NLP Advanced Topics & Theory 5 — Questions and Answers
Question 1: What is 'zero-shot chain-of-thought' prompting?
- Providing zero training examples and using a chain of pre-trained models in sequence
- Appending a phrase like 'Let's think step by step' to elicit reasoning without any examples (Correct answer)
- Generating chains of paraphrases to augment zero-shot training data
- A decoding algorithm that explores reasoning chains via beam search
Correct answer: Appending a phrase like 'Let's think step by step' to elicit reasoning without any examples
Zero-shot CoT adds a simple reasoning trigger phrase to the prompt, causing LLMs to produce intermediate reasoning steps before answering.
Question 2: Which evaluation challenge is specific to open-ended text generation compared to classification tasks?
- Open-ended generation has no ground-truth labels, making automatic evaluation inherently difficult (Correct answer)
- Generation models cannot be evaluated on held-out test sets
- Classification metrics like F1 generalize perfectly to generation
- Open-ended generation is always evaluated via human ratings, never automatically
Correct answer: Open-ended generation has no ground-truth labels, making automatic evaluation inherently difficult
Unlike classification, many valid outputs exist for generation tasks and no single reference captures all correct answers, making standard metrics insufficient.
Question 3: What is the primary advantage of retrieval-augmented generation (RAG) over purely parametric LLMs?
- RAG models train faster by skipping the pre-training phase
- RAG grounds responses in retrieved external documents, reducing hallucination and enabling knowledge updates without retraining (Correct answer)
- RAG eliminates the need for attention mechanisms
- RAG stores all knowledge in a key-value cache inside the model weights
Correct answer: RAG grounds responses in retrieved external documents, reducing hallucination and enabling knowledge updates without retraining
RAG dynamically retrieves relevant passages at inference time, allowing the model to cite current or domain-specific knowledge beyond what was memorized during training.
Question 4: In NLP, what is 'distributional semantics'?
- Distributing computation across multiple GPUs for faster embedding training
- The hypothesis that words with similar meanings appear in similar contexts (Correct answer)
- Mapping discrete tokens to continuous distributions via variational autoencoders
- A method of distributing a vocabulary evenly across transformer attention heads
Correct answer: The hypothesis that words with similar meanings appear in similar contexts
Distributional semantics—'you shall know a word by the company it keeps'—underpins word2vec and GloVe by representing meaning through co-occurrence statistics.
Question 5: What distinguishes 'extractive' from 'abstractive' summarization?
- Extractive models generate new sentences; abstractive models copy spans verbatim
- Extractive models select and copy spans from the source; abstractive models generate novel text (Correct answer)
- Extractive summarization uses neural networks; abstractive summarization uses rule-based methods
- Extractive models require fine-tuning; abstractive models work zero-shot only
Correct answer: Extractive models select and copy spans from the source; abstractive models generate novel text
Extractive summarization lifts sentences directly from the document, while abstractive summarization paraphrases and synthesizes information into new text.
Question 6: What is the 'vanishing gradient' problem and why did it historically limit RNN performance on long sequences?
- Gradients grew too large during backpropagation, causing weight explosion in long sequences
- During backpropagation through time, gradients diminish exponentially, making it hard to learn long-range dependencies (Correct answer)
- The optimizer converged to local minima because gradients were too noisy on short sequences
- Vanishing gradients occur only in feed-forward networks, not recurrent ones
Correct answer: During backpropagation through time, gradients diminish exponentially, making it hard to learn long-range dependencies
Multiplying many small Jacobians through long RNN unrolled steps makes gradients shrink toward zero, preventing updates to early time-step weights.
Question 7: Which of the following best describes 'constitutional AI' as an alignment approach?
- Training an LLM exclusively on texts derived from legal constitutions
- Using a set of written principles to guide a model to critique and revise its own outputs without human labeling (Correct answer)
- A hardware-level constraint that prevents models from generating harmful tokens
- An RL method where the reward function is derived from constitutional law
Correct answer: Using a set of written principles to guide a model to critique and revise its own outputs without human labeling
Constitutional AI (Anthropic) has the model self-critique responses against a written list of principles and revise them, scaling alignment feedback without exhaustive human annotation.
NLP Advanced Topics & Theory 5
What is 'zero-shot chain-of-thought' prompting?