Abstract:
Queries received by tutors on the Dr Math mathematics tutoring service are created in a domain-specific form of microtext. The aim of the service is to help South African school learners to master mathematical concepts, but not all of the queries received on the service contain content relevant to the tutoring process.
This paper contrasts various methods to classify learner queries automatically as relevant or not, in order to determine whether such a process could approximate human judgement. A back-propagation artificial neural network, a decision tree, a Bayesian filter, a k-means clustering algorithm and a rule-based filter are compared.
The results of the classification techniques are contrasted with the results of three human coders, using the metrics of precision, recall, F-measure and the Pearson correlation co-efficient. Both the rule-based filter and neural network deliver classification results which closely reflect the classifications made by the human coders.
