Check how robust a sentiment classifier is to random typos in our dataset.
Given sentence A:
I just love the new Honda I bought from Crystal.
we can introduce a random typo to get sentence A':
I jpst love the new Honda I bought from Crystal.
Now we test whether our sentiment classifer S gives the same answer for S(A) and S(A'). If S(A) == S(A'), that's good! We can do this over hundreds or thousands of example sentences and estimate what percent of examples S(A) == S(A').