Automated Essay Scoring dengan Fitur Feedback pada Essay Bahasa Indonesia Berbasis BERT

Abstract

Developing an effective and accurate Automated Essay Scoring (AES) system for assessing essays in Bahasa Indonesia is the focus of this research. In the context of higher education, Course Learning Outcomes (CLO) serve as a reference for instructors to evaluate students' abilities. One common evaluation method used is essay assessment, which, however, is time-consuming and requires proper score categorization to provide appropriate feedback. To address this challenge, an automated essay scoring and feedback system is needed, utilizing the Bidirectional Encoder Representations for Transformers (BERT) approach, capable of sequence classification tasks to obtain the true value of an essay. This research leverages Word2Vec for vector representation to assess similarity between essay answers and established reference essays. The dataset used in this study is sourced from the Kaggle competition titled "The Hewlett Foundation: Automated Essay Scoring," translated into Bahasa Indonesia, consisting of 3,258 essays and human-assigned scores. The base model utilized in this research is the "bert-base-uncase" model pre-trained on text classification tasks. This model is then fine-tuned using the dataset until achieving a perfect accuracy score of 1 and a satisfactory kappa score of 0.82. Subsequently, the model is employed to predict essay scores using text similarity, rubric scores, and BERT. The predicted scores provide appropriate feedback aligned with the author's final score.