How to Do Multilabel Text Classification in NLP

To do multilabel text classification in NLP, prepare your text data with multiple labels per example, convert text to features (like TF-IDF or embeddings), and train a model that supports multilabel outputs such as LogisticRegression with OneVsRestClassifier or neural networks with sigmoid activation. Use appropriate loss functions like binary_crossentropy and evaluate with metrics like f1_score for multilabel tasks.

Syntax

Here is the typical syntax pattern for multilabel text classification using scikit-learn:

• OneVsRestClassifier: Wraps a base classifier to handle multilabel by training one classifier per label.
• fit(X_train, y_train): Trains the model on text features and multilabel targets.
• predict(X_test): Predicts multilabel outputs for new text data.
• MultiLabelBinarizer: Converts label lists to binary indicator format for multilabel.

python

from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.multiclass import OneVsRestClassifier
from sklearn.linear_model import LogisticRegression
from sklearn.preprocessing import MultiLabelBinarizer

texts = ["sample text one", "sample text two"]
labels = [["label1", "label2"], ["label2"]]

# Convert labels to binary format
mlb = MultiLabelBinarizer()
y = mlb.fit_transform(labels)

# Convert text to features
vectorizer = TfidfVectorizer()
X = vectorizer.fit_transform(texts)

# Create multilabel classifier
model = OneVsRestClassifier(LogisticRegression(max_iter=1000))

# Train model
model.fit(X, y)

# Predict
predictions = model.predict(X)

Example

This example shows how to train a multilabel text classifier on sample data, predict labels, and evaluate performance using F1 score.

python

from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.multiclass import OneVsRestClassifier
from sklearn.linear_model import LogisticRegression
from sklearn.preprocessing import MultiLabelBinarizer
from sklearn.metrics import f1_score
from sklearn.model_selection import train_test_split

# Sample data
texts = [
    "I love programming in Python",
    "Python and Java are popular languages",
    "I enjoy machine learning and AI",
    "AI is the future of technology",
    "Java is used in many applications"
]
labels = [
    ["programming", "python"],
    ["programming", "java"],
    ["machine learning", "ai"],
    ["ai"],
    ["programming", "java"]
]

# Convert labels to binary format
mlb = MultiLabelBinarizer()
y = mlb.fit_transform(labels)

# Split data
X_train, X_test, y_train, y_test = train_test_split(texts, y, test_size=0.4, random_state=42)

# Vectorize text
vectorizer = TfidfVectorizer()
X_train_vec = vectorizer.fit_transform(X_train)
X_test_vec = vectorizer.transform(X_test)

# Train multilabel classifier
model = OneVsRestClassifier(LogisticRegression(max_iter=1000))
model.fit(X_train_vec, y_train)

# Predict
y_pred = model.predict(X_test_vec)

# Evaluate
f1 = f1_score(y_test, y_pred, average='micro')

print("Predicted labels:", mlb.inverse_transform(y_pred))
print(f"Micro F1 score: {f1:.2f}")

Output

Predicted labels: [('ai',), ('programming', 'java')]
Micro F1 score: 1.00

Common Pitfalls

Common mistakes when doing multilabel text classification include:

• Using single-label classifiers without multilabel wrappers like OneVsRestClassifier.
• Not converting label lists to binary format with MultiLabelBinarizer.
• Using softmax activation or categorical cross-entropy loss which assume single-label classification.
• Ignoring multilabel evaluation metrics like micro/macro F1 score.

python

from sklearn.linear_model import LogisticRegression
from sklearn.preprocessing import LabelEncoder

# Wrong: Using LabelEncoder for multilabel (only works for single-label)
labels = [["python", "ai"], ["java"]]
le = LabelEncoder()
# y_wrong = le.fit_transform(labels)  # This will raise an error for multilabel lists

# Right: Use MultiLabelBinarizer for multilabel
from sklearn.preprocessing import MultiLabelBinarizer
labels_multi = [["python", "ai"], ["java"]]
mlb = MultiLabelBinarizer()
y_right = mlb.fit_transform(labels_multi)

Quick Reference

• Use MultiLabelBinarizer to convert multilabel lists to binary arrays.
• Use OneVsRestClassifier with a base classifier like LogisticRegression for multilabel tasks.
• Vectorize text with TfidfVectorizer or embeddings.
• Use sigmoid activation and binary cross-entropy loss in neural networks.
• Evaluate with multilabel metrics like micro/macro F1 score.

Key Takeaways

Convert multilabel targets to binary format using MultiLabelBinarizer before training.

Wrap classifiers with OneVsRestClassifier to handle multilabel classification.

Use text vectorization methods like TF-IDF to convert text into features.

Choose evaluation metrics suited for multilabel tasks, such as micro or macro F1 score.

Avoid single-label assumptions like softmax activation or LabelEncoder for multilabel data.