Readerbench – automated feedback generation for essays in Romanian

The development of online environments has transformed written communication into one of the most frequently used types of interactions between individuals;this effect has increased even more during the COVID-19 pandemic, which imposed physical distancing restrictions. As writing is a key skill in everyday activities, it is important for people to have strong skills and to be capable to communicate their thoughts and beliefs in a structured form. This paper introduces automated scoring and feedback mechanisms for Romanian, derived from an online collection of freely available essays, and integrated in the ReaderBench platform. Several regression models are evaluated in terms of essay scoring accuracy, out of which Gradient Boosting Regression was selected based on its performance (R2 = .42, MAE = 1.10 on a 10-point scale). The feedback mechanisms provide suggestions for improving the quality of writings based on several rules, which in turn rely on the textual complexity indices computed by the ReaderBench framework, together with meaningful components generated from a Principal Component Analysis. © 2021, Politechnica University of Bucharest. All rights reserved.