Dense Passage Retrieval Quiz

Answers and explanations appear as you take the quiz.

1. How does Dense Passage Retrieval decide which passages match a query?
   • Comparing dense vector representations of queries and passages by similarity
   • Counting how many query words appear in each document
   • Translating the query into SQL before searching
   • Ranking documents purely by how recently they were published
2. DPR uses a 'bi-encoder' architecture. What does that mean?
   • It encodes text twice and averages the results
   • It only works on bilingual documents
   • It encodes both characters and words simultaneously
   • It uses two separate encoders, one for queries and one for passages
3. Why can passage vectors be computed in advance in DPR?
   • Because the passages never change once written
   • Because the query encoder is disabled at search time
   • Because each passage is encoded independently of any query
   • Because passages are stored as raw text only
4. What training trick lets DPR generate many negative examples cheaply?
   • Random label flipping
   • In-batch negatives, using other passages in the same batch
   • Dropping out half the neurons
   • Manually labeling every wrong passage
5. For what vocabulary mismatch can learned DPR representations be useful compared with lexical matching alone?
   • The passage is unrelated but repeats the query's terms
   • Every query contains only an exact serial number
   • The encoder is guaranteed to retrieve the right answer for all inputs
   • A relevant passage expresses a query's idea using different words
6. What tool is commonly used to make DPR's nearest-neighbor search fast at scale?
   • An approximate nearest-neighbor index such as FAISS
   • A spreadsheet sort
   • A regular expression engine
   • A relational database join
7. Why can BM25-retrieved nonanswer passages be useful hard negatives when training DPR?
   • They guarantee that every retrieved passage is irrelevant
   • Their lexical overlap can make them challenging distractors that the retriever must distinguish from the answer passage
   • They remove the need for a positive answer passage
   • They force the passage encoder to count words without learning
8. After changing DPR's passage encoder, what should happen to a stored passage-vector index?
   • Keep the old embeddings because encoder weights cannot affect vectors
   • Rename the index without recalculating any representation
   • Recompute the passage embeddings with the intended compatible encoder and update the index
   • Disable the query encoder and search raw filenames instead