Classic NLP Tasks

pre- vs post-LLM paradigms for four classic tasks. (1) Text classification — before: TF-IDF/BOW + LR/GBDT feature engineering, or BERT's [CLS] into $\text{softmax}(W^\top h_{\text{[CLS]}})$; after: zero/few-shot prompts assign labels, or distill into a small model for cost. (2) NER — before: sequence labeling BiLSTM-CRF over BIO tags; after: LLMs emit JSON entity lists, unconstrained label sets, new entity types with zero training. (3) MT — before: SMT (phrase statistics) to NMT (Seq2Seq + Attention, BLEU); after: one multilingual model plus glossary constraints in-context. (4) Summarization — before: extractive (TextRank) and abstractive (BART/Pegasus, ROUGE); after: Map-Reduce chunked summarization for very long documents.

tech-selection when migrating to LLMs; keep traditional small models where throughput, latency, or privacy dominate; a classic interview question — classifying 10M customer-support tickets, balancing LLM vs fine-tuned small models.

the traditional paradigm needs task-specific labeled data, models, and feature engineering — slow cold starts per task. The LLM paradigm reduces task differences to prompt differences: few-shot cold start and reuse, at the cost of latency/cost/stability — which is why production often runs the hybrid "LLM annotates → distill to a small model" route.

LLM vs BERT for classification: accuracy, cost, latency, and cold-start trade-offs?

Sequence-labeling NER (BIO, BiLSTM-CRF) vs LLM extraction: pros and cons of each?

Why was NMT replacing SMT a milestone, and what role did BLEU play?

How do you design Map-Reduce summarization for very long documents, and how do you mitigate information loss at merge?

LLM-to-small-model distillation pipeline: how do you build training data and measure distillation loss?