Beat the AI: Investigating Adversarial Human Annotation for Reading Comprehension
This addresses the problem of improving reading comprehension models by testing adversarial annotation methods, but it is incremental as it builds on existing trends in annotation methodology.
The paper investigates adversarial human annotation for reading comprehension, where humans create questions to challenge models, collecting 36,000 samples with progressively stronger models. It finds that training on adversarial data generalizes well to non-adversarial datasets, with performance deteriorating as models in the loop strengthen, and stronger models can learn from data collected with weaker models, achieving 39.9 F1 vs. 41.0 F1 in comparisons.
Innovations in annotation methodology have been a catalyst for Reading Comprehension (RC) datasets and models. One recent trend to challenge current RC models is to involve a model in the annotation process: humans create questions adversarially, such that the model fails to answer them correctly. In this work we investigate this annotation methodology and apply it in three different settings, collecting a total of 36,000 samples with progressively stronger models in the annotation loop. This allows us to explore questions such as the reproducibility of the adversarial effect, transfer from data collected with varying model-in-the-loop strengths, and generalisation to data collected without a model. We find that training on adversarially collected samples leads to strong generalisation to non-adversarially collected datasets, yet with progressive performance deterioration with increasingly stronger models-in-the-loop. Furthermore, we find that stronger models can still learn from datasets collected with substantially weaker models-in-the-loop. When trained on data collected with a BiDAF model in the loop, RoBERTa achieves 39.9F1 on questions that it cannot answer when trained on SQuAD - only marginally lower than when trained on data collected using RoBERTa itself (41.0F1).