Skills extraction is crucial for understanding labor market trends and aligning workforce capabilities with industry demands, but existing methods often lack generalizability and struggle with domain-specific skill taxonomy. This research introduces a novel methodology for skills extraction that identifies and maps skills from unstructured text to a predefined skill database or taxonomy. Our approach consists of three key steps: (1) employing BERT-based span extractors to identify skill mentions, (2) using sentence embedding models for semantic mapping to an existing skills taxonomy, and (3) applying a re-ranking model to assess skill importance based on relevance. In addition, we implemented several optimizations to enhance inference speed without compromising performance. Experimental results confirm that our optimized approach improves skills extraction accuracy and processing efficiency across domains, offering a replicable framework for internal taxonomies.