Entropy-based Recognition of Anomalous Answers for Efficient Grading of Short Answers with an Evolutionary Clustering Algorithm

Short answer question is a common assessment tool eliciting a specific textual response for knowledge assessment. The divide-and-grade is an automated grading approach that uses clustering to assign answers into sets each of which is supposed to be sufficiently similar to receive the same grade. This approach can potentially produce accurate grading with significantly reduced manual grading effort. Many current clustering methods are known to suffer from the presence of anomalies, answers that are deviated from the modes of model answers or common misconceptions. The freedom afforded to the composition of answers often lead to these anomalies, and in particular, contextual anomalous answers that are marginally correct or wrong. This paper proposes an evolutionary clustering method for the divide-and-conquer approach. The method is coupled with anomalous answer recognition based on an entropy formulation of the cluster membership of answers for rectifying misclustered answers. The method has been evaluated with an open short answer grading dataset and the accuracy compares favorably with existing algorithms. Result analysis has suggested that the method can effectively leverage small manual grading effort on issues of great impact on short answer grading accuracy.