Machine Learning Natural Language Processing 4 — Questions and Answers
Question 1: What is the role of positional encoding in the original Transformer model?
- To normalize token embeddings before attention
- To inject sequence order information since attention is permutation-invariant (Correct answer)
- To reduce the dimensionality of token embeddings
- To initialize the softmax output layer
Correct answer: To inject sequence order information since attention is permutation-invariant
Positional encodings add order information to token embeddings because self-attention treats all positions equivalently without them.
Question 2: In text classification, what does the macro-averaged F1 score compute?
- F1 weighted by class frequency
- The unweighted mean of per-class F1 scores (Correct answer)
- F1 on only the majority class
- F1 computed on the full token sequence
Correct answer: The unweighted mean of per-class F1 scores
Macro-averaged F1 calculates F1 independently for each class and averages them equally, treating all classes as equally important regardless of size.
Question 3: Which NLP task involves assigning a structured predicate-argument representation to a sentence, identifying who did what to whom?
- Semantic Role Labeling (SRL) (Correct answer)
- Named Entity Recognition
- Sentiment Analysis
- Textual Entailment
Correct answer: Semantic Role Labeling (SRL)
SRL labels arguments of predicates (verbs) with semantic roles like Agent, Patient, and Instrument, revealing the meaning structure of a sentence.
Question 4: What does 'hallucination' mean in the context of large language models?
- The model refusing to generate text
- The model generating fluent but factually incorrect or fabricated content (Correct answer)
- The model repeating input tokens verbatim
- The model producing very short outputs
Correct answer: The model generating fluent but factually incorrect or fabricated content
Hallucination refers to LLMs confidently generating plausible-sounding text that is factually wrong or unsupported by the provided context.
Question 5: Which decoding strategy introduces randomness by sampling from the top-p cumulative probability mass rather than the full vocabulary?
- Greedy decoding
- Beam search
- Nucleus (top-p) sampling (Correct answer)
- Temperature=0 sampling
Correct answer: Nucleus (top-p) sampling
Nucleus sampling restricts the sampling pool to the smallest set of tokens whose cumulative probability exceeds p, balancing diversity and coherence.
Question 6: In Retrieval-Augmented Generation (RAG), what is the primary role of the retriever component?
- To fine-tune the language model on domain data
- To fetch relevant documents from an external knowledge base given a query (Correct answer)
- To rerank the model's output tokens by likelihood
- To tokenize the user query before generation
Correct answer: To fetch relevant documents from an external knowledge base given a query
The retriever fetches semantically relevant passages from an indexed corpus, which are then provided as context to the generator to ground its responses.
Question 7: What is the key difference between extractive and abstractive summarization?
- Extractive summarization uses neural networks; abstractive does not
- Extractive copies spans from the source; abstractive generates novel text (Correct answer)
- Abstractive summarization is always shorter than extractive
- Extractive models require more training data
Correct answer: Extractive copies spans from the source; abstractive generates novel text
Extractive summarization selects and concatenates sentences from the original document, while abstractive summarization paraphrases and generates new phrasing.
Machine Learning Natural Language Processing 4
What is the role of positional encoding in the original Transformer model?