Natural Language Processing Machine Translation and Sequence-to-Sequence Models 2 — Questions and Answers
Question 1: What is text summarization in NLP?
- The task of automatically producing a shorter version of a document that retains the most important information (Correct answer)
- The task of classifying a document into one of several topic categories
- The process of extracting all named entities from a document
- A method for detecting duplicate sentences in a corpus
Correct answer: The task of automatically producing a shorter version of a document that retains the most important information
Summarization can be extractive (selecting key sentences) or abstractive (generating new sentences), both aiming to condense source content while preserving meaning.
Question 2: What is the difference between extractive and abstractive summarization?
- Extractive summarization selects existing sentences from the source, while abstractive summarization generates novel sentences that may not appear verbatim in the source (Correct answer)
- Extractive summarization is always better quality than abstractive
- Abstractive summarization cannot handle multi-document inputs
- Extractive summarization requires a pre-trained language model while abstractive uses rules
Correct answer: Extractive summarization selects existing sentences from the source, while abstractive summarization generates novel sentences that may not appear verbatim in the source
Extractive systems copy salient spans directly; abstractive systems paraphrase and synthesize information, producing more fluent but harder-to-evaluate summaries.
Question 3: What is the ROUGE metric used for in NLP evaluation?
- Evaluating text summarization quality by measuring n-gram recall overlap between generated and reference summaries (Correct answer)
- Evaluating machine translation using n-gram precision
- Measuring named entity recognition accuracy
- Assessing the grammaticality of generated text
Correct answer: Evaluating text summarization quality by measuring n-gram recall overlap between generated and reference summaries
ROUGE (Recall-Oriented Understudy for Gisting Evaluation) measures the overlap of n-grams and longest common subsequences between system and reference summaries.
Question 4: What is the copy mechanism in sequence-to-sequence models?
- A mechanism that allows the decoder to directly copy tokens from the source input sequence rather than always generating from the vocabulary (Correct answer)
- A caching layer that stores previously computed hidden states
- A method for duplicating training examples to balance classes
- A technique for initializing decoder weights from the encoder
Correct answer: A mechanism that allows the decoder to directly copy tokens from the source input sequence rather than always generating from the vocabulary
The copy mechanism (pointer networks) lets the model point to and copy source tokens, which is critical for tasks like summarization where proper nouns or rare words should be reproduced exactly.
Question 5: What is the teacher forcing technique in training sequence-to-sequence models?
- Feeding the ground-truth previous token as input to the decoder at each step during training, rather than the model's own previous prediction (Correct answer)
- Using a teacher model to supervise a student model during knowledge distillation
- Forcing the model to attend to the teacher's annotations during training
- Providing the model with extra labeled examples from human annotators
Correct answer: Feeding the ground-truth previous token as input to the decoder at each step during training, rather than the model's own previous prediction
Teacher forcing speeds up convergence by providing correct previous tokens during training, though it can cause exposure bias since the model sees its own errors only at inference time.
Question 6: What is dialogue state tracking in task-oriented dialogue systems?
- Maintaining a structured representation of the conversation's goals and slot values across multiple turns (Correct answer)
- Generating the next system response at each dialogue turn
- Detecting the user's intent from a single utterance
- Evaluating the fluency of system responses using human raters
Correct answer: Maintaining a structured representation of the conversation's goals and slot values across multiple turns
Dialogue state tracking updates a belief state (slot-value pairs) after every user turn so the system always knows what the user wants and what has been confirmed.
Natural Language Processing Machine Translation and Sequence-to-Sequence Models 2
What is text summarization in NLP?