Language Modeling DATA-SCIENCE-OPTIMIZE Practice Questions (Page 4)

Questions 16–20

1. 16expert · hard

   A data scientist is building a bigram model for a corpus with a vocabulary of 10,000 words. The model will be used for a real-time application where speed is critical. The scientist is considering Laplace smoothing and Kneser-Ney smoothing. Which factor is most important in this decision?

   • ALaplace smoothing is computationally simpler and faster, which is critical for real-time applications
   • BKneser-Ney is always more accurate than Laplace
   • CLaplace smoothing is always better for large vocabularies
   • DKneser-Ney requires more memory, which is not available
   Reveal Solution Discussion Select an answer first
2. 17application · medium

   A research team is evaluating two neural language models for a machine translation task. Model X is an RNN, and Model Y is a transformer. Both are trained on the same data. On the validation set, Model Y achieves a lower perplexity than Model X. What does this indicate?

   • AThe perplexity difference is not meaningful because the models are different architectures
   • BModel Y is better because lower perplexity indicates it assigns higher probability to the validation data
   • CModel X is better because RNNs are more interpretable
   • DModel X is better because RNNs are faster to train
   Reveal Solution Discussion Select an answer first
4. 18application · medium

   A data scientist is building a bigram language model for a small corpus of customer support tickets. The model must assign a non-zero probability to every possible word pair, even those that never appear in the training data. Which approach should the data scientist use?

   • AApply Laplace smoothing to the bigram counts
   • BUse a unigram model instead of a bigram model
   • CUse maximum likelihood estimation without any smoothing
   • DIncrease the corpus size until all bigrams appear
   Reveal Solution Discussion Select an answer first
5. 19application · medium

   A team is evaluating a bigram language model on a test set of 100 sentences. The model's perplexity is 200. What does this value indicate about the model's performance?

   • AThe model is uncertain, on average, among 200 equally likely choices for the next word
   • BThe model is worse than a unigram model
   • CThe model is perfect because perplexity is always positive
   • DThe model assigns a probability of 200 to each word
   Reveal Solution Discussion Select an answer first
6. 20foundation · easy

   Which smoothing technique adds a small constant (typically 1) to the count of every n-gram?

   • ALaplace smoothing
   • BKneser-Ney smoothing
   • CInterpolation
   • DBackoff
   Reveal Solution Discussion Select an answer first