Sentence-BERT for embeddings in NLP - Model Metrics & Evaluation

Metrics & Evaluation - Sentence-BERT for embeddings

Which metric matters for Sentence-BERT embeddings and WHY

Sentence-BERT creates vector representations (embeddings) of sentences. To check how good these embeddings are, we often use cosine similarity. This measures how close two sentence vectors are, showing if they mean similar things. For tasks like sentence similarity or clustering, cosine similarity helps us see if the model groups related sentences well.

When Sentence-BERT is used for classification or retrieval, metrics like accuracy, precision, and recall become important. These tell us how well the embeddings help the model find or classify the right sentences.

Confusion matrix example for Sentence-BERT in classification

      Actual \ Predicted | Positive | Negative
      ---------------------------------------
      Positive           |   TP=85  |  FN=15
      Negative           |   FP=10  |  TN=90
    

Here, TP means sentences correctly matched as similar, FP means sentences wrongly matched, FN means missed similar sentences, and TN means correctly identified as not similar.

Precision vs Recall tradeoff with Sentence-BERT

Imagine you use Sentence-BERT to find similar customer questions in a help center.

• High precision: Most found questions are truly similar. Good if you want to avoid showing unrelated answers.
• High recall: You find almost all similar questions, even if some are less related. Good if you want to make sure no relevant question is missed.

Choosing between precision and recall depends on your goal. For example, if showing wrong answers is bad, prioritize precision. If missing any related question is bad, prioritize recall.

What good vs bad metric values look like for Sentence-BERT embeddings

• Good: Precision and recall above 0.8, F1 score above 0.8, cosine similarity scores clearly separate similar and dissimilar pairs.
• Bad: Precision or recall below 0.5, F1 score low, cosine similarity scores overlap a lot between similar and dissimilar sentences, making it hard to tell them apart.

Common pitfalls when evaluating Sentence-BERT embeddings

• Accuracy paradox: High accuracy can be misleading if classes are imbalanced (e.g., many dissimilar pairs).
• Data leakage: Using test sentences seen during training inflates metrics falsely.
• Overfitting: Embeddings work well on training data but poorly on new sentences.
• Ignoring threshold tuning: Cosine similarity needs a good cutoff to decide similarity; wrong thresholds hurt precision and recall.

Self-check question

Your Sentence-BERT model finds similar sentences with 98% accuracy but only 12% recall on similar pairs. Is it good for production? Why or why not?

Answer: No, it is not good. The low recall means the model misses most similar sentences, even if it is usually correct when it does find one. For tasks needing to find all similar sentences, missing many is a big problem.

Key Result

Cosine similarity is key for Sentence-BERT embeddings; precision and recall show how well similar sentences are found.

Practice

1. What is the main purpose of Sentence-BERT embeddings in NLP?

easy

A. To count the number of words in a sentence

B. To translate sentences into different languages

C. To generate random sentences for data augmentation

D. To convert sentences into numbers that capture their meaning

Solution

1. Step 1: Understand Sentence-BERT's role

   Sentence-BERT creates embeddings, which are numbers representing sentence meaning.

2. Step 2: Compare options with Sentence-BERT's function

   Only To convert sentences into numbers that capture their meaning describes converting sentences into meaningful numbers, matching Sentence-BERT's purpose.

3. Final Answer:

   To convert sentences into numbers that capture their meaning -> Option D

4. Quick Check:

   Sentence-BERT embeddings = meaningful numbers [OK]

Hint: Remember: embeddings = numbers capturing meaning [OK]

Common Mistakes:

• Confusing embeddings with translation
• Thinking embeddings count words
• Assuming embeddings generate sentences

2. Which Python code snippet correctly loads a pre-trained Sentence-BERT model using the sentence-transformers library?

easy

A. from sentence_transformers import SentenceTransformer
model = SentenceTransformer('all-MiniLM-L6-v2')

B. import sentence_transformers
model = sentence_transformers.load('all-MiniLM-L6-v2')

C. from transformers import SentenceBert
model = SentenceBert.load('all-MiniLM-L6-v2')

D. import sbert
model = sbert.SentenceTransformer('all-MiniLM-L6-v2')

Solution

1. Step 1: Recall correct import and model loading syntax

   The sentence-transformers library uses 'from sentence_transformers import SentenceTransformer' and then creates a model instance with the model name.

2. Step 2: Check each option for correctness

   from sentence_transformers import SentenceTransformer model = SentenceTransformer('all-MiniLM-L6-v2') matches the correct syntax. Options A, B, and D use incorrect imports or methods.

3. Final Answer:

   from sentence_transformers import SentenceTransformer model = SentenceTransformer('all-MiniLM-L6-v2') -> Option A

4. Quick Check:

   Correct import and model load = from sentence_transformers import SentenceTransformer model = SentenceTransformer('all-MiniLM-L6-v2') [OK]

Hint: Use 'from sentence_transformers import SentenceTransformer' [OK]

Common Mistakes:

• Using wrong import statements
• Calling non-existent load methods
• Confusing transformers library with sentence-transformers

3. Given the code below, what is the output shape of embeddings?

from sentence_transformers import SentenceTransformer
model = SentenceTransformer('all-MiniLM-L6-v2')
sentences = ['Hello world', 'How are you?']
embeddings = model.encode(sentences)
print(embeddings.shape)

medium

A. (384, 2)

B. (2, 384)

C. (2, 768)

D. (1, 384)

Solution

1. Step 1: Understand input and output of model.encode()

   Input is 2 sentences, so output embeddings will have 2 rows, one per sentence.

2. Step 2: Know embedding dimension of 'all-MiniLM-L6-v2'

   This model produces embeddings of size 384 per sentence.

3. Final Answer:

   (2, 384) -> Option B

4. Quick Check:

   2 sentences x 384 dims = (2, 384) [OK]

Hint: Output shape = (number of sentences, embedding size) [OK]

Common Mistakes:

• Swapping dimensions in output shape
• Assuming embedding size is 768
• Forgetting batch size dimension

4. You run this code but get an error: AttributeError: module 'sentence_transformers' has no attribute 'load'. What is the likely cause?

import sentence_transformers
model = sentence_transformers.load('all-MiniLM-L6-v2')

medium

A. The model file is missing from local directory

B. The model name 'all-MiniLM-L6-v2' is incorrect

C. The sentence_transformers module does not have a 'load' function

D. You need to import SentenceTransformer class explicitly

Solution

1. Step 1: Analyze the error message

   The error says 'sentence_transformers' has no attribute 'load', meaning 'load' is not a valid function in this module.

2. Step 2: Understand correct usage

   The correct way is to import SentenceTransformer class and instantiate it with the model name, not use 'load'.

3. Final Answer:

   The sentence_transformers module does not have a 'load' function -> Option C

4. Quick Check:

   AttributeError means wrong function call [OK]

Hint: Use SentenceTransformer(), not load() [OK]

Common Mistakes:

• Calling non-existent 'load' method
• Not importing SentenceTransformer class
• Assuming model loads from local file by default

5. You want to find the most similar sentence to 'I love machine learning' from a list using Sentence-BERT embeddings. Which approach is best?

hard

A. Encode all sentences and query, then find the sentence with highest cosine similarity to the query embedding

B. Count common words between query and each sentence, pick the highest count

C. Use a pre-trained translation model to translate sentences before comparison

D. Encode only the query sentence and compare it to raw text sentences

Solution

1. Step 1: Understand how Sentence-BERT embeddings are used for similarity

   Sentence-BERT embeddings represent sentence meaning as vectors; similarity is measured by cosine similarity between vectors.

2. Step 2: Evaluate options for similarity search

   Encode all sentences and query, then find the sentence with highest cosine similarity to the query embedding correctly encodes all sentences and compares embeddings using cosine similarity. Other options do not use embeddings properly or rely on less effective methods.

3. Final Answer:

   Encode all sentences and query, then find the sentence with highest cosine similarity to the query embedding -> Option A

4. Quick Check:

   Embedding + cosine similarity = best similarity search [OK]

Hint: Compare embeddings with cosine similarity for best match [OK]

Common Mistakes:

• Comparing raw text instead of embeddings
• Using word count instead of semantic similarity
• Encoding only query, not all sentences