Healing Unsafe Dialogue Responses with Weak Supervision Signals