Natural Language Processing Natural Language Processing MCQ 3 — Questions and Answers
Question 1: Which loss function is most commonly used for training neural machine translation models?
- Mean squared error
- Cross-entropy loss (Correct answer)
- Hinge loss
- Kullback-Leibler divergence
Correct answer: Cross-entropy loss
Cross-entropy loss measures the difference between predicted token probability distributions and the one-hot target distributions during NMT training.
Question 2: What problem does 'label smoothing' address in sequence-to-sequence NLP models?
- Gradient vanishing in deep decoder stacks
- Overconfident predictions that hurt generalization (Correct answer)
- Misalignment between source and target sentence lengths
- Slow convergence of the attention mechanism
Correct answer: Overconfident predictions that hurt generalization
Label smoothing replaces hard 0/1 targets with soft distributions, preventing the model from becoming overconfident on training examples.
Question 3: In the ROUGE metric used for text summarization evaluation, what does ROUGE-L specifically measure?
- Longest common subsequence between candidate and reference (Correct answer)
- Number of unigram overlaps divided by reference length
- Bigram precision and recall averaged with F1
- Lexical diversity of the generated summary
Correct answer: Longest common subsequence between candidate and reference
ROUGE-L measures the longest common subsequence (LCS) between candidate and reference summaries, capturing sentence-level structure.
Question 4: What is 'zero-shot classification' in the context of large language models?
- Classifying text without any labeled training examples using natural language prompts (Correct answer)
- Training a classifier from scratch on a single GPU
- Using transfer learning with zero frozen layers
- Evaluating model performance before any fine-tuning
Correct answer: Classifying text without any labeled training examples using natural language prompts
Zero-shot classification uses an LLM's pre-trained knowledge and natural language descriptions to classify text without task-specific training examples.
Question 5: Which component of a transformer architecture is responsible for the 'parallelization advantage' over RNNs?
- Feed-forward network layers
- Self-attention mechanism (Correct answer)
- Layer normalization
- Residual connections
Correct answer: Self-attention mechanism
Self-attention computes relationships between all token pairs simultaneously, unlike RNNs which process tokens sequentially and cannot be parallelized across time steps.
Question 6: What is 'semantic role labeling' (SRL) in NLP?
- Assigning sentiment polarity to phrases in a sentence
- Identifying the predicate-argument structure to label who did what to whom (Correct answer)
- Clustering semantically similar sentences together
- Tagging each token with its grammatical part of speech
Correct answer: Identifying the predicate-argument structure to label who did what to whom
SRL identifies the roles of words in a sentence relative to predicates, such as Agent, Patient, and Instrument, answering who did what to whom.
Question 7: What distinguishes 'extractive' summarization from 'abstractive' summarization?
- Extractive uses neural networks while abstractive uses rule-based systems
- Extractive selects and copies existing sentences while abstractive generates new text (Correct answer)
- Extractive produces longer summaries while abstractive produces shorter ones
- Extractive requires labeled data while abstractive is unsupervised
Correct answer: Extractive selects and copies existing sentences while abstractive generates new text
Extractive summarization selects verbatim sentences from the source document, whereas abstractive summarization generates novel paraphrased text.
Natural Language Processing Natural Language Processing MCQ 3
Which loss function is most commonly used for training neural machine translation models?