130 likes | 390 Views
Project Halo . Chris Kirkland Bret Wilson. Background. Vulcan, Inc.’s Goal Pilot Phase (2002) 30-50% correct (SRI, Cycorp , Ontoprise ) Halo Phase II Initiated 2004 Intermediate Testing 2006 Final Testing 2008-9. Halo Phase II. AP Chemistry, Biology, and Physics Questions (4 types):
E N D
Project Halo Chris Kirkland Bret Wilson
Background • Vulcan, Inc.’s Goal • Pilot Phase (2002) • 30-50% correct (SRI, Cycorp, Ontoprise) • Halo Phase II • Initiated 2004 • Intermediate Testing 2006 • Final Testing 2008-9
Halo Phase II • AP Chemistry, Biology, and Physics • Questions (4 types): • Conceptual Questions • Mathematical Equations • Tables* • Diagrams**
AURA • Automated User-Centered Reasoning and Acquisition System • Layout • Document Base • Knowledge Base • Inference Engine • User Interface
Document Base • Biology : 44 pp (23%) • Cell structure, function, and division; DNA rep. and protein synthesis • Chemistry : 67 (11%) • Stoichiometry, Chemical Equilibria, aqueous reactions, acids and bases • Physics : 78 (15%) • Kinematics and Newtonian Dynamics
Knowledge Machine (KM) • Prototypes • Semantic Nets • Unification Mapping (UMAP) • Component Library (CLP) • Independent Library • “Attach, Penetrate, Physical Object, Location…” (37)
Inference Engine • Pattern Matching • Equation Solver • Inference Tracker* • Explanation Generator* • Outputs “chain of reasoning” to user
Computer Processable Language (CPL) • Simplified English syntax • Multiple choice questions split into list of true/false questions • Complex questions broken into simpler parts • Simple format: • Subject + Verb + Complements + Adjuncts • Not allowed: “probably”, “mostly”
User Querying • User enters question in CPL • System notifies of any errors • Correct questions are shown as graphical feedback • Answer contains: • Simple direct answer • Explanation in basic English
BBN Evaluation • Results • Biology • with nonexpert KF, nonexpert QF outperformed expert QF (why?) • Overall 47% success • Chemistry • no significant differences • Overall 18% success • Physics • Expert KF, QF outperformed nonexpert KF, QF • Overall 36% success
Experts vs. non-experts • Experts • Expert in domain (physics, chemistry, biology) • Extensive training, previous experience with AURA • Collaboration with AURA team members • Non-experts • Graduate-level experience in domain (KF) • Undergraduate-level experience in domain (QF) • Limited training, no previous experience with AURA
MUKE (India) • Multi-User Knowledge Entry • High performance (75% success rate on all novel questions in biology)
Difficulties / Concerns • CPL format is limited and handles multiple choice inelegantly • KF not automated • No handling for diagrams in Document Base • Optimized to small subset of curriculum • Difficult to encode physics vector equations