Word2Vec (CBOW and Skip-gram) in NLP - Practice Problems & Coding Challenges

Challenge - 5 Problems

🧠 Conceptual

intermediate

Difference between CBOW and Skip-gram in Word2Vec

Which statement correctly describes the main difference between the CBOW and Skip-gram models in Word2Vec?

A CBOW predicts the target word from surrounding context words, while Skip-gram predicts surrounding context words from the target word.

B CBOW is used only for large datasets, while Skip-gram works only on small datasets.

C CBOW uses one-hot encoding for words, while Skip-gram uses word embeddings directly as input.

D CBOW predicts the next word in a sentence, while Skip-gram predicts the previous word.

💡 Hint

Think about which model uses context to predict the center word and which uses the center word to predict context.

CBOW takes the surrounding words as input and tries to predict the center word. Skip-gram does the opposite: it takes the center word as input and tries to predict the surrounding words.

❓ Predict Output

intermediate

Output of a simple Skip-gram training step

Given the following simplified Skip-gram training code snippet, what will be the shape of the output vector representing the predicted context word probabilities?

NLP

import numpy as np

vocab_size = 10
embedding_dim = 5

# Random embedding matrix
embeddings = np.random.rand(vocab_size, embedding_dim)

# One-hot encoded center word (index 3)
center_word = np.zeros(vocab_size)
center_word[3] = 1

# Compute hidden layer (embedding lookup)
hidden = embeddings.T @ center_word  # shape (embedding_dim,)

# Output weights
output_weights = np.random.rand(vocab_size, embedding_dim)

# Compute output layer
output = output_weights @ hidden  # shape ?

print(output.shape)

A (5,)

B (10,)

C (1, 10)

D (10, 5)

💡 Hint

Consider the matrix multiplication dimensions: output_weights (vocab_size x embedding_dim) times hidden (embedding_dim,).

The output_weights matrix has shape (10, 5) and hidden vector has shape (5,). Multiplying them results in a vector of shape (10,), representing scores for each word in the vocabulary.

❓ Model Choice

advanced

Choosing Word2Vec model for rare words

You want to train word embeddings on a small dataset with many rare words. Which Word2Vec model is generally better at learning good embeddings for rare words?

A CBOW, because it averages context and smooths rare word signals.

B Skip-gram, because it ignores rare words during training.

C Skip-gram, because it predicts context from the target word and better captures rare word representations.

D CBOW, because it uses hierarchical softmax which is faster for rare words.

💡 Hint

Think about which model focuses more on individual target words and their contexts.

Skip-gram predicts context words from the target word, which helps it learn better embeddings for rare words by focusing on their contexts individually. CBOW averages context which can dilute rare word signals.

❓ Hyperparameter

advanced

Effect of window size in Word2Vec training

In Word2Vec training, what is the effect of increasing the window size parameter?

A It controls the learning rate decay during training.

B It decreases the number of context words, focusing on very close neighbors only.

C It changes the embedding dimension size, making vectors longer.

D It increases the number of context words considered, capturing broader semantic relationships but may add noise.

💡 Hint

Window size defines how many words around the target word are used as context.

A larger window size means more surrounding words are included as context, which helps capture broader semantic relations but can introduce irrelevant words as noise.

❓ Metrics

expert

Evaluating Word2Vec embeddings with analogy task

After training Word2Vec embeddings, you want to evaluate them using the analogy task: "king is to queen as man is to ?". Which metric best measures the quality of the embeddings on this task?

A Cosine similarity between the vector (queen - king + man) and all other word vectors to find the closest match.

B Euclidean distance between the vector (king + queen) and (man + woman).

C Dot product between the embeddings of 'king' and 'queen' only.

D Mean squared error between predicted and true word indices.

💡 Hint

Think about how analogy tasks use vector arithmetic and similarity measures.

The analogy task uses vector arithmetic: vector('queen') - vector('king') + vector('man'). The best matching word vector is found by highest cosine similarity to this result, indicating semantic relationships.