Quiz: Classic NLP

Let's find out what's stuck. Ten questions, on the classic tools of NLP. Pause on each one, and commit to an answer. Then see how you did.

• Ten questions
• Commit before you hear the answer
• Keep score

Question 1

Question 1. What's the best tool for finding order numbers with a fixed format? A, A large language model. B, A regular expression. C, Topic modelling. D, Word2Vec. Pause the video and choose. The answer is B. Fixed patterns are cheap and exact with regex.

Question 2

Question 2. Why can removing stop words hurt sentiment analysis? A, It adds noise. B, It removes words like not. C, It slows training. D, It breaks tokenisers. Pause the video and choose. The answer is B. Not and never can reverse the meaning.

Question 3

Question 3. What does TF-IDF reward? A, Long documents. B, Words that are common here, but rare elsewhere. C, Very common words. D, Punctuation. Pause the video and choose. The answer is B. That's the mix that carries meaning.

Question 4

Question 4. Lemmatising differs from stemming because it does what? A, Uses a dictionary for the real base form. B, Removes all vowels. C, Only works on names. D, Translates words. Pause the video and choose. The answer is A. Studies becomes study, not studi.

Question 5

Question 5. What does cosine similarity measure? A, Document length. B, The angle between two vectors. C, Spelling. D, File size. Pause the video and choose. The answer is B. A smaller angle means more similar.

Question 6

Question 6. What does BIO tagging help with? A, Marking multi-word entities. B, Translation. C, Compressing text. D, Spell checking. Pause the video and choose. The answer is A. Begin, inside, and outside turn spans into token labels.

Question 7

Question 7. Two labellers agree seventy five percent of the time. What should you expect from a model? A, 99 percent is easy. B, Around that level is a realistic limit. C, It doesn't matter. D, Zero. Pause the video and choose. The answer is B. Models can't reliably beat the agreement of the humans who labelled the data.

Question 8

Question 8. What did Word2Vec learn from? A, Dictionaries. B, Which words appear near each other. C, Grammar rules. D, Human definitions. Pause the video and choose. The answer is B. You shall know a word by the company it keeps.

Question 9

Question 9. What's the classic baseline for text classification? A, TF-IDF with logistic regression. B, A hand-built transformer. C, Random guessing only. D, Stop word lists. Pause the video and choose. The answer is A. It's cheap, fast, readable, and hard to beat.

Question 10

Question 10. Which is a sign that accuracy is hiding a problem? A, Macro average far below weighted average. B, Fast training. C, Few classes. D, Short texts. Pause the video and choose. The answer is A. Rare classes are being missed.

Summary