Document similarity ranking in NLP - Practice Problems & Coding Challenges

Challenge - 5 Problems

🧠 Conceptual

intermediate

What does cosine similarity measure in document ranking?

Imagine you have two documents represented as vectors. What does cosine similarity tell you about these documents?

A The angle between the two document vectors, indicating how similar their content is.

B The total number of common words between the two documents.

C The difference in length between the two documents measured by word count.

D The sum of the frequencies of all words in both documents.

💡 Hint

Think about how vectors relate to each other in space.

Cosine similarity measures the cosine of the angle between two vectors. A smaller angle means the documents are more similar in content.

❓ Predict Output

intermediate

Output of cosine similarity calculation

What is the output of this Python code that calculates cosine similarity between two document vectors?

NLP

import numpy as np

doc1 = np.array([1, 2, 3])
doc2 = np.array([4, 5, 6])

cos_sim = np.dot(doc1, doc2) / (np.linalg.norm(doc1) * np.linalg.norm(doc2))
print(round(cos_sim, 2))

A 0.75

B 1.00

C 0.97

D 0.87

💡 Hint

Calculate dot product and norms carefully.

The cosine similarity between [1,2,3] and [4,5,6] is approximately 0.97 after rounding.

❓ Model Choice

advanced

Best model for semantic document similarity

You want to rank documents by meaning, not just word overlap. Which model is best for this?

A Bag-of-words model with Euclidean distance

B Count vectorizer with Jaccard similarity

C TF-IDF vectorizer with cosine similarity

D Pretrained transformer embeddings with cosine similarity

💡 Hint

Consider models that understand context and meaning.

Pretrained transformer embeddings capture semantic meaning better than simple count or TF-IDF models.

❓ Hyperparameter

advanced

Effect of embedding dimension on similarity ranking

How does increasing the embedding vector size affect document similarity ranking?

A It can improve accuracy but may cause overfitting or slow computation.

B It reduces accuracy because larger vectors are harder to compare.

C It has no effect on similarity ranking performance.

D It always improves ranking accuracy by capturing more details.

💡 Hint

Think about trade-offs between detail and complexity.

Larger embeddings can capture more information but may overfit or slow down similarity calculations.

❓ Metrics

expert

Choosing the best metric for ranking evaluation

You have a list of documents ranked by similarity to a query. Which metric best measures how well the ranking matches user relevance?

A Mean Squared Error (MSE)

B Precision at K (P@K)

C Root Mean Squared Logarithmic Error (RMSLE)

D Confusion Matrix

💡 Hint

Consider metrics that evaluate ranked lists and relevance.

Precision at K measures how many of the top K ranked documents are relevant, which fits ranking evaluation.