Automatic Essay Scoring: A Review on the Feature Analysis Techniques

Automatic Essay Scoring (AES) is the automatic process of identifying scores for a particular essay answer. Such a task has been extensively addressed by the literature where two main learning paradigms have been utilized: Supervised and Unsupervised. Within these paradigms, there is a wide range of feature analyses has been utilized, Morphology, Frequencies, Structure, and semantics. This paper aims at addressing these feature analysis types with their subcomponent and corresponding approaches by introducing a new taxonomy. Consequentially, a review of recent AES studies is being conducted to highlight the utilized techniques and feature analysis. The finding of such a critical analysis showed that the traditional morphological analysis of the essay answer would lack semantic analysis. Whereas, utilizing a semantic knowledge source such as ontology would be restricted to the domain of the essay answer. Similarly, utilizing semantic corpus-based techniques would be impacted by the domain of the essay answer as well. On the other hand, using essay structural features and frequencies alone would be insufficient, but rather as an auxiliary to another semantic analysis technique would bring promising results. The state-of-the-art in AES research concentrated on neural-network-based-embedding techniques. Yet, the major limitations of these techniques are represented as (i) finding an adequate sentence-level embedding when using models such as Word2Vec and Glove, (ii) ‘out-of-vocabulary when using models such as Doc2Vec and GSE, and lastly, (iii) ‘catastrophic forgetting’ when using BERT model.