Applying latent semantic analysis to computer assisted assessment in the Computer Science domain: a framework, a tool, and an evaluation

Abstract

This dissertation argues that automated assessment systems can be useful for both students and educators provided that the results correspond well with human markers. Thus, evaluating such a system is crucial. I present an evaluation framework and show how and why it can be useful for both producers and consumers of automated assessment systems. The framework is a refinement of a research taxonomy that came out of the effort to analyse the literature review of systems based on Latent Semantic Analysis (LSA), a statistical natural language processing technique that has been used for automated assessment of essays. The evaluation framework can help developers publish their results in a format that is comprehensive, relatively compact, and useful to other researchers. The thesis claims that, in order to see a complete picture of an automated assessment system, certain pieces must be emphasised. It presents the framework as a jigsaw puzzle whose pieces join together to form the whole picture. The dissertation uses the framework to compare the accuracy of human markers and EMMA, the LSA-based assessment system I wrote as part of this dissertation. EMMA marks short, free text answers in the domain of computer science. I conducted a study of five human markers and then used the results as a benchmark against which to evaluate EMMA. An integral part of the evaluation was the success metric. The standard inter-rater reliability statistic was not useful; I located a new statistic and applied it to the domain of computer assisted assessment for the first time, as far as I know. Although EMMA exceeds human markers on a few questions, overall it does not achieve the same level of agreement with humans as humans do with each other. The last chapter maps out a plan for further research to improve EMMA.