Dynabench: Rethinking Benchmarking in NLP
This addresses the need for more robust and informative benchmarks in the NLP community, though it is incremental as it builds on existing benchmarking concepts with a new dynamic approach.
The paper tackles the problem of NLP models achieving high benchmark performance but failing on simple challenges and real-world scenarios by introducing Dynabench, a platform for dynamic dataset creation and benchmarking, which has been applied to four initial NLP tasks to demonstrate its potential.
We introduce Dynabench, an open-source platform for dynamic dataset creation and model benchmarking. Dynabench runs in a web browser and supports human-and-model-in-the-loop dataset creation: annotators seek to create examples that a target model will misclassify, but that another person will not. In this paper, we argue that Dynabench addresses a critical need in our community: contemporary models quickly achieve outstanding performance on benchmark tasks but nonetheless fail on simple challenge examples and falter in real-world scenarios. With Dynabench, dataset creation, model development, and model assessment can directly inform each other, leading to more robust and informative benchmarks. We report on four initial NLP tasks, illustrating these concepts and highlighting the promise of the platform, and address potential objections to dynamic benchmarking as a new standard for the field.