Automated Essay Scoring and the Deep Learning Black Box: How Are Rubric Scores Determined?

This article investigates the feasibility of using automated scoring methods to evaluate the quality of student-written essays. In 2012, Kaggle hosted an Automated Student Assessment Prize contest to find effective solutions to automated testing and grading. This article: a) analyzes the datasets from the contest - which contained hand-graded essays - to measure their suitability for developing competent automated grading tools; b) evaluates the potential for deep learning in automated essay scoring (AES) to produce sophisticated testing and grading algorithms; c) advocates for thorough and transparent performance reports on AES research, which will facilitate fairer comparisons among various AES systems and permit study replication; d) uses both deep neural networks and state-of-the-art NLP tools to predict finer-grained rubric scores, to illustrate how rubric scores are determined from a linguistic perspective, and to uncover important features of an effective rubric scoring model. This study's findings first highlight the level of agreement that exists between two human raters for each rubric as captured in the investigated essay dataset, that is, 0.60 on average as measured by the quadratic weighted kappa (QWK). Only one related study has been found in the literature which also performed rubric score predictions through models trained on the same dataset. At best, the predictive models had an average agreement level (QWK) of 0.53 with the human raters, below the level of agreement among human raters. In contrast, this research's findings report an average agreement level per rubric with the two human raters' resolved scores of 0.72 (QWK), well beyond the agreement level between the two human raters. Further, the AES system proposed in this article predicts holistic essay scores through its predicted rubric scores and produces a QWK of 0.78, a competitive performance according to recent literature where cutting-edge AES tools generate agreement levels between 0.77 and 0.81, results computed as per the same procedure as in this article. This study's AES system goes one step further toward interpretability and the provision of high-level explanations to justify the predicted holistic and rubric scores. It contends that predicting rubric scores is essential to automated essay scoring, because it reveals the reasoning behind AIED-based AES systems. Will building AIED accountability improve the trustworthiness of the formative feedback generated by AES? Will AIED-empowered AES systems thoroughly mimic, or even outperform, a competent human rater? Will such machine-grading systems be subjected to verification by human raters, thus paving the way for a human-in-the-loop assessment mechanism? Will trust in new generations of AES systems be improved with the addition of models that explain the inner workings of a deep learning black box? This study seeks to expand these horizons of AES to make the technique practical, explainable, and trustable.