Cheating Automatic Short Answer Grading: On the Adversarial Usage of Adjectives and Adverbs

Abstract

Automatic grading models are valued for the time and effort saved during the instruction of large student bodies. Especially with the increasing digitization of education and interest in large-scale standardized testing, the popularity of automatic grading has risen to the point where commercial solutions are widely available and used. However, for short answer formats, automatic grading is challenging due to natural language ambiguity and versatility. While automatic short answer grading models are beginning to compare to human performance on some datasets, their robustness, especially to adversarially manipulated data, is questionable. Exploitable vulnerabilities in grading models can have far-reaching consequences ranging from cheating students receiving undeserved credit to undermining automatic grading altogether -even when most predictions are valid. In this paper, we devise a black-box adversarial attack tailored to the educational short answer grading scenario to investigate the grading models' robustness. In our attack, we insert adjectives and adverbs into natural places of incorrect student answers, fooling the model into predicting them as correct. We observed a loss of prediction accuracy between 10 and 22 percentage points using the state-of-the-art models BERT and T5. While our attack made answers appear less natural to humans

Short-Text Semantic Similarity (STSS): Techniques, Challenges and Future Perspectives

“…Several studies import datasets using Google Colab, Python 3.6, and GPU TeslaK80 with 12 GB RAM. The SciEntsBank further includes unseen answers (UA), unseen questions (UQ), and an unseen domain (UD) [133][134][135][136]. Moreover, Table 12 illustrates the summary of the datasets.…”

Evaluating Spatial Visualization Learning Through Digitized Sketches: A Case Study of Engineering Students' Orthographic Projection Errors

“…There is no evidence that the student looked at solutions provided by a classmate, but this hypothesis is worth considering especially since the student repeatedly submitted the correct solution in a single attempt in a short amount of time. It should be noted that, student's gaming a CBL system is a recognized area of study ( [34], [35]).…”

“…Thus, the student was able to get the advantage of a Peek, but lost only a single star rather than 2 stars that a Peek would have cost. It should be noted that "gaming the system" is a recognized issue with CBL systems ( [34], [35]). Interestingly the student chose to scribble only in the Top View rather than all views even though scribbling everywhere would have shown the complete solution with the same star penalty.…”