SCENE: Self-Labeled Counterfactuals for Extrapolating to Negative Examples
This addresses the costly and domain-specific challenge of manually collecting negative examples for models in tasks like question answering and entailment, offering a scalable solution.
The paper tackles the problem of detecting negative examples in natural language understanding tasks by proposing SCENE, an automatic method for synthesizing training data from only positive examples, which closes 69.6% of the performance gap on SQuAD 2.0 and improves generalization to out-of-domain benchmarks.
Detecting negatives (such as non-entailment relationships, unanswerable questions, and false claims) is an important and challenging aspect of many natural language understanding tasks. Though manually collecting challenging negative examples can help models detect them, it is both costly and domain-specific. In this work, we propose Self-labeled Counterfactuals for Extrapolating to Negative Examples (SCENE), an automatic method for synthesizing training data that greatly improves models' ability to detect challenging negative examples. In contrast with standard data augmentation, which synthesizes new examples for existing labels, SCENE can synthesize negative examples zero-shot from only positive ones. Given a positive example, SCENE perturbs it with a mask infilling model, then determines whether the resulting example is negative based on a self-training heuristic. With access to only answerable training examples, SCENE can close 69.6% of the performance gap on SQuAD 2.0, a dataset where half of the evaluation examples are unanswerable, compared to a model trained on SQuAD 2.0. Our method also extends to boolean question answering and recognizing textual entailment, and improves generalization from SQuAD to ACE-whQA, an out-of-domain extractive QA benchmark.