Sentence-BERT Embeddings Quiz

Answers and explanations appear as you take the quiz.

1. What core problem with plain BERT does Sentence-BERT solve?
   • Comparing many sentences pairwise with BERT is far too slow
   • BERT cannot read more than one sentence
   • BERT only works in English
   • BERT cannot be fine-tuned
2. What network architecture does Sentence-BERT use?
   • A single autoregressive decoder
   • A siamese network with two weight-sharing BERT towers
   • A convolutional image network
   • A reinforcement-learning agent
3. Which pooling strategy did the original Sentence-BERT paper evaluate for creating sentence embeddings?
   • Max pooling over attention heads
   • Using only the first character
   • Mean pooling over the token embeddings
   • Random selection of one token
4. Once sentences are embedded, how is their similarity typically measured?
   • Running BERT again on both
   • Counting shared characters
   • Comparing file sizes
   • Cosine similarity (a dot product) between vectors
5. What training data did the original Sentence-BERT paper use for NLI fine-tuning?
   • Natural Language Inference examples labeled for entailment, contradiction and neutral relations
   • Only satellite images with land-cover labels
   • Only audio recordings with speaker labels
   • Only source code and bug reports
6. In a possible semantic-retrieval workflow, what can a Sentence-BERT bi-encoder do?
   • Cross-attend to every query-candidate pair before any encoding
   • Encode a query and candidate texts separately so vector similarity can rank candidates
   • Guarantee retrieval quality for every corpus
   • Replace the need to evaluate relevance
7. What does Matryoshka Representation Learning enable for an embedding?
   • It guarantees identical quality at every vector size
   • It converts embeddings into labeled text tokens
   • A single representation can be truncated to different dimensions to trade storage or compute for task performance
   • It removes the need for similarity comparisons
8. In a common hybrid retrieval pipeline, what happens after a bi-encoder retrieves candidate passages?
   • The candidate text is converted into audio
   • The bi-encoder is discarded and no ranking occurs
   • The system averages all documents into one sentence
   • A cross-encoder can score query-passage pairs to rerank the candidates