Because the extrinsic evaluations are realistic but expensive, you have learned an intrinsic evaluation metric, the perplexity PP(W) = (W)/", where n indicates the length of the test data W = ww2...wn by concatenating all sentences. You measured q(W) by a language model learned on a training data. In particular, you have two different but same amount of datasets D: whether reports and D: phone conversations. Two trigram models 9 and 92 are learned on D and D2, respectively. Suppose you measure the perplexity of an unseen weather reports data with 9, and the perplexity of an unseen phone conversation data of the same length with q2. Explain which perplexity is likely higher than the other. Q3.2 5 Points Now you measure the perplexity of an unseen weather reports data with 92, and the perplexity of an unseen phone conversation data of the same length with q. If the perplexity measured with 91 is lower, can you conclude that q is a better model than 92? Yes O No Q3.3 5 Points When you are evaluating the perplexity as stated above, what will happen if the test data consists of some unseen words that you have never observed in the training data? (select all correct) If use smoothing method, p(unseen word)>0 If use smoothing method, p(unseen word)=0 If use smoothing method, we can compute perplexity If use smoothing method, we can not compute perplexity