Soft-prompt Tuning for Large Language Models to Evaluate Bias