Classical NLP & Text Classification

This lesson is about getting text into a classifier and knowing whether the result is any good. Tokenization internals (BPE merges, vocabulary size, token budgets) are Tokenization and Context Windows, prompting an LLM is Prompt Engineering, and how dense vectors are learned and compared is Embeddings & Representation Learning; here we use them as ingredients. The scoring vocabulary (precision, recall, F1, ROC and PR curves) is Model Evaluation Metrics, and the linear model itself is Logistic Regression & Linear Classifiers.

Text interviews follow a pattern: "classify 40 imbalanced classes", "evaluate an NER model", "would you use an LLM or a small model?" Each reduces to a few decisions: how text becomes numbers, which model family, how labels are obtained, how you measure, and what it costs to run. A TF-IDF plus linear model is still the right first answer in many products, because it trains in seconds, serves in microseconds, and gives a floor that any heavier model must beat on the same split. Every number marked "measured" below comes from scripts in /tmp/nlp/ run on a synthetic support-ticket corpus described in the worked example; they illustrate directions and mechanisms, not benchmarks, and anything not run is labelled as an assumption.