Automatic Short Answer Grading onHigh School’s E-Learning Using Semantic Similarity Methods

Abstract

Grading students’ answers has always been a daunting task which takes a lot of teachers’ time. The aim of this study is to grade students’ answers automatically in a high school’s e-learning system. The grading process must be fast, and the result must be as close as possible to the teacher assigned grades. We collected a total of 840 answers from 40 students for this study, each already graded by their teachers. We used Python library sentence-transformers and three of its latest pre-trained machine learning models (all-mpnet-base-v2, all-distilroberta-v1, all-MiniLM-L6-v2) for sentence embeddings. Computer grades were calculated using Cosine Similarity. These grades were then compared with teacher assigned grades using both Mean Absolute Error and Root Mean Square Error. Our results showed that all-MiniLM-L6-v2 gave the most similar grades to teacher assigned grades and had the fastest processing time. Further study may include testing these models on more answers from more students, also fine tune these models using more school materials.