CourseAI Engineering: The Complete Course · Module 5: NLP for Applications · part 39 of 105
Part 39 · Module 5: NLP for Applications

Lesson 5: Scores that lie, and how to catch them

19 min read·13 Oct 2026
Module 5 · Lesson 5 · About 40 minutes · Needs: Module 3 Lesson 3 (precision and recall) and Module 5 Lesson 2 (sentence embeddings)

The situation

Priya has three months of real tickets, and the mix is not what the forty labelled tickets suggested: warranty questions are one in eleven, not one in four. Dev reruns the Lesson 3 classifier on a sample with that mix and reports 76 percent accuracy. Priya replies that the repairs team has not received a single ticket from it all week.

She also wants a one-line summary of each help-centre article, and asks how you will know the summaries are good. Both questions are about a score that looks fine and is not.

The rest of this course is yours to keep

This course is bought on its own, once, and stays readable afterwards, including the parts added to it later.