NLP deep dive: IE, multilingual, code generation

Interview questions to prep

1. Pick between Dijkstra, Bellman-Ford, Floyd-Warshall, MST (Prim/Kruskal), or topo sort — defend the choice.
2. What does this problem assume about edge weights (non-negative? integer? bounded?) — and what breaks if those don't hold?
3. Walk me through complexity in V and E, and the data-structure choice (heap vs Fibonacci heap vs array).

Interview questions to prep

1. Compare span-based vs sequence-tagging IE — when does each win?
2. How would you design a relation-extraction system that handles entities not seen in training?

Interview questions to prep

1. How does mBERT/XLM-R enable cross-lingual transfer, and where does it break?
2. What's the curse of multilinguality, and how do you mitigate it for low-resource languages?

Interview questions to prep

1. How is training a code model different from training a natural-language model?
2. How would you evaluate a code generation model — pass@k, HumanEval, SWE-bench?

Interview questions to prep

1. When would you choose BERT-style encoder-only models over GPT-style decoder-only models?
2. When is encoder-decoder still the cleanest choice for summarization, translation, or extraction?
3. How would latency and serving cost affect that architecture choice?