Software to Support Inductive Learning of Text-Markup Rules from Examples
Nick Benson

--

The task of forming assessments based on students' written responses can
be a daunting one, especially when presented with a large body of material
for analysis.
 
This talk describes work undertaken with Adam Carlson and Steve Tanimoto
on a facility within the INFACT online learning environment aimed at
easing the burden of this task.  By automatically proposing text-markup
rules to the instructor based on the instructor's previous manual markup
activity within an online discussion forum, the system is capable of
helping to identify possible occurrences of common misconceptions.
 
Our method makes use of the version-space machine learning approach and a
rule language based on word occurrences and distances between word pairs.
My work has focused primarily on integrating the rule-learning software
into the INFACT system, a large code base written in Java and Perl.

Major challenges encountered so far include dealing with the experimental
nature of the learning algorithm, constraints imposed by the existing
software system, and the unpredictiablity of student responses, in terms
of both content and syntax. Additionally, the system's "over
the shoulder" interface requires that great care be taken to provide 
useful and constructive information to the user, as opposed hindering the
user with unhelpful distractions.