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.

via “zero-shot classification via hypothesis reformulation”

zero-shot-classification model by undefined. 80,926 downloads.

Unique: Repurposes NLI task (premise-hypothesis entailment) as a general-purpose zero-shot classification mechanism by treating input text as premise and category labels as hypotheses, enabling classification without task-specific fine-tuning or labeled data

vs others: More flexible than traditional zero-shot classifiers (e.g., CLIP for images) because it works with arbitrary text categories defined at inference time; more accurate than keyword/regex-based classification because it understands semantic relationships; requires no labeled data unlike supervised classifiers