CRESST ONR/NETC Meetings, July 2003, v1 17 July, 2003 ONR Advanced Distributed Learning Greg Chung Bill Bewley UCLA/CRESST Ontologies and Bayesian Networks in Assessment 2003 Regents of the University of California
CRESST ONR/NETC Meetings, July 2003, v1 1 Problem Statement How do you link information from assessments to individualized instructional recommendations in a DL context? –Content is going online –Assessments are going online –Couple content and assessment
CRESST ONR/NETC Meetings, July 2003, v1 2 Ontologies An ontology is a conceptual representation of a domain expressed in terms of concepts and the relationships among the concepts –Support knowledge capture, representation, sharing –Fielded technology Medical, engineering, e-commerce, military… Development tools and APIs available today (e.g., Protégé) Support rapid development and testing of prototypes
CRESST ONR/NETC Meetings, July 2003, v1 3 Bayesian Networks Graphical modeling of the causal structure of a phenomenon in terms of nodes and relations –Nodes represent states, links represent the influence relations –Supports fusion of observable data (e.g., “correct on item 1”) into high-level hypotheses (e.g., “understands breath control”) –Fielded technology Development tools available (HUGIN, MSBNx) ONR-funded research
CRESST ONR/NETC Meetings, July 2003, v1 4 Assessment Application Example Content recommendation –Deliver individualized instructional content based on assessment results Approach –Use a domain ontology to represent content –Use assessments to measure students’ knowledge of the domain –Use a Bayesian network to model knowledge dependencies
CRESST ONR/NETC Meetings, July 2003, v1 5 Rifle Marksmanship Ontology Capture hierarchical structure of a domain –Field manuals, doctrine, training videos –Bind content to structure (text, video, graphics) Capture conceptual representation –Experts (coaches, snipers, rifle team) –Upper-level ontology captured using knowledge maps
CRESST ONR/NETC Meetings, July 2003, v1 6 Based on corpus of marksmanship literature and doctrine Currently 168 concepts (classes) Content directly bound to each node –Important if you want to make use of the information Hierarchical Representation
CRESST ONR/NETC Meetings, July 2003, v1 7 Binding Content to Structure
CRESST ONR/NETC Meetings, July 2003, v1 8 Application of Ontology Marksmanship ontology serves as testbed to evaluate feasibility of approach –Pilot test of approach — 2nd Lts. undergoing entry-level marksmanship training –Design Individualized content recommendation vs. control (no recommendation) –Examine shooting outcome, learning outcomes, changes in BN due to instruction, Marines’ perceptions of learning
CRESST ONR/NETC Meetings, July 2003, v1 9 Linking Assessment and Instruction Approach –Depict knowledge dependencies among marksmanship concepts using a Bayesian network –Administer assessment to gather information on Marines’ understanding of rifle marksmanship –Take assessment results — item-level data — and update BN
CRESST ONR/NETC Meetings, July 2003, v1 10 Linking Assessment and Instruction Approach (continued) –Identify concepts that have low probabilities in BN — interpreted as poor understanding –Make use of cognitive demands of tasks and items to infer depth of a Marine’s understanding –Deliver different content based on depth of understanding
CRESST ONR/NETC Meetings, July 2003, v1 11 Content Recommendation
CRESST ONR/NETC Meetings, July 2003, v1 12 Example of Feedback
CRESST ONR/NETC Meetings, July 2003, v1 13 Preliminary Results Within-group analyses –BN probabilities increased for concepts that had instructional content served up –BN probabilities did not change for concepts that did not have instructional content –High-level BN topics correlated with measure it was derived from as well as reasoning measure –BN “scores” corresponded with Marines’ self- ratings of their level of knowledge (80% agreement)
CRESST ONR/NETC Meetings, July 2003, v1 14 Preliminary Results Between-group analyses inconclusive –Small sample size (n = 16) –Experimental-condition Marines Qualified in thunderstorm Learned more from classroom training than expected (i.e., > 70% of topics “correct”) Knowledge map scores appear to be increasing at a faster rate than the control group, but differences not statistically significant
CRESST ONR/NETC Meetings, July 2003, v1 15 Summary An important opportunity of online assessments is the potential to measure many aspects of human behavior under a variety of different conditions An important challenge is extracting meaningful information from (potentially) voluminous amounts of data Bayesian networks and ontologies may be one approach to address