Word similarity and analogies in NLP - Practice Problems & Coding Challenges

Challenge - 5 Problems

🧠 Conceptual

intermediate

Understanding cosine similarity in word embeddings

Which statement best describes what cosine similarity measures between two word vectors?

A The difference in magnitude between the two vectors, ignoring their direction

B The Euclidean distance between the two vectors, showing how far apart they are in space

C The sum of the element-wise multiplication of the two vectors without normalization

D The angle between the two vectors, indicating how similar their directions are regardless of length

💡 Hint

Think about how cosine similarity focuses on direction rather than length.

Cosine similarity measures the cosine of the angle between two vectors. It shows how similar their directions are, ignoring their lengths. This is useful in word embeddings because words with similar meanings tend to point in similar directions.

❓ Predict Output

intermediate

Output of word analogy vector operation

Given the following word vectors, what is the result of the operation: vector('king') - vector('man') + vector('woman')? Vectors: king = [0.8, 0.6] man = [0.7, 0.5] woman = [0.6, 0.7]

NLP

import numpy as np

king = np.array([0.8, 0.6])
man = np.array([0.7, 0.5])
woman = np.array([0.6, 0.7])

result = king - man + woman
print(result)

A [0.7 0.8]

B [0.9 0.8]

C [0.6 0.7]

D [0.7 0.6]

💡 Hint

Perform the vector addition and subtraction step by step.

Subtracting 'man' from 'king' and then adding 'woman' results in [0.8-0.7+0.6, 0.6-0.5+0.7] = [0.7, 0.8].

❓ Model Choice

advanced

Choosing the best model for capturing word analogies

Which word embedding model is best known for capturing linear relationships that allow solving analogies like 'king' - 'man' + 'woman' = 'queen'?

A Bag-of-Words model

B Word2Vec Skip-gram model

C TF-IDF vectorizer

D One-hot encoding

💡 Hint

Think about models that learn dense vector representations with semantic relationships.

Word2Vec Skip-gram learns dense vectors where semantic relationships are encoded as linear offsets, enabling analogy tasks. Bag-of-Words and TF-IDF do not capture such relationships, and one-hot encoding is sparse and does not encode meaning.

❓ Metrics

advanced

Evaluating word similarity with Spearman correlation

You have a list of human-rated word similarity scores and cosine similarity scores from your model. Which metric best measures how well your model's similarity rankings match human judgments?

A Spearman rank correlation coefficient

B Mean squared error

C Accuracy

D Precision

💡 Hint

Consider metrics that compare rankings rather than exact values.

Spearman rank correlation measures how well the order of similarity scores from the model matches the order of human judgments, which is ideal for similarity tasks. Mean squared error measures exact value differences, which is less meaningful here.

🔧 Debug

expert

Debugging incorrect analogy results

You run this code to find the word closest to the vector result of 'king' - 'man' + 'woman' but get an unrelated word. What is the most likely cause? Code snippet: import numpy as np result = embeddings['king'] - embeddings['man'] + embeddings['woman'] closest_word = None max_sim = -1 for word, vec in embeddings.items(): sim = np.dot(result, vec) / (np.linalg.norm(result) * np.linalg.norm(vec)) if sim > max_sim: max_sim = sim closest_word = word print(closest_word)

A The embeddings dictionary includes the words 'king', 'man', and 'woman', so no issue there

B The embeddings vectors are not normalized before the operation, causing invalid results

C The code does not exclude the input words from the search, so it may return 'king' or 'woman' instead of the correct analogy

D The cosine similarity calculation is incorrect because it uses dot product without normalization

💡 Hint

Think about what happens if the search includes the original words used in the vector operation.

The code searches over all words including 'king', 'man', and 'woman'. Since the result vector is close to these, it may return one of them instead of the intended analogy word. Excluding input words from the search fixes this.