Text Classification

Text classification is the oldest useful thing done with language data and it is still everywhere.

The shape never changes: a document arrives, and one label from a fixed list must come out.

Spam filtering, support ticket routing, content moderation and intent detection are all instances of it.

The classical pipeline preprocesses the text, converts it to TF-IDF vectors, and fits a classifier.

Naive Bayes and logistic regression were the workhorses, and linear models remain remarkably competitive.

That approach is fast, interpretable, needs no GPU, and gives you a strong baseline in an afternoon.

Its weakness is that it never sees meaning: a review saying not good looks positive to it.

Fine-tuning a model such as BERT fixed that by classifying from contextual representations instead.

It reads the whole sentence at once, so negation and word order finally count.

Prompting a large model removes the training step entirely and works with no labelled data at all.

The trade is cost and latency per document, which matters when you classify millions a day.

Whatever the model, the evaluation trap is the same one: class imbalance.

Real label distributions are lopsided, and a classifier that always predicts the common class can look excellent.

Precision and recall on the rare class, and the confusion matrix behind them, are what tell the truth.

Which of the two you optimise is a product decision: a missed fraud costs more than a false alarm.