Task-based Example Miner for Intelligent Tutoring Systems.

摘要

Intelligent tutoring systems (ITS) aim to provide customized resources or feedback on a subject (commonly known as domain in ITS) to students in real-time, emulating the behavior of an actual teacher in a classroom. This thesis designs an ITS based on an instructional strategy called example-ba sed learning (EBL), that focuses primarily on students devoting their time and cognitive capacity to studying worked-out examples so that they can enhance their learning and apply it to similar graded problems or tasks. A task is a graded problem or question that an ITS assigns to students (e.g. task T1 in C programming domain defined as 'Write an assignment instruction in C that adds 2 integers'). A worked-out example refers to a complete solution of a problem or question in the domain. Existing ITS systems such as NavEx and PADS, that use EBL to teach their domain suffer from several limitations such as (1) methods used to extract knowledge from given tasks and worked-out examples require highly trained experts and are not easily applicable or extendable to other problem domains (e.g. Math), either due to use of manual knowledge extraction methods (such as Item Objective Consistency (IOC)) or highly complex automated methods (such as syntax tree generation) (2) recommended worked-out examples are not customized for assigned tasks and therefore are ineffective in improving student success rate.;This thesis proposes a new modular model for an EBL-based ITS called Example Recommendation System (ERS). ERS extracts knowledge in terms of basic learning units (LU) (e.g. scanf is a LU in the domain of C programming) from all task solutions and worked-out examples in its domain by using regular expression analysis and represents this knowledge in vector space. The prime contribution of knowledge extraction method of ERS is its extendibility to new domains without requiring highly trained experts. Experiments on two different domains show that LUs are extracted with 81correctness for domain 1 (Programming in C) and 95% for domain 2 (Programming in Miranda). Knowledge extraction also serves as a crucial data pre-processing step for ERS, which then uses the extracted knowledge to mine its repository of worked-out examples using data mining methods such as k-nearest neighbors, in order to generate customized list of examples for each task in its domain. The accuracy of ERS's customization model is 93%, while its f_score is 88%. An evaluation of ERS demonstrates that the key elements (simpler and efficient automated knowledge extraction, extendibility to other domains, task-based customization, and clear integration of all components) have been accomplished and the overall goal of optimizing learning has been achieved. Experiments show that students score an average of 89% in tasks for which ERS recommends worked-out examples, compared to an average of 73% for tasks that students attempt without using any such examples.