LG-CAV: Train Any Concept Activation Vector with Language Guidance | Read Paper on Bytez