THE GRAMMAR TUTOR Ben Parent Haibei Zhang Huong Huang Jasdeep Singh
Introduction GA is a grammar tutor for non-native English learners / TOEFL test takers One-to-one tutoring has been proven to be effective Transfer
Theoretical Background Intelligent tutoring system (ITS) is a computer program that simulates human, one-to-one tutoring
Theoretical Background Learning by being told: Clear articulation Clear diagnosis
Theoretical Background Structure: Expert knowledge module Student knowledge module Tutoring module User interface
Theoretical Background Design principles of Anderson tutors: Production rules Immediate error-feedback Path-control actions
Theoretical Background An example structure: Miss. Lindquist: Pre-test* Next question Level 0 Level 1 Level 2 r r r w w w “r” denotes “production rule based on a right answer” “w” denotes “production rule based on a wrong answer” ”*” indicates “correct answer detected, any more wrong answers are considered typos.” Level 0* r w w r Pre-test r w
Initial Design Concerns Mimic TOEFL-style questions Test one knowledge point per question Fit into our model-tracing, multiple- choice scenario “ Grammar & Structure ” Section
3-Part Structure Pre-Test Error-Logging Screen Tutoring Consistent with Anderson ’ s three main tenets of the model-tracing strategy Production rules, Immediate error- feedback, Path-control actions
Version 1 screen shot Log-in screen and welcome screen
Pre-Test Short group of questions (20-30) drawn from all knowledge points Designed to identify and make explicit the most common error category
Version 1 screen shot Pre-test question
Error-Logging Screen Compares user ’ s answers with correct answers Identifies main area of difficulty Logs pre-test score for later review by either user, teacher or designer
Version 1 screen shot Error-Logging Screen
Tutoring Each package designed to tutor one knowledge point Consists of short lesson and a problem set 3 levels of difficulty
Version 1 tutoring structure Brief Lesson Level 3 Level 2 Level 1 Congratulations Resources r r r r r r r w w ww w w w “r” denotes “right answer” “w” denotes “wrong answer”
Problems Enormity of domain (grammar) with thousands of production rules Not like Ms. Lindquist, which works in a very specific domain (symbolization) Many rule exceptions, which are simply declarative knowledge points Possible solution: limit knowledge points covered to “ most difficult ”
Problems Informant design needed Informant design necessary to determine “ most difficult ” points Informant design would be necessary to develop an effective hinting structure
Problems Program tutors only one knowledge point before exiting user Virtually inconceivable that a user would need tutoring on only one knowledge point
Version 2: current design Decision to limit knowledge points to “ most-difficult ” Some informant design done Pre-test elimintated
Version 2: current design “ Rolling question ” approach that tests user on all knowledge points determine a hierarchy of difficulty for the user (among knowledge points), Focuses user first on the knowledge points he/she has the most difficulty with.
Rolling Question Approach Top question level most difficult User stays at top level until question answered wrong # of correct questions record, once five of one category answered, that knowledge point drops out When user gets a question wrong, he/she immediately drops into tutoring
Version 2 tutoring structure
Version 2 screen shot Welcome screen
Version 2 screen shot Initial question
Version 2 screen shot If your answer is wrong …
Version 2 screen shot Hinting screen
Version 2 screen shot Now, try another simpler question
Version 2 screen shot Congratulations, you get it correct!
Version 2 screen shot Auxiliary function: user activity log
Version 2 screen shot Auxiliary function: save/open user profile
Version 3: Proposed future design Blah …
Version 3 tutoring structure Blah …