In Naive Bayes, we use probabilities of particular features values rather than calculating the probability of the Evidence as a whole. We do this because we assume that there will be many occurrences of each e_i. However, if those occurrence counts are small (or zero!) our probabilities are likely to be underestimated. We've seen this problem before in other methods for class probability estimation. Use the same correction method to calculate an estimated p(e_i) given that the count of e_i in your training data is just 1 out of the 1000 training examples.