Neural Text Sanitization with Privacy Risk Indicators: An Empirical Analysis