This paper studies semantic textual relevance (STR) for job title matching, a critical task in resume recommendation systems. To improve STR beyond simple lexical similarity and capture nuanced relationships, we propose a self-supervised hybrid architecture combining dense sentence embeddings and domain-specific knowledge graphs (KGs). Unlike previous studies, we perform a hierarchical evaluation by segmenting the STR score continuum into low, medium, and high semantic relevance regions, allowing for a detailed analysis of model performance. Evaluating various embedding models with and without KGs, we demonstrate that a fine-tuned SBERT model with KGs significantly improves performance, reducing root mean square error (RMSE) by up to 25% in high-STR regions. This highlights the effectiveness of combining KGs with text embeddings and the importance of analyzing region-specific performance.
