Zero-Shot Learning in NLP: Generalization Without Examples

Module: Natural Language Processing | Difficulty: Advanced

Zero-Shot Classification

Chain-of-Thought Prompting

Instruction Tuning

Few-Shot vs Zero-Shot

| Method | MNLI | SST-2 | QNLI | |--------|------|-------|------| | Zero-Shot | 72.3 | 85.6 | 68.1 | | One-Shot | 78.1 | 89.2 | 75.3 | | Few-Shot | 82.5 | 91.8 | 80.2 |

def zero_shot_classify(model, tokenizer, text, labels):
    scores = []
    for label in labels:
        prompt = f"{text} This is about {label}."
        inputs = tokenizer(prompt, return_tensors='pt')
        with torch.no_grad():
            outputs = model(**inputs)
            log_prob = outputs.logits[0, -1, inputs.input_ids[0, -1]]
        scores.append(log_prob.item())
    return labels[scores.index(max(scores))]

Research Insight: Chain-of-thought prompting improves zero-shot reasoning by 20-40% on math and logic tasks. The key insight is that intermediate reasoning steps provide a scaffold for the model to build upon, similar to how humans solve complex problems.