Development of an Automated Scoring Model Using SentenceTransformers for Discussion Forums in Online Learning Environments

Abstract

Due to the limitations of public datasets, research on automatic essay scoring in Indonesian has been restrained and resulted in suboptimal accuracy. In general, the main goal of the essay scoring system is to improve execution time, which is usually done manually with human judgment. This study uses a discussion forum in online learning to generate an assessment between the responses and the lecturer's rubric in the automated essay scoring. A SentenceTransformers pre-trained model that can construct the highest vector embedding was proposed to identify the semantic meaning between the responses and the lecturer's rubric. The effectiveness of monolingual and multilingual models was compared. This research aims to determine the model's effectiveness and the appropriate model for the Automated Essay Scoring (AES) used in paired sentence Natural Language Processing tasks. The distiluse-base-multilingual-cased-v1 model, which employed the Pearson correlation method, obtained the highest performance. Specifically, it obtained a correlation value of 0.63 and a mean absolute error (MAE) score of 0.70. It indicates that the overall prediction result is enhanced when compared to the earlier regression task research. ACM CCS (2012) Classification: Computing methodologies → Modeling and simulation → Model development and analysis → Model verification and validation Applied computing → Education → Distance learning.