Q1. Which statement best describes Natural Language Processing (NLP)?

1. The study of programming languages for software engineering.
2. A field of AI that enables computers to process, understand, and generate human language.
3. A subfield of computer graphics focused on text rendering.
4. A branch of networking for transmitting textual data.

Select one:

A
B
C
D

Q2. Which of the following is not typically considered a core challenge in NLP?

1. Ambiguity of words and sentences.
2. Dependence on context and world knowledge.
3. Variability of language, dialects, and informal usage.
4. Limited floating-point arithmetic precision in GPUs.

Select one:

A
B
C
D

Q3. Which pair correctly matches a linguistic level with its description?

1. Phonology – meaning of sentences in context.
2. Morphology – structure of words and their internal components.
3. Syntax – sound patterns in spoken language.
4. Pragmatics – word segmentation in raw text.

Select one:

A
B
C
D

Q4. (Short answer) Briefly explain the difference between a token and a type in corpus linguistics.

Q5. Which of the following operations is most closely associated with lemmatisation rather than stemming?

1. Truncating “running” to “run” by simple suffix stripping.
2. Mapping “better” to “good” using vocabulary and POS information.
3. Removing all punctuation symbols from the text.
4. Converting all characters to lowercase.

Select one:

A
B
C
D

Q6. Which representation best describes a bag-of-words (BoW) model?

1. A sequence model that preserves full word order.
2. A multiset of terms where word order is ignored and only counts matter.
3. A character-level model that encodes every letter individually.
4. A tree structure representing full syntactic parses.

Select one:

A
B
C
D

Q7. In classical text classification, which of the following combinations is most standard and effective?

1. One-hot encoding + k-means clustering.
2. TF–IDF features + linear SVM or logistic regression.
3. Raw characters + decision tree.
4. POS tags only + Naïve Bayes.

Select one:

A
B
C
D

Q8. (Short answer) Name one key advantage of Naïve Bayes for text classification, and one important limitation.

Q9. The distributional hypothesis underlying word embeddings is often summarised as:

1. “Words are fully independent from their context.”
2. “You shall know a word by the company it keeps.”
3. “Semantics is purely syntactic.”
4. “Each word has exactly one meaning.”

Select one:

A
B
C
D

Q10. Which of the following is a key characteristic of Word2Vec embeddings?

1. They are hand-crafted rule-based features.
2. They are dense, low-dimensional vectors learned by predicting word co-occurrence.
3. They are sparse vectors using TF–IDF weighting.
4. They directly encode parse-tree structures.

Select one:

A
B
C
D

Q11. What is the main advantage of FastText compared with standard Word2Vec?

1. FastText uses only character-level information and discards all word-level context.
2. FastText represents words as bags of character n-grams, enabling better handling of rare and out-of-vocabulary words.
3. FastText does not require training data.
4. FastText is purely rule-based and does not learn parameters.

Select one:

A
B
C
D

Q12. (Short answer) Static word embeddings assign a single vector per word type. Briefly explain why this is problematic for polysemous words.

Q13. Which of the following problems is most closely associated with vanilla RNNs when modelling long sequences?

1. Inability to process variable-length input.
2. Vanishing and exploding gradients during backpropagation through time.
3. Deterministic outputs given random inputs.
4. Complete inability to learn any temporal dependencies.

Select one:

A
B
C
D

Q14. Long Short-Term Memory (LSTM) networks introduce gates primarily to:

1. Reduce the model size by removing hidden states.
2. Control information flow, enabling the network to learn what to remember and what to forget.
3. Directly compute parse trees from text.
4. Eliminate the need for backpropagation.

Select one:

A
B
C
D

Q15. In the scaled dot-product attention mechanism, what is the purpose of dividing the dot product QKᵀ by √dk before applying softmax?

1. To reduce memory usage by compressing matrices.
2. To normalise the attention scores and prevent large values from causing extremely peaked softmax distributions.
3. To guarantee that attention weights sum to one.
4. To remove all negative values from attention scores.

Select one:

A
B
C
D

Q16. (Short answer) Briefly describe the role of the encoder and the decoder in a sequence-to-sequence (seq2seq) architecture for machine translation.

Q17. Which of the following is a key reason why Transformers replaced RNNs in many NLP tasks?

1. Transformers cannot model long-range dependencies.
2. Transformers rely on self-attention, enabling parallel computation and effective modelling of global context.
3. Transformers require no training data.
4. Transformers only work for images, not text.

Select one:

A
B
C
D

Q18. Large Language Models (LLMs) like GPT are typically trained with which primary objective?

1. Predicting the next character in binary machine code only.
2. Predicting the next token in a sequence given all previous tokens (autoregressive language modelling).
3. Labelling each token with a part-of-speech tag.
4. Minimising the size of the vocabulary.

Select one:

A
B
C
D

Q19. Which of the following is an example of a risk associated with deploying LLMs in real-world applications?

1. They always produce perfectly accurate outputs, so no risk exists.
2. They may generate fluent but factually incorrect content (“hallucinations”).
3. They are mathematically incapable of learning any biases from data.
4. They can only process numeric data, not natural language.

Select one:

A
B
C
D

Q20. (Short answer) Give one example of how prompt engineering can improve the output quality of an LLM on a complex task.