Why Cosine similarity in NLP? - Purpose & Use Cases

The Big Idea

Discover how a simple angle can unlock the secrets of text similarity!

The Scenario

Imagine you have a huge pile of documents and you want to find which ones talk about similar topics. Doing this by reading each document and comparing them word by word would take forever!

The Problem

Manually checking similarity is slow and tiring. It's easy to miss connections or make mistakes because human brains can't quickly compare thousands of documents or long texts accurately.

The Solution

Cosine similarity turns texts into numbers and measures the angle between them. This way, it quickly tells how close two texts are in meaning without reading every word.

Before vs After

✗ Before

for doc1 in docs:
  for doc2 in docs:
    compare_words(doc1, doc2)

✓ After

similarity = cosine_similarity(vector1, vector2)

What It Enables

It lets machines quickly find how alike two pieces of text are, enabling smart search, recommendations, and understanding.

Real Life Example

When you search for a product online, cosine similarity helps find items with descriptions similar to your query, even if the exact words differ.

Key Takeaways

Manual text comparison is slow and error-prone.

Cosine similarity measures text closeness using math, not reading.

This speeds up tasks like search and recommendation.

Practice

1. What does cosine similarity measure between two vectors?

easy

A. The difference in vector lengths

B. How close the vectors point in the same direction

C. The sum of vector elements

D. The distance between vector origins

Solution

1. Step 1: Understand vector comparison

   Cosine similarity compares the angle between two vectors, not their length or sum.

2. Step 2: Interpret cosine similarity meaning

   A value close to 1 means vectors point in the same direction, showing similarity.

3. Final Answer:

   How close the vectors point in the same direction -> Option B

4. Quick Check:

   Cosine similarity = direction closeness [OK]

Hint: Cosine similarity checks angle, not length or sum [OK]

Common Mistakes:

• Confusing cosine similarity with Euclidean distance
• Thinking it measures vector length difference
• Assuming it sums vector values

2. Which of the following is the correct formula for cosine similarity between vectors A and B?

easy

A. \( \frac{\|A\|}{\|B\|} \)

B. \( \|A - B\| \)

C. \( \frac{A \cdot B}{\|A\| \times \|B\|} \)

D. \( A + B \)

Solution

1. Step 1: Recall cosine similarity formula

   Cosine similarity is the dot product of vectors divided by the product of their lengths.

2. Step 2: Match formula to options

   \( \frac{A \cdot B}{\|A\| \times \|B\|} \) matches the formula \( \frac{A \cdot B}{\|A\| \times \|B\|} \), others do not.

3. Final Answer:

   \( \frac{A \cdot B}{\|A\| \times \|B\|} \) -> Option C

4. Quick Check:

   Cosine similarity = dot product / product of norms [OK]

Hint: Look for dot product over product of lengths [OK]

Common Mistakes:

• Choosing Euclidean distance formula
• Adding vectors instead of dot product
• Dividing norms instead of multiplying

3. Given vectors A = [1, 2, 3] and B = [4, 5, 6], what is the cosine similarity (rounded to 2 decimals)?

medium

A. 0.97

B. 0.83

C. 0.74

D. 0.56

Solution

1. Step 1: Calculate dot product of A and B

   Dot product = 1*4 + 2*5 + 3*6 = 4 + 10 + 18 = 32

2. Step 2: Calculate norms of A and B

   Norm A = sqrt(1^2 + 2^2 + 3^2) = sqrt(14) ≈ 3.74; Norm B = sqrt(4^2 + 5^2 + 6^2) = sqrt(77) ≈ 8.77

3. Step 3: Compute cosine similarity

   Cosine similarity = 32 / (3.74 * 8.77) ≈ 32 / 32.83 ≈ 0.9749 rounded to 0.97

4. Step 4: Check closest option

   0.97 matches the value rounded to 2 decimals.

5. Final Answer:

   0.97 -> Option A

6. Quick Check:

   Dot product / (norms product) ≈ 0.97 [OK]

Hint: Calculate dot product and divide by product of lengths [OK]

Common Mistakes:

• Forgetting to take vector norms
• Mixing up dot product with element-wise multiplication
• Rounding too early causing wrong answer

4. What is wrong with this Python code to compute cosine similarity?

import numpy as np

def cosine_sim(a, b):
    return np.dot(a, b) / np.linalg.norm(a + b)

A = np.array([1, 0])
B = np.array([0, 1])
print(cosine_sim(A, B))

medium

A. It should add vectors before dot product

B. It uses np.dot instead of np.cross

C. It misses normalizing vectors before dot product

D. It divides by norm of sum instead of product of norms

Solution

1. Step 1: Analyze denominator in code

   The code divides by norm of (a + b), but cosine similarity requires product of norms of a and b.

2. Step 2: Understand correct formula

   Correct denominator is np.linalg.norm(a) * np.linalg.norm(b), not norm of sum.

3. Final Answer:

   It divides by norm of sum instead of product of norms -> Option D

4. Quick Check:

   Denominator must be product of norms [OK]

Hint: Denominator is product of norms, not norm of sum [OK]

Common Mistakes:

• Using norm of sum instead of product
• Confusing dot product with cross product
• Normalizing vectors before dot product unnecessarily

5. You have two text documents represented as TF-IDF vectors: doc1 = [0, 1, 2, 0] and doc2 = [1, 0, 1, 1]. Which step is best to improve cosine similarity comparison for very sparse vectors?

hard

A. Normalize vectors to unit length before computing cosine similarity

B. Add the vectors element-wise before similarity

C. Use Euclidean distance instead of cosine similarity

D. Ignore zero elements in vectors

Solution

1. Step 1: Understand sparse vector challenges

   Sparse vectors have many zeros; normalizing to unit length ensures fair angle comparison.

2. Step 2: Identify best practice for cosine similarity

   Normalizing vectors before cosine similarity avoids bias from vector length differences.

3. Final Answer:

   Normalize vectors to unit length before computing cosine similarity -> Option A

4. Quick Check:

   Normalization improves cosine similarity on sparse data [OK]

Hint: Always normalize vectors before cosine similarity [OK]

Common Mistakes:

• Adding vectors instead of comparing
• Switching to Euclidean distance without reason
• Ignoring zeros instead of normalizing