Improving Automatic Concept Extraction for Global Model Explainability