Q1
Walk me through how you would approach building an end-to-end NLP pipeline for a text classification task. What preprocessing steps would you include, and how would you decide between using a pre-trained transformer model versus fine-tuning one from scratch?
Why they ask this:* They're assessing your practical understanding of NLP workflow design, trade-offs between model complexity and computational resources, and your familiarity with modern transformer architectures commonly used in production systems.
Q2
Describe your experience with handling imbalanced datasets in NLP tasks. What techniques have you used, and how did you evaluate whether they improved model performance on minority classes?
Why they ask this:* This tests your awareness of real-world data challenges and your ability to select appropriate metrics (precision, recall, F1, AUC-ROC) beyond accuracy—critical for production NLP systems where class imbalance is common.
Q3
How would you implement and optimize a named entity recognition (NER) model for a specific domain? Walk me through your choice of architecture, training approach, and how you'd handle limited labeled data in that domain.
Why they ask this:* They want to understand your hands-on experience with sequence labeling tasks, transfer learning strategies, and practical solutions for domain adaptation—a frequent challenge in industry NLP projects.
Q4
Explain the difference between absolute positional encoding and relative positional encoding in transformer models. When would you choose one over the other, and what impact would this have on inference speed and memory usage?