Zero Shot Classification Via Hypothesis Reformulation

via “zero-shot task reformulation via entailment”

text-classification model by undefined. 5,13,435 downloads.

Unique: Leverages MNLI fine-tuning to generalize inference patterns to arbitrary task formulations without task-specific training. The disentangled attention mechanism enables the model to reason about semantic relationships in novel hypothesis-premise pairs, making zero-shot reformulation more robust than models trained only on generic language modeling objectives.

vs others: Outperforms zero-shot classification with generic language models (GPT-2, BERT) because inference-specific training enables better reasoning about entailment relationships; more efficient than prompting large language models (GPT-3) for zero-shot tasks due to smaller model size and lower latency.