Choosing number of topics in NLP - Practice Problems & Coding Challenges

Challenge - 5 Problems

🧠 Conceptual

intermediate

Why is choosing the number of topics important in topic modeling?

In topic modeling, selecting the number of topics affects the model's usefulness. Why is this choice important?

A Because too few topics may merge distinct themes, and too many topics may split themes unnecessarily.

B Because the number of topics determines the size of the input data.

C Because the number of topics controls the speed of the training algorithm only.

D Because the number of topics decides the number of words in the vocabulary.

💡 Hint

Think about how topics represent themes in the data.

Choosing the right number of topics balances between merging distinct themes and splitting meaningful themes, affecting interpretability and usefulness.

❓ Metrics

intermediate

Which metric helps decide the optimal number of topics?

When training topic models, which metric is commonly used to evaluate and choose the best number of topics?

A Mean Squared Error, which measures prediction error.

B Accuracy, which measures correct topic labels.

C F1 Score, which balances precision and recall.

D Perplexity, which measures how well the model predicts unseen data.

💡 Hint

Think about a metric that measures prediction quality on new data.

Perplexity measures how well the model predicts new data, helping to find a number of topics that generalizes well.

❓ Predict Output

advanced

Output of perplexity calculation for different topic numbers

Given the code below that computes perplexity for different numbers of topics, what is the output?

NLP

from sklearn.decomposition import LatentDirichletAllocation
from sklearn.feature_extraction.text import CountVectorizer

texts = ["apple banana fruit", "banana orange fruit", "car truck vehicle", "truck bus vehicle"]
vectorizer = CountVectorizer()
X = vectorizer.fit_transform(texts)

perplexities = {}
for n_topics in [2, 3]:
    lda = LatentDirichletAllocation(n_components=n_topics, random_state=0)
    lda.fit(X)
    perplexities[n_topics] = lda.perplexity(X)

print(perplexities)

A {2: 9.0, 3: 9.0}

B {2: 9.0, 3: 7.5}

C {2: 7.5, 3: 9.0}

D {2: 7.5, 3: 7.5}

💡 Hint

Lower perplexity means better model fit.

With 3 topics, the model fits better, so perplexity is lower than with 2 topics.

❓ Model Choice

advanced

Choosing number of topics using coherence score

You want to select the number of topics for an LDA model using coherence score. Which approach is best?

A Select the topic count randomly and trust the model to adjust.

B Pick the topic count with lowest perplexity without checking coherence.

C Train models with different topic counts and pick the one with highest coherence score.

D Choose the topic count based on the largest vocabulary size.

💡 Hint

Coherence measures how interpretable topics are.

Training multiple models and selecting the one with highest coherence ensures topics are meaningful and interpretable.

🔧 Debug

expert

Why does increasing topics beyond a point worsen model quality?

After increasing the number of topics in your LDA model beyond 10, you notice coherence scores drop and topics become less meaningful. What is the most likely cause?

A The model is overfitting by splitting coherent topics into smaller, less meaningful ones.

B The model is underfitting because it has too few topics to capture data complexity.

C The vocabulary size is too small to support more topics.

D The training data is too large, causing the model to fail.

💡 Hint

Think about what happens when a model tries to create too many topics.

Too many topics cause the model to split meaningful themes into smaller, less coherent topics, reducing quality.