Sentence-BERT for embeddings in NLP - Cheat Sheet & Quick Revision

Recall & Review

beginner

What is Sentence-BERT (SBERT)?

Sentence-BERT is a model that creates meaningful sentence embeddings by fine-tuning BERT with a siamese network structure, enabling efficient comparison of sentence meanings.

beginner

Why is Sentence-BERT better than vanilla BERT for sentence similarity tasks?

SBERT produces fixed-size sentence embeddings that can be compared quickly using cosine similarity, while vanilla BERT requires expensive pairwise token-level comparisons.

intermediate

How does SBERT generate embeddings for two sentences?

SBERT uses a siamese network to encode each sentence separately into vectors, then compares these vectors using simple distance metrics like cosine similarity.

beginner

What is a common use case for Sentence-BERT embeddings?

SBERT embeddings are used for tasks like semantic search, clustering similar sentences, and paraphrase detection by comparing sentence meanings efficiently.

beginner

What metric is typically used to compare SBERT embeddings?

Cosine similarity is commonly used to measure how close two SBERT sentence embeddings are in meaning.

What does Sentence-BERT primarily produce?

A Named entity labels

B Word-level token embeddings

C Fixed-size sentence embeddings

D Part-of-speech tags

Sentence-BERT creates fixed-size vectors representing whole sentences for easy comparison.

Which architecture does SBERT use to encode sentence pairs?

A Siamese network

B Recurrent neural network

C Convolutional neural network

D Transformer decoder only

SBERT uses a siamese network to encode sentences separately but in parallel.

Why is cosine similarity used with SBERT embeddings?

A It counts the number of matching words

B It measures the angle between vectors, showing semantic similarity

C It measures Euclidean distance only

D It normalizes sentence length

Cosine similarity measures how close two vectors point in the same direction, indicating similar meaning.

Which task is NOT a typical use case for SBERT embeddings?

A Image classification

B Semantic search

C Sentence clustering

D Paraphrase detection

SBERT embeddings are for text tasks, not image classification.

What problem does SBERT solve compared to vanilla BERT for sentence similarity?

A Inability to process sentences

B Lack of word embeddings

C Poor spelling correction

D High computational cost of pairwise token comparisons

SBERT reduces computation by encoding sentences separately, avoiding costly token-level comparisons.

Explain how Sentence-BERT creates embeddings and why this is useful for comparing sentence meanings.

Think about how SBERT avoids comparing tokens directly.

Describe common applications where Sentence-BERT embeddings improve performance.

Consider tasks that need quick understanding of sentence meaning.