Text Classification: Approaches and Trade-offs

Text classification maps text → label. The right approach depends on how much labeled data you have and what inference cost/latency you can afford.

Approach 1: Zero-Shot Generative Prompt a capable LLM with the text and label options.

Classify this customer message into one of: [billing, shipping, returns, other]
Message: 'I never got my package and it's been 3 weeks'
Category:

• Data needed: 0 examples
• Latency: 1-5 seconds
• Cost: $0.001-0.05 per text
• Accuracy: excellent for ≤10 clear categories, degrades at 20+
• Best for: prototyping, rare labels, categories needing reasoning

Approach 2: Zero-Shot via Embedding Similarity Embed the text and each label string. Assign the closest label.

label_embeddings = {label: embed(label) for label in labels}
text_emb = embed(text)
predicted = max(labels, key=lambda l: cosine_sim(text_emb, label_embeddings[l]))

• Data needed: 0 examples
• Latency: < 100ms (no LLM generation)
• Accuracy: good for semantically clear labels, poor for domain-specific jargon

Approach 3: Embedding + Classifier Head Embed all labeled examples once, then train a classical ML classifier. Steps: embed corpus → split train/test → fit LogReg/SVM → evaluate → serve

• Data needed: 100+ examples (1K+ for reliable results)
• Latency: < 10ms at inference (no embedding API call if cached)
• Cost: negligible at inference (classifier only, no LLM)
• Accuracy: excellent for well-defined tasks with sufficient data

Approach 4: Fine-Tuned Encoder (BERT/RoBERTa) Update the encoder weights on your labeled data (full fine-tune or adapter).

• Data needed: 1K-100K examples
• Training cost: GPU hours
• Accuracy: highest ceiling, captures domain-specific patterns
• Use when: production-critical, domain is very specialized, have sufficient data

Few-Shot with In-Context Examples For generative approaches, include 3-5 (text, label) examples before the target. Improves accuracy significantly when labels are ambiguous or examples clarify edge cases.