NLP with Python: Transformers, LLMs, Text Classification

About this course

This course contains the use of artificial intelligence.

Most NLP courses show you a method, print a number, and move on. You are never told what that number would have been if the model had learned nothing at all. This course publishes that floor next to every result, and the data is synthetic and seeded so you can reproduce each figure on your own machine to the digit.

Here is what that buys. A classifier on this corpus scores 69.83 per cent accuracy across seven categories. On its own that sounds respectable. The majority-class floor is 25.60 per cent, so the model earned about forty-four points, not sixty-nine. Break it out per class and precision holds steady while recall collapses from 0.7474 to 0.5368 - the model is not confused about rare categories, it is reluctant to commit to them. None of that is visible in the headline, and the headline is what most courses report.

Text preparation as a measurable job. The shipped corpus arrives dirty: HTML, doubled encoding, signature blocks, quoted replies, duplicate clusters and empty bodies. A correct deduplication removes 398 rows, not 412, because the first of each cluster is a real ticket - and getting that off by the cluster count is the commonest error in a cleaning pipeline.

Representation, derived rather than imported. Bag of words and TF-IDF including the exact formula scikit-learn uses, and why a third of this vocabulary is noise that inverse document frequency rewards hardest. Word embeddings as the by-product of a prediction task. Contextual embeddings and the single change that matters - one vector per occurrence rather than one per word type.

Then the methods, each scored. Classification with imbalance taken apart rather than mentioned. Topic modelling measured against known truth, where two planted topics share eighteen of their twenty-two words and no algorithm, seed or component count will separate them.

Read the full description on Udemy on Udemy (opens in a new tab)

Prices, discounts and availability are set by Udemy. We may earn a commission when you purchase through links on this site.