Text Classification

Text classification is the task of assigning a category or label to a piece of text. It is one of the most widely deployed AI applications in business, powering everything from email spam filters to customer service routing to content moderation.

Common examples

• Spam detection: Classifying emails as spam or legitimate
• Sentiment analysis: Classifying reviews as positive, negative, or neutral
• Topic tagging: Classifying news articles by topic (politics, sports, technology)
• Support ticket routing: Classifying customer queries by department (billing, technical, sales)
• Content moderation: Classifying user-generated content as safe, flagged, or prohibited
• Intent detection: Classifying chatbot messages by user intent (question, complaint, request)

Traditional approach

Before LLMs, text classification required:

1. Collecting thousands of labelled examples
2. Converting text into numerical features (bag of words, TF-IDF, embeddings)
3. Training a classification model (Naive Bayes, SVM, or neural network)
4. Evaluating on held-out test data
5. Deploying and monitoring

This process was effective but time-consuming and required machine learning expertise.

The LLM revolution

Modern LLMs have dramatically simplified text classification. Instead of training a custom model, you can write a prompt:

"Classify the following customer email into one of these categories: billing, technical support, account management, general enquiry. Return only the category name."

This zero-shot approach requires no training data and works surprisingly well. For higher accuracy, you can provide examples in the prompt (few-shot classification).

When to use which approach

• LLM prompting: Best for prototyping, low-volume tasks, or when categories change frequently. Simple to implement but more expensive per classification.
• Fine-tuned models: Best for high-volume production use. Lower cost per classification, higher consistency, but requires labelled data and ML expertise.
• Hybrid: Use LLMs to generate initial labels for training data, then train a lightweight model for production.

Evaluation metrics

• Accuracy: Percentage of correct classifications
• Precision: Of items classified as X, how many actually were X
• Recall: Of all actual X items, how many were correctly classified
• F1 score: Harmonic mean of precision and recall

Multi-label vs multi-class

Multi-class classification assigns exactly one label (an email is either spam or not). Multi-label classification can assign multiple labels (an article can be tagged as both "technology" and "business").