Automatic scoring of Bangla language essay using generalized latent semantic analysis

Abstract

Automated Essay Grading (AEG) is a very important research area in educational assessment. Several AEG systems have been developed using statistical, Bayesian Text Classification Technique, Natural Language Processing (NLP), Artificial Intelligence (AI), and amongst many others. Latent Semantic Analysis (LSA) is an information retrieval technique used for automated essay grading. LSA forms a word by document matrix and the matrix is decomposed using Singular Value Decomposition (SVD) technique. It does not consider the word order in a sentence. Existing AEG systems based on LSA cannot achieve higher level of performance to be a replica of human grader. Moreover most of the essay grading systems are used for grading pure English essays or essays written in pure European languages. We have developed a Bangla essay grading system using Generalized Latent Semantic Analysis (GLSA) which uses n-gram by document matrix instead of word by document matrix of LSA. We have also developed an architecture for training essay set generation and evaluation of submitted essays by using the training essays. We have evaluated this system using real and synthetic datasets. We have developed training essay sets for three domains: standard Bangla essays titled “বাংলােদেশর sাধীনতা সংgাম”, “কািরগির িশkা” and descriptive answers of S.S.C level Bangla literature. We have gained 89% to 95% accuracy compared to human grader. This accuracy level is higher than that of the existing AEG systems.