Review: Modern NLP

Now the modern toolkit. This round covers embeddings, retrieval, and working with language models.

• The modern toolkit
• Embeddings and retrieval
• Language models

Recap: modern tools

Embeddings place meaning in a space of numbers. Attention lets each word look at the others. RAG retrieves first, then generates. Chunking decides what can be found. Hybrid search combines meaning and keywords. Rerankers read the question and document together. And citations plus refusal make answers trustworthy.

Question 1

Question 1. Clustering tickets scored 0.05 with word counts and 1.0 with embeddings. Why? A, More data. B, Short texts share few words, but similar meanings. C, A better algorithm. D, Random luck. Pause and choose. The answer is B. The algorithm was the same.

Question 2

Question 2. A customer asks, can I change it? What do you do before searching? A, Search as written. B, Rewrite it to stand alone. C, Refuse. D, Translate it. Pause and choose. The answer is B. Follow-ups depend on context.

Question 3

Question 3. Why rerank after vector search? A, To save disk. B, A reranker reads query and document together. C, To add noise. D, To remove citations. Pause and choose. The answer is B. Embeddings are compared separately.

Question 4

Question 4. What's the best test of a RAG system? A, A question with no answer in the sources. B, Its speed. C, Its colour. D, A greeting. Pause and choose. The answer is A. Does it refuse, or invent?

Question 5

Question 5. Why can't BLEU judge a single sentence well? A, It is slow. B, Same meaning in different words can score zero. C, It needs a GPU. D, It only works on French. Pause and choose. The answer is B. It counts matching phrases.

Question 6

Question 6. How do you stop a bot guessing? A, Higher temperature. B, A similarity threshold, then hand over. C, Longer prompts. D, More emojis. Pause and choose. The answer is B. I don't know is a good answer.

Summary