IEEE Standard for Robustness Evaluation Test Methods for a Natural Language Processing Service That Uses Machine Learning

The natural language processing (NLP) services using machine learning have rich applications in solving various tasks and have been widely deployed and used, usually accessible by application programming interface (API) calls. The robustness of the NLP services is challenged by various well-known general corruptions and adversarial attacks. Inadvertent or random deletion, addition, or repetition of characters or words are examples of general corruptions. Adversarial characters, words, or sentence samples are generated by adversarial attacks, causing the models underpinning the NLP services to produce incorrect results. A method for quantitatively evaluating the robustness the NLP services is proposed by this standard. Under the method, different cases the evaluation needs to perform against are specified. Robustness metrics and their calculation are defined. With the standard, understanding of the robustness of the services can be developed by the service stakeholders including the service developer, service providers, and service users. The evaluation can be performed during various phases in the life cycle of the NLP services, the testing phase, in the validation phase, after deployment, and so forth.