A hybrid approach for measuring semantic similarity in lexically identical but ambiguous sentences

Abstract

This study addresses the critical challenge of semantic similarity and lexical disambiguation in natural language processing, focusing on sentences with structural and lexical ambiguities. We introduce an innovative hybrid approach that synergistically combines symbolic and neural methods to better align with human judgment. Our methodology dynamically integrates fuzzy Jaccard’s lexical precision with SBERT embeddings’ contextual sensitivity, enabling adaptive semantic ambiguity resolution. Experimental evaluation on 33 ambiguous sentences demonstrates that our approach significantly outperforms conventional artificial intelligence (AI) systems, achieving an 11.7% reduction in mean absolute error compared to reference models, with statistical analysis confirming robust results (d = -0.80, p < 0.001). This represents a 65% improvement in human evaluation alignment over existing methods. Our research contributes to advancing the field by showing that architectural intelligence can surpass mere parameter scaling, offering an effective solution for applications requiring both precision and interpretability, with promising directions for multilingual extension and explainable AI integration.