A process receives a user input in a human-to-machine interaction. The process generates, with a natural language generation engine, one or more response candidates. Further, the process measures, with the natural language generation engine, the semantic similarity of the one or more response candidates. In addition, the process selects, with the natural language generation engine, a response candidate from the one or more response candidates. The process measures, with the natural language generation engine, an offensiveness measurement and a politeness measurement of the selected response. The process determines, with the natural language generation engine, that the offensiveness measurement or the politeness measurement lacks compliance with one or more predefined criteria. The process selects, with the natural language generation engine, an additional response candidate from the one or more response candidates that has a higher semantic similarity measurement than remaining response candidates from the one or more response candidates.