1.
Analyze the architectural differences between CBOW and Skip-gram models in
Word2Vec. Given the sentence "The quick brown fox jumps over the lazy dog", with a
window size of 2, illustrate the training pairs generated by each model and explain in which
scenarios each model is more effective.
(a) Architectural diagram/description of CBOW and Skip-gram (3)
(b) Correct enumeration of training pairs for both models from the given sentence (2)
2. Analyze the vanishing gradient problem in Recurrent Neural Networks (RNNs). Using the
gating mechanism of LSTM (Long Short-Term Memory), explain with equations and a
diagram how LSTM overcomes this problem. Apply your analysis to a language modelling
task involving long-distance dependencies.
(a) Explanation of vanishing gradient problem in RNNs with mathematical justification (3)
(b) Application to language modelling with long-distance dependency example (2)
3. Evaluate how GloVe differs from Word2Vec in its training objective and co-occurrence
statistics. Then, assess how word embeddings can be applied to measure text coherence
in discourse segmentation. Provide a concrete example of coherence scoring using cosine
similarity between sentence embeddings.
(a) GloVe objective function and global co-occurrence matrix construction explained (5)
(b) Comparison with Word2Vec local context window approach (2)
(c) Application of embeddings to coherence scoring with cosine similarity example (3)
4. Evaluate the pre-training strategy differences between BERT and RoBERTa. Then,
design a fine-tuning pipeline using BERT for a sentiment classification task on movie
reviews. Justify the choice of architecture modifications, loss function, and hyperparameter
settings.
(a) Comparison of BERT (MLM + NSP) vs RoBERTa (dynamic masking, no NSP) pre-
training (3)
(b)Fine-tuning pipeline design: tokenization, [CLS] representation, classification head (4)
(c) Loss function choice (cross-entropy) and hyperparameter justification (lr, epochs, batch)
(3)
5. Design an end-to-end summarization system that dynamically selects between extractive
and abstractive summarization strategies based on document type (e.g., news article vs.
scientific paper). Justify your design decisions with appropriate models, selection criteria,
and evaluation metrics. Propose how you would evaluate the system's performance.
(a) System architecture with decision logic for extractive vs abstractive selection. (4)
(b) Model selection rationale (e.g., TextRank for extractive; BART/T5 for abstractive) (3)
(c) Evaluation strategy using ROUGE, BERTScore, human evaluation framework. (2)
(d) Creativity, feasibility, and clarity of the proposed system design. (2)
6. You are tasked with building a customer service chatbot for an e-commerce platform that
must handle both retrieval-based and generative responses. Create a hybrid architecture,
justify when to use retrieval vs. generation, and critically evaluate how you would
incorporate machine translation to make it multilingual. Propose metrics for end-to-end
evaluation.
(a) Hybrid chatbot architecture: intent classification, retrieval module, generative
fallback. (3)
(b) Decision logic for retrieval vs. generation with examples.(2)
(c) Machine translation integration (MT pipeline: source → translate → respond →
translate back). (3)
Evaluation framework: BLEU for MT, F1 for intent, human MOS for chatbot quality.(3)
[Link] ELMo and static word embeddings (Word2Vec/GloVe) in terms of
contextualization. Given the sentences “He went to the bank to deposit money” and “The
river bank was flooded”, analyze how ELMo generates different embeddings for “bank”.
Design a pipeline to use ELMo embeddings in a Named Entity Recognition (NER) task.
(a) Explanation of contextual embeddings in ELMo vs static embeddings (3)
(b) Illustration of context-dependent representations with example (2)
(c) Pipeline design for NER using ELMo embeddings (tokenization, BiLSTM, CRF layer)
(4)
(d) Evaluation metrics for NER (precision, recall, F1-score) (1)
8. Analyze the attention mechanism in sequence-to-sequence models. Using a machine
translation example, derive the attention score computation (dot-product or additive
attention) and visualize alignment between source and target sentences. Evaluate how
attention improves translation quality over vanilla encoder-decoder models.
(a) Mathematical formulation of attention mechanism (4)
(b) Example with alignment visualization (3)
(c) Comparison with vanilla Seq2Seq without attention (2)
(d) Discussion of performance improvements (1)
[Link] the Transformer architecture with emphasis on self-attention and positional
encoding. Given a short sentence, demonstrate how self-attention computes contextual
representations. Compare Transformers with RNNs in terms of parallelization and long-
range dependency handling.
(a) Architecture explanation: multi-head attention, feed-forward layers (4)
(b) Self-attention computation example (3)
(c) Role of positional encoding (2)
(d) Comparison with RNNs (1)
[Link] a text classification system using traditional machine learning approaches (TF-
IDF + SVM). Then compare it with a deep learning approach (BERT-based classifier).
Evaluate both approaches on accuracy, interpretability, and computational cost.
(a) TF-IDF feature extraction and SVM pipeline (3)
(b) BERT-based classification architecture (3)
(c) Comparative evaluation (accuracy, interpretability, efficiency) (3)
(d) Justification of use-case suitability (1)
[Link] Named Entity Recognition (NER) using Conditional Random Fields (CRF).
Given a tagged sentence, compute the most probable tag sequence using Viterbi decoding.
Compare CRF with BiLSTM-CRF models in terms of feature learning and performance.
(a) CRF formulation and sequence labeling explanation (4)
(b) Viterbi decoding example (3)
(c) Comparison with BiLSTM-CRF (2)
(d) Practical applications of NER (1)
[Link] different language modelling approaches: n-grams, RNNs, and Transformers.
Given a sample corpus, compute bigram probabilities and compare them with neural
language model predictions. Discuss perplexity as an evaluation metric.
(a) Bigram probability computation with example (3)
(b) Comparison with RNN and Transformer language models (3)
(c) Explanation of perplexity and its interpretation (2)
(d) Limitations of n-gram models (2)
[Link] a question answering (QA) system using a pre-trained Transformer (e.g., BERT).
Explain how context and question are encoded, and how answer spans are predicted.
Evaluate performance using Exact Match (EM) and F1 score.
(a) QA architecture using Transformer encoders (4)
(b) Input representation (question + context) (2)
(c) Span prediction mechanism (2)
(d) Evaluation metrics (EM, F1) (2)
[Link] sentiment analysis using lexicon-based and deep learning approaches. Given
a set of sentences, compute sentiment scores using a lexicon method and compare with
predictions from an LSTM-based model. Discuss challenges such as sarcasm and domain
adaptation.
(a) Lexicon-based sentiment scoring method (3)
(b) LSTM-based sentiment model architecture (3)
(c) Comparative analysis with examples (3)
(d) Challenges in sentiment analysis (1)
[Link] a neural machine translation (NMT) system and compare it with statistical
machine translation (SMT). Explain encoder-decoder architecture with attention and
evaluate translation quality using BLEU score.
(a) NMT architecture with encoder-decoder and attention (4)
(b) Comparison with SMT approaches (3)
(c) BLEU score computation and interpretation (2)
(d) Limitations and improvements in NMT (1)