Handling imbalanced text data in NLP

Introduction

Imbalanced text data means some categories have many examples, and others have very few. Handling this helps models learn better and make fair predictions.

When classifying emails as spam or not spam, but spam emails are much fewer.

When detecting rare events in customer reviews, like complaints versus normal feedback.

When sorting news articles into topics where some topics appear less often.

When building sentiment analysis models with many neutral texts but few positive or negative ones.

Syntax

NLP

from sklearn.utils import resample

# Upsample minority class
minority_upsampled = resample(minority_class_data, replace=True, n_samples=desired_count, random_state=42)

# Downsample majority class
majority_downsampled = resample(majority_class_data, replace=False, n_samples=desired_count, random_state=42)

Use resample to increase or decrease samples in classes.

Set random_state for reproducible results.

Examples

This makes the smaller class as big as the larger one by repeating samples.

NLP

from sklearn.utils import resample

# Upsample minority class to match majority
minority_upsampled = resample(minority_data, replace=True, n_samples=len(majority_data), random_state=1)

This reduces the larger class size by randomly picking fewer samples.

NLP

from sklearn.utils import resample

# Downsample majority class to match minority
majority_downsampled = resample(majority_data, replace=False, n_samples=len(minority_data), random_state=1)

SMOTE creates new synthetic samples for the minority class to balance data.

NLP

from imblearn.over_sampling import SMOTE

smote = SMOTE(random_state=42)
X_resampled, y_resampled = smote.fit_resample(X, y)

Sample Model

This example shows how to balance a small text dataset by repeating the minority class samples, then train a simple model and check its performance.

NLP

from sklearn.feature_extraction.text import CountVectorizer
from sklearn.utils import resample
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import classification_report

# Sample imbalanced text data
texts = ["good product", "bad product", "excellent", "poor quality", "nice", "terrible", "awesome", "bad", "great", "awful"]
labels = [1, 0, 1, 0, 1, 0, 1, 0, 1, 0]  # 1=positive, 0=negative

# Separate majority and minority classes
texts_pos = [t for t, l in zip(texts, labels) if l == 1]
texts_neg = [t for t, l in zip(texts, labels) if l == 0]

# Upsample minority class (positive) to match majority (negative)
texts_pos_upsampled = resample(texts_pos, replace=True, n_samples=len(texts_neg), random_state=42)
labels_pos_upsampled = [1] * len(texts_pos_upsampled)

# Combine balanced data
texts_balanced = texts_neg + texts_pos_upsampled
labels_balanced = [0] * len(texts_neg) + labels_pos_upsampled

# Vectorize text
vectorizer = CountVectorizer()
X = vectorizer.fit_transform(texts_balanced)
y = labels_balanced

# Train logistic regression
model = LogisticRegression(random_state=42)
model.fit(X, y)

# Predict on training data
predictions = model.predict(X)

# Print classification report
report = classification_report(y, predictions, zero_division=0)
print(report)

Important Notes

Upsampling can cause overfitting because it repeats data.

Downsampling may lose useful information by removing data.

Try synthetic methods like SMOTE for better balance without duplicates.

Summary

Imbalanced text data can hurt model learning and fairness.

Use upsampling, downsampling, or synthetic sampling to balance classes.

Check model performance carefully after balancing data.