CS224N: Natural Language Processing
Spring 2008
Weekly Quiz
Quiz 2
1: The beam search technique discussed in J&M, when deciding which sentences to
expand at each step, compares scores between candidate translation sentences that
have different numbers of words by:
Normalizing the score by the length of the sentence
Comparing scores on sub-sections of the sentences that have equal length
Comparing only scores between sentences with the same number of words
None of those above
All of those above
2: Which of the following is true about EM?
If you run it long enough, it eventually reaches a global optimum.
Different initial condition may lead to different results.
You have to have completely observed data to do EM.
The data likelihood sometimes goes down after an iteration.
None of those above
3: Which of the following is FALSE about IBM Models?
Adding the fertility model makes it hard to train IBM Model 3 efficiently
IBM Model 1, even though simple, takes a long time because it has to sum
over all possible alignment structures of a sentence pair.
IBM Model 2 can be efficiently computed even though it makes use of a
distortion model
With IBM Model 3, you need to initialize the model parameters with some
good estimates instead of just random or uniform ones.
None of the above
4: Which of the following is FALSE?
The MT automatic evaluation metric BLEU is a weighted average of the
number of N-gram overlaps with the human translations.
Having multiple human reference translation makes BLEU more reliable
Since BLEU is computing n-gram overlaps, a system can get a high BLEU
score by translating every foreign sentence to one very common word (e.g.,
"the"), obtaining a high unigram precision.
