenges. Classification-based methods often struggle with rare or "overshadowed"entities, as their knowledge is implicitly stored in model parameters. Generative methods, conversely, can overlook the rich semantic information in entity descriptions crucial for resolving fine-grained ambiguity, often relying on surface-level entity titles. To address these gaps, we introduce KIMV, a novel Knowledge-Integrated MultiView framework. KIMV features two key innovations: (1) a Word Knowledge Tree (WKT), a purpose-built prefix tree that efficiently integrates rich background knowledge for candidate entities prior to encoding, enhancing the input representation in a manner distinct from output-constrained decoding approaches; and (2) a multi-view dual-encoder that computes a composite relevance score by synergistically combining dense semantic matching, sparse lexical alignment, and a token-level late-interaction mechanism. Experimental evaluations on six benchmark datasets, including AIDA, MSNBC, and AQUAINT, demonstrate that KIMV achieves a new state-of-the-art average InKB Micro F1 score of 89.2%. The model shows particularly strong performance on challenging out-of-domain datasets like AQUAINT and CWEB, underscoring the effectiveness of its explicit, pre-encoder knowledge integration strategy in achieving superior generalization.