NLP Machine Translation 2 — Questions and Answers
Question 1: Which decoding strategy in neural MT generates the most probable translation by greedily selecting the highest-probability token at each step?
- Beam search
- Greedy decoding (Correct answer)
- Top-k sampling
- Nucleus sampling
Correct answer: Greedy decoding
Greedy decoding selects the single highest-probability token at each step without exploring alternatives.
Question 2: What does the BLEU score primarily measure in machine translation evaluation?
- Semantic similarity between source and translation
- N-gram precision of the translation against reference(s) (Correct answer)
- Recall of content words from the source
- Fluency as rated by human judges
Correct answer: N-gram precision of the translation against reference(s)
BLEU measures n-gram precision of the MT output against one or more human reference translations.
Question 3: In the transformer encoder-decoder architecture for MT, what is the role of cross-attention?
- Allows the encoder to attend to itself
- Allows the decoder to attend to encoder outputs (Correct answer)
- Computes attention between source and target vocabularies
- Normalizes attention scores across heads
Correct answer: Allows the decoder to attend to encoder outputs
Cross-attention in the decoder lets each target position attend over all encoder hidden states to gather source context.
Question 4: Which phenomenon occurs when an NMT model repeatedly generates the same phrase or omits parts of the source sentence?
- Exposure bias
- Over-translation and under-translation (Correct answer)
- Catastrophic forgetting
- Attention collapse
Correct answer: Over-translation and under-translation
Over-translation (repeated content) and under-translation (omitted content) are common NMT failure modes caused by imperfect attention alignment.
Question 5: What is 'back-translation' used for in low-resource neural machine translation?
- Translating the model weights back to source language
- Generating synthetic source sentences from target monolingual data (Correct answer)
- Reversing the encoder-decoder order during inference
- Validating translations by re-translating to the source
Correct answer: Generating synthetic source sentences from target monolingual data
Back-translation uses a reverse MT system to translate target-language monolingual text into the source, creating synthetic parallel data for training.
Question 6: Which metric penalizes MT output that is shorter than the reference translation?
- METEOR
- BLEU (brevity penalty) (Correct answer)
- TER
- chrF
Correct answer: BLEU (brevity penalty)
BLEU includes a brevity penalty that reduces the score when the hypothesis is shorter than the reference.
Question 7: In statistical machine translation, what does an 'alignment model' estimate?
- The probability of a target sentence given source
- The correspondence between source and target words (Correct answer)
- The language model score for the target sentence
- The number of phrase pairs in the translation table
Correct answer: The correspondence between source and target words
An alignment model (e.g., IBM Models) estimates the probability that source word i generated target word j.
NLP Machine Translation 2
Which decoding strategy in neural MT generates the most probable translation by greedily selecting the highest-probability token at each step?