Publication

Abstract : The reliability of automated essay scoring (AES) has been the subject of debate among educators. Most systems treat essays as a bag of words and evaluate them based on LSA, LDA or other means. Many also incorporate syntactic information about essays such as the number of spelling mistakes, number of words and so on. Towards this goal, a challenging problem is to correctly understand the semantics of the essay to be evaluated so as to differentiate the intended meaning of terms used in the context of a sentence. We incorporate an unsupervised word sense disambiguation (WSD) algorithm which measures similarity between sentences as a preprocessing step to our existing AES system. We evaluate the enhanced AES model with the Kaggle AES dataset of 1400 pre-scored text answers that were manually scored by two human raters. Based on kappa scores, while both models had weighted kappa scores comparable to the human raters, the model with the WSD outperformed the model without the WSD.