Robustness Gym is a simple and extensible toolkit for robustness testing of NLP models that supports the entire spectrum of evaluation methodologies, from adversarial attacks to rule-based data augmentations. The Robustness Gym project is an ongoing collaboration between Stanford Hazy Research, Salesforce Research and UNC Chapel-Hill.

Despite impressive performance on standard benchmarks, deep neural networks often fail when deployed to real-world systems, due to distribution shifts, training artifacts, and noisy data. To address these vulnerabilities, Robustness Gym has been developed to overcome challenges in evaluating machine learning models today. It is a Python evaluation toolkit for natural language processing.

Entitled "Robustness Gym: Unifying the NLP Evaluation Landscape," the paper has been co-authored by Nazneen Rajani, Mohit Bansal, Karan Goel and others. They state: “Advances in natural language processing (NLP) have led to models that achieve high accuracy when train and test data are independent and identically distributed. However, analyses suggest that these models are not robust to data corruptions, distribution shifts, or harmful data manipulations, and they may rely on spurious patterns for prediction.”

“In this work, we identify challenges with evaluating NLP systems and propose a solution in the form of Robustness Gym a simple and extensible evaluation toolkit that unifies four standard evaluation paradigms: subpopulations, transformations, evaluation sets, and adversarial attacks. By providing a common platform for evaluation, Robustness Gym enables practitioners to compare results from all four evaluation paradigms with just a few clicks, and to easily develop and share novel evaluation methods using a built-in set of abstractions,” states the research paper.

Robustness Gym can be used to conduct new research analyses with ease. To validate this, the researchers have conducted the first study of academic and commercially available named entity linking (NEL) systems, as well as a study of the fine-grained performance of summarisation models. Commercial APIs from Microsoft, Google and Amazon have been compared to open-source systems BOOTLEG, WAT and REL across two benchmark datasets ie. Wikipedia and AIDA. NEL is a fundamental component of both search and question-answering systems such as conversational assistants, and has a widespread impact on the performance of commercial technology.

Robustness Gym supports a broad set of evaluation idioms and can be used for collaboratively building and sharing evaluations and results. A promising tool for researchers and practitioners, it has been embedded into the Contemplate → Create → Consolidate continual evaluation loop.