Advanced Topics & Theory Flashcards
7 cards from real NLP practice questions. Tap to flip, then mark Knew It or Still Learning ā missed cards come back until you master them.
Read the first 7 Advanced Topics & Theory flashcards as text
What is the key idea behind Reinforcement Learning from Human Feedback (RLHF) in LLM alignment?
Answer: Training a reward model on human preferences, then using RL to optimize the LLM against that reward
RLHF trains a reward model on human comparisons of outputs, then uses PPO or similar RL algorithms to steer the LLM toward higher-reward responses.
What does the term 'hallucination' mean in the context of large language models?
Answer: The model producing fluent but factually incorrect or fabricated content
Hallucination refers to confident, fluent model outputs that are factually wrong or entirely made up, a central reliability challenge for LLMs.
Which NLP task involves identifying the opinion expressed about specific aspects of a product or service?
Answer: Aspect-based sentiment analysis (ABSA)
ABSA extracts sentiment polarity at a fine-grained level for specific attributes (e.g., 'The battery life is great but the screen is dim').
What is 'prompt injection' in the context of LLM security?
Answer: Crafting input that overrides system instructions to make the model perform unintended actions
Prompt injection attacks embed adversarial instructions in user input that hijack the model's behavior by overriding or confusing its system prompt.
What does 'sparse attention' aim to achieve compared to full (dense) self-attention?
Answer: Reduce quadratic complexity by restricting each token to attend to a subset of positions
Sparse attention (e.g., Longformer, BigBird) limits each token to attending to local windows or global tokens, reducing O(nĀ²) cost for long sequences.
In dependency parsing, what does a 'head' word determine?
Answer: The syntactically dominant word to which a dependent word is attached
In dependency grammar, each dependent word has exactly one head, and the arc between them encodes a grammatical relation such as subject or object.
What problem does byte-pair encoding (BPE) solve in NLP tokenization?
Answer: It balances vocabulary size and out-of-vocabulary coverage by merging frequent character pairs into subwords
BPE iteratively merges the most frequent adjacent byte or character pairs to build a subword vocabulary, handling rare and unseen words gracefully.