NLP Machine Translation 5 — Questions and Answers
Question 1: What is the primary purpose of the 'language model' component in a traditional statistical MT system?
- Ensuring the source sentence is grammatical
- Assigning probability to target-language word sequences to favor fluent output (Correct answer)
- Selecting phrase pairs from the translation table
- Computing alignment between source and target words
Correct answer: Assigning probability to target-language word sequences to favor fluent output
The language model scores candidate translations by their fluency in the target language, preferring grammatically natural word sequences.
Question 2: Which challenge is unique to translating morphologically rich languages (e.g., Finnish, Turkish) in NMT?
- Lack of parallel corpora for any language pair
- Data sparsity due to exponentially many word forms (Correct answer)
- Inability to use subword tokenization
- The transformer cannot model long-range dependencies
Correct answer: Data sparsity due to exponentially many word forms
Morphologically rich languages have vast inflectional paradigms, causing data sparsity since many word forms appear rarely or not at all in training data.
Question 3: In constrained decoding for MT, what is a 'soft constraint' as opposed to a 'hard constraint'?
- A constraint applied only during training, not inference
- A preference that biases the model toward certain terms without guaranteeing their inclusion (Correct answer)
- A constraint on the maximum output length
- A constraint derived from a soft-alignment model
Correct answer: A preference that biases the model toward certain terms without guaranteeing their inclusion
Soft constraints adjust the probability distribution to favor particular terms but do not guarantee they appear, unlike hard constraints that force specific tokens.
Question 4: What is 'adaptive MT' in a professional translation workflow?
- An MT system that adapts its language model at training time only
- A system that updates its model in real time based on translator corrections (Correct answer)
- MT that adapts beam width based on sentence complexity
- Translating adaptively long documents paragraph by paragraph
Correct answer: A system that updates its model in real time based on translator corrections
Adaptive MT systems update their parameters on-the-fly as human translators post-edit output, personalizing the model to a specific user or domain during a session.
Question 5: Which property of the transformer makes it more parallelizable during training compared to RNN-based MT models?
- Use of convolutional layers instead of recurrent connections
- Self-attention operates on all positions simultaneously rather than sequentially (Correct answer)
- The encoder processes tokens from right to left
- Positional encodings replace word embeddings
Correct answer: Self-attention operates on all positions simultaneously rather than sequentially
Transformer self-attention computes relationships between all token pairs in parallel, whereas RNNs must process tokens one at a time sequentially.
Question 6: What is the role of 'length normalization' in beam search for MT?
- Ensuring source and target sentences have the same number of tokens
- Dividing the log-probability score by sentence length to avoid penalizing longer hypotheses (Correct answer)
- Normalizing attention weights so they sum to one
- Rescaling embeddings by the square root of model dimension
Correct answer: Dividing the log-probability score by sentence length to avoid penalizing longer hypotheses
Without length normalization, beam search favors short translations because each additional token multiplies (reduces) the probability; dividing by length corrects this bias.
Question 7: Which evaluation approach asks bilingual judges to rate MT output on adequacy and fluency separately?
- Automatic post-editing evaluation
- Human evaluation using adequacy/fluency scales (Correct answer)
- BLEU with multiple references
- Direct Assessment (DA) on a 0-100 scale
Correct answer: Human evaluation using adequacy/fluency scales
The classic human MT evaluation protocol uses separate adequacy (meaning preserved?) and fluency (is target text natural?) rating scales judged by bilingual assessors.
NLP Machine Translation 5
What is the primary purpose of the 'language model' component in a traditional statistical MT system?