Predict first: If a word never appeared in class c during training, what happens to the entire document likelihood when α = 0?
Training corpus
Test document
Class scores
Predicted class—
Posterior-like normalized score—
Why does word order not matter?
The bag-of-words model uses counts, not positions. Under the multinomial Naïve Bayes assumptions, the same multiset of words gives the same class score even after shuffling.
Laplace smoothing prevents a single unseen word from forcing the class likelihood to zero.