Natural Language Processing — DU B.Sc. (Hons) Computer Science Previous Year Question Papers

(b) What are the major ethical issues in Large Language Models? Explain any four with real-world implications. (5) (c) Explain the three types of Transformer-based language model architectures with suitable diagram. Parts of a question should be attempted together. 1. (a) What are stopwords? Why are stopwords removed in text processing? (3) (b) What is tokenization in natural language processing and why is it required? (3) (c) State the Markoy assumption that underlies bigram language models and write the bigram probability formula it produces. Give one limitation of the bigram language model. (3) (d) Explain the Bag-of-Words (Bow) model and its limitations. How does Word2Vec address these limitations? (3) (e) Why are word embeddings important in NLP, and how do they differ from one-hot encoding? (3) (I) State one application for each of the three types of Transformer-based models. (3) (g) What is pooling in neural NLP models, and how do mean pooling and max pooling work? (c) Explain the XOR problem in neural networks. Why is it not linearly separable, and how does a multi-layer perceptron overcome this limitation? (5) 6. (a) Write short notes on Any two of the following: I. Feed forward Neural Networks 11. Recurrent Neural Networks Ill. Acti.vation Functions (6) (b) What is self-attention in Transformer models, and why is it important for language modeling? Also, explain how the Query, Key, and Value matrices are used to compute attention. (6) (c) What is prompting in large language models? Explain prompt engineering with one example. 7. (a) Explain how text is generated in language models. Describe any one sampling methods used during generation. (4) (b) Compute the TF, IDF, and TF-IDF score for the term "ccz/" in document Dl given the corpus: D1 -- "the cat sat", D2. = "the cat ate", D3 ~~ C¢the dog sat". (c) Compute the cosine similarity between the vectors A=[1,2,3] and 8=[2,1,2]. Show the dot product and magnitude calculations, and interpret the result. 5. (a) What is part-of-Speech (POS) tagging? Perform 'i r for the seritence ``The quick brown fox jumps `i over /fee /czz)/ dog. " Assign an appropriate tag to each word. (5) Pichai joined Google in Mountain View on A4lo7cczey. identify all named entities and assign appropriate 810 tags along with their entity types. (h) Explain perplexity as a metric for language model evaluation. If a bigram model assigns probability 0.001 to a test sentence of 10 words, compute its perplexity. (3 ) (i) Why is Miningum Edit Distance computed using dynamic pro.gramming instead of a brute-force approach, and what are the edit operations involved in calculating it? (3) 0.) Define Term Frequency-Inverse Document Frequency (TF-IDF). Exp`1ain its significance in text representation. (3) 2. (a) Given the regular expression r'\b[A-Z][a-z]+\b' and the text "Alice went to Paris in January". I. List all the words matched by the pattern. IIe Explain the purpose of each part of the pattern (\b, [A-Z], [a-z]+).