Assessing the robustness of automated scoring of divergent thinking tasks with adversarial examples

Automated scoring with machine learning has received considerable attention, especially for the Alternative Uses Test (AUT) of divergent thinking. In particular, supervised learning on word embeddings and approaches based on large language models predict human ratings with sufficient accuracy for many practical purposes. However, since automated systems will potentially be used without a ``human in the loop'', the robustness of any model is crucial, not only for the validity of basic research, but also for applications. We investigate the potential of adversarial examples (AEs) as robustness checks. AEs are synthetic responses obtained by subtle permutations of original responses that are assigned a substantially different score by the automated system. Specifically, we propose to employ synonyms to obtain semantically close synthetic responses. A synthetic response is an AE, when the deviation of the automated scores of the synthetic response and the original response is at least as large as the prediction error of the automated system. A random forest trained on GloVe word embeddings predicts human ratings of 2,690 distinct original AUT answers (stimuli: brick, paperclip) and for 42% AEs can be found, including both single and multi-word responses. On average, the AEs have slightly higher automated scores than the originals. While a small fraction of the AEs are invalid AUT responses, pronounced large deviations indicate that the prediction model is non-robust. Retraining including the AEs in the training data improves the robustness. We recommend to use AEs routinely to assess the robustness of automated scoring systems.