Adversarial Training for Commonsense Inference
This work addresses the challenge of improving commonsense reasoning in NLP models, but it is incremental as it builds on existing fine-tuning methods with a novel regularization technique.
The paper tackled the problem of commonsense inference in reading comprehension by proposing an adversarial training algorithm that perturbs word embeddings to regularize the model, achieving competitive results on multiple datasets without external resources.
We propose an AdversariaL training algorithm for commonsense InferenCE (ALICE). We apply small perturbations to word embeddings and minimize the resultant adversarial risk to regularize the model. We exploit a novel combination of two different approaches to estimate these perturbations: 1) using the true label and 2) using the model prediction. Without relying on any human-crafted features, knowledge bases, or additional datasets other than the target datasets, our model boosts the fine-tuning performance of RoBERTa, achieving competitive results on multiple reading comprehension datasets that require commonsense inference.