You can test on a tester of use some of the examples from the checklist paper. In that paper they might add "I hate you" to some random data and assert that sentiment doesn't improve.