A framework for evaluating and explaining the robustness of NLP models

The standard practice for evaluating the generalisation of supervised machine learning models in NLP tasks is to use previously unseen (i.e. held-out) data and report the performance on it using various metrics such as accuracy. Whilst metrics reported on held-out data summarise a model's performance, ultimately these results represent aggregate statistics on benchmarks and do not reflect the nuances in model behaviour and robustness when applied in real-world systems.

We propose a robustness evaluation framework for NLP models concerned with arguments and facts, which encompasses explanations for robustness failures to support systematic and efficient evaluation. We will develop novel methods for simulating real-world texts stemming from existing datasets, to help evaluate the stability and consistency of models when deployed in the wild. The simulation methods will be used to challenge NLP models through text-based transformations and distribution shifts on datasets as well as on data sub-sets that capture linguistic patterns, to provide a systematic coverage of real-world linguistic phenomena. Furthermore, our framework will shed insights into a model's robustness by generating explanations for robustness failures along the lexical, morphological, and syntactic dimensions, extracted from the various dataset simulations and data sub-sets, thus departing from current approaches that solely provide a metric to quantify robustness. We will focus on two NLP research areas, argument mining and fact verification, however, several simulation methods and the robustness explanations are also scalable to other NLP tasks.