Handling imbalanced text data in NLP - Practice Problems & Coding Challenges

Challenge - 5 Problems

🧠 Conceptual

intermediate

Why use SMOTE for imbalanced text data?

You have a text classification task with very few examples of one class. Why might you use SMOTE (Synthetic Minority Over-sampling Technique) on the text features?

A To create synthetic examples of the minority class by interpolating feature vectors, helping balance the dataset.

B To remove noisy examples from the majority class to reduce imbalance.

C To convert text data into numerical vectors using TF-IDF.

D To randomly duplicate minority class examples without changing their features.

💡 Hint

Think about how SMOTE creates new data points rather than just copying existing ones.

SMOTE generates new synthetic samples by interpolating between existing minority class feature vectors, which helps the model learn better decision boundaries.

❓ Predict Output

intermediate

Output of class distribution after random oversampling

Given the following code that uses RandomOverSampler on text data features, what will be the printed class distribution?

NLP

from collections import Counter
from sklearn.feature_extraction.text import CountVectorizer
from imblearn.over_sampling import RandomOverSampler

texts = ['good', 'bad', 'good', 'bad', 'bad', 'good', 'good', 'bad', 'bad', 'bad']
labels = [1, 0, 1, 0, 0, 1, 1, 0, 0, 0]

vectorizer = CountVectorizer()
X = vectorizer.fit_transform(texts)

ros = RandomOverSampler(random_state=42)
X_res, y_res = ros.fit_resample(X, labels)

print(Counter(y_res))

A Counter({0: 6, 1: 6})

B Counter({0: 5, 1: 5})

C Counter({0: 7, 1: 3})

D Counter({0: 6, 1: 4})

💡 Hint

RandomOverSampler balances classes by duplicating minority class samples until counts match.

The original counts are 0:6 and 1:4. RandomOverSampler duplicates minority class (1) samples to match majority class (0) count of 6.

❓ Model Choice

advanced

Best model choice for imbalanced text classification

You have a highly imbalanced text dataset with 95% negative and 5% positive labels. Which model choice is best to handle this imbalance?

A A K-Nearest Neighbors model with k=3 and no class weighting.

B A simple neural network without any class weighting or sampling.

C A decision tree with default parameters and no imbalance handling.

D A logistic regression model with class_weight='balanced' parameter.

💡 Hint

Consider models that can adjust learning to pay more attention to minority class.

Logistic regression with class_weight='balanced' adjusts the loss to give more importance to minority class, helping with imbalance.

❓ Hyperparameter

advanced

Choosing the right threshold for imbalanced text classification

After training a binary text classifier on imbalanced data, you notice low recall for the minority class. Which hyperparameter adjustment can help improve recall?

A Increase the learning rate to speed up training.

B Lower the classification threshold below 0.5 to predict more positives.

C Increase the batch size to stabilize gradients.

D Use early stopping to prevent overfitting.

💡 Hint

Recall improves when the model predicts more positive cases, even if some are false positives.

Lowering the threshold makes the model label more samples as positive, increasing recall but possibly lowering precision.

❓ Metrics

expert

Choosing the best metric for imbalanced text data evaluation

You trained a text classifier on imbalanced data. Which metric is best to evaluate model performance focusing on minority class detection?

A Accuracy, because it shows overall correct predictions.

B Log Loss, because it measures probability calibration.

C F1-score, because it balances precision and recall for the minority class.

D Mean Squared Error, because it measures prediction error.

💡 Hint

Accuracy can be misleading when classes are imbalanced.

F1-score combines precision and recall, giving a better sense of minority class detection quality than accuracy.