Why similarity measures find related text in NLP - Challenge Your Understanding

Challenge - 5 Problems

🧠 Conceptual

intermediate

Why do cosine similarity scores close to 1 indicate related text?

Cosine similarity measures the angle between two text vectors. Why does a score close to 1 mean the texts are related?

A Because the vectors point in very similar directions, showing similar word usage patterns.

B Because the vectors have very different lengths, indicating unrelated content.

C Because the vectors are orthogonal, meaning they share no common words.

D Because the vectors have zero magnitude, so similarity is undefined.

💡 Hint

Think about what it means when two arrows point the same way.

Cosine similarity measures the angle between vectors. When the angle is small, cosine similarity is close to 1, meaning the texts share similar word patterns.

❓ Predict Output

intermediate

Output of cosine similarity between two text vectors

What is the output of the following code that computes cosine similarity between two text vectors?

NLP

from sklearn.feature_extraction.text import CountVectorizer
from sklearn.metrics.pairwise import cosine_similarity

texts = ['apple orange banana', 'banana orange apple']
vectorizer = CountVectorizer()
X = vectorizer.fit_transform(texts)
sim = cosine_similarity(X[0], X[1])
print(round(sim[0][0], 2))

A 0.0

B 0.5

C 1.0

D 0.33

💡 Hint

Both texts have the same words but in different order.

Cosine similarity ignores word order and measures vector angle. Since both texts have the same words, similarity is 1.0.

❓ Model Choice

advanced

Best similarity measure for short text snippets

You want to find relatedness between very short texts like tweets. Which similarity measure is best?

A Jaccard similarity on sets of words

B Euclidean distance on raw word counts

C Manhattan distance on character counts

D Cosine similarity on TF-IDF vectors

💡 Hint

Consider a measure that accounts for word importance and ignores length differences.

Cosine similarity on TF-IDF vectors captures word importance and normalizes length, making it suitable for short texts.

❓ Hyperparameter

advanced

Effect of stopword removal on similarity scores

How does removing stopwords before vectorizing text affect similarity scores?

A It increases similarity scores by focusing on meaningful words.

B It decreases similarity scores by removing common words that link texts.

C It has no effect because stopwords are ignored by similarity measures.

D It causes errors because vectors become empty.

💡 Hint

Think about what words carry meaning in text.

Removing stopwords removes common words that add noise, so similarity focuses on meaningful words, often increasing scores for related texts.

🔧 Debug

expert

Why does this similarity code produce zero similarity for related texts?

Given two related texts, this code outputs zero similarity. What is the cause?

NLP

from sklearn.feature_extraction.text import CountVectorizer
from sklearn.metrics.pairwise import cosine_similarity

texts = ['cat and dog', 'dog and cat']
vectorizer = CountVectorizer()
X = vectorizer.fit_transform(texts)
sim = cosine_similarity(X[0], X[1])
print(sim[0][0])

A The vectors are sparse matrices and need to be converted to dense arrays before similarity.

B The code is correct and should output 1.0; zero means an environment error.

C The cosine_similarity function expects 1D arrays, but gets 2D sparse matrices causing zero output.

D The CountVectorizer default token pattern excludes all words, resulting in empty vectors.

💡 Hint

Check if the code runs as expected in a normal Python environment.

The code is correct and outputs 1.0 because the texts have the same words. Zero similarity would indicate an external issue.