In the lecture, the model is defined as Pi (P(wi|Pos) / P(wi|Neg)). But given a tweet, since we already knows the condition that wi is in the tweet, and we want to measure whether the sentiment is positive or negative, why don’t we use Pi (P(Pos | wi) / P(Neg | wi))?
As I continue to the log likelihood video, I saw that the score is extended to
In that way, it is actually equivalent to Pi (P(pos|wi) / P(neg|wi)). So I think that answered my question. 
2 Likes