Overview

1. Ontology-Based Query and Answering in Chemistry: OntoNova @ Project HaloJürgen Angele - Ontoprise, Karlsruhe, Germany

2. Overview • Project • Scenario / Participants • Technical Approach • Performance strategies • Metareasoning / Justifications • Encoding / Knowledge Base • Encoding Method • Architecture of the KB • Challenge • Results • Question encoding, fidelity • Failures, brittleness

3. Vulcan Inc: OntoBroker passes Advanced Placement Test • Background • A multistage project towards the Development of a Digital Aristotle • Funded by Vulcan Inc. Seattle • Phase 1 successfully closed in 2003 • Phase 2 since January 2004 • Functions (Halo 1) • Capturing of extensive set of chemical knowledge • System passed the „Advanced Placement Test“ • Query is answered and answer is explained

4. Stage 1: Halo 1 • task: development of a query answering system containing the knowledge of about 80 pages of chemistry book in 4 months • evaluation: • sequestration of system • 160 novel questions (AP exam) • encoding of the questions • chemistry professors graded answers and explanations • participants: SRI, Cyc Corp, Ontoprise

5. Overview • Project • Scenario / Participants • Technical Approach • Performance • Metareasoning / Justifications • Encoding / Knowledge Base • Encoding Method • Architecture of the KB • Challenge • Results • Question encoding, fidelity • Failures, brittleness

6. Implementation - F –logic • Knowledge Encoding in F-Logic • Query-Encoding in F-Logic • Evaluation in a batch-run by Ontobroker

7. High Performance Inferencing

8. Inference Server Inference Server C o n n e k t o r s C o n n e k t o r s Inference Server Inference Server F-Logic Compiler F-Logic Compiler Prolog Compiler Prolog Compiler RDF Compiler RDF Compiler .... Compiler .... Compiler Inference kernel Inference kernel Builtins Builtins Inference kernel Inference kernel Internal Database Internal Database Internal Database Internal Database Metareasoning & Answer Justification Answer = A explanation knowledge base knowledge base

9. Metareasoning & Answer Justification Explanations the products of this reaction are PbI2 and Na because PbI2 precipitates out of the solution an ionic molecule consisting of cation Pb and anion I is not known to be soluble and is thus guessed to be unsoluble

10. Metareasoning & Answer Justification Reasoning for generating explanations • integrating additional knowledge into explanations • generating abstractions • avoiding redundancies • considering context and user profile

11. Overview • Project • Scenario / Participants • Technical Approach • Performance strategies • Metareasoning / Justifications • Encoding / Knowledge Base • Encoding Method • Architecture of the KB • Challenge • Results • Question encoding, fidelity • Failures, brittleness

12. Verify by syllabus questions Model ground operations Encoding Modelling procedure refinement ofmodeling refinement ofexplanationstesting addexplanationrules Verify ground operations by example questions from Brown et al. 40 test cases 2002/2003 dry run challenge run

13. Question encoding A 0.3M solution of acetic acid has a pH of 2.63. The ionization constant of this acid is a) 1.8 x 10-5 b) 7.0 x 10-4 c) 1.1 x 10-6 d) 7.8 x 10-3 d) 1.9 x 10-6

14. Question encoding – multiple choice strategy m1:Mixture[hasComponents->>{"HCl","Ba(OH)2"}]. m2:Mixture[hasComponents->>{"HCl","CaCO3"}]. m3:Mixture[hasComponents->>{"HCl","CuSO4"}]. m4:Mixture[hasComponents->>{"HCl","Na3PO4"}]. m5:Mixture[hasComponents->>{"HCl","NaCl"}]. answer("A") <- exists P P:GaseousReaction[fromMixture->>m1]. answer("B") <- exists P P:GaseousReaction[fromMixture->>m2]. answer("C") <- exists P P:GaseousReaction[fromMixture->>m3]. answer("D") <- exists P P:GaseousReaction[fromMixture->>m4]. answer("E") <- exists P P:GaseousReaction[fromMixture->>m5]. FORALL X <- answer(X). Input facts definition of alternatives ask for alternative

15. Question encoding – Detailed Answer Section MC48 1.0 L of a buffer formed by mixing 0.25 moles of ammonia solution with 0.25 moles of ammonium nitrate has a pH of (For ammonia, Kb = 1.8 x 10-5) m1:BufferSolution[ hasComponents->> {ammonia ,ammonium_nitrate,}; hasMole@(ammonium_nitrate)->0.25; hasMole@(ammonia )->0.25; hasVolume->1.0]. FORALL Ph <- m1[hasPHValue->Ph].

16. Encoding - Basic Chemical Operations Classify compound as ionic (aequous.flo, utils.flo) Balance chemical equation (balancing.flo) Determine solubility (acidbase.flo) Determine equilibrium expression Rank strength of metal ions as lewis acids ..... ..... ..... Determine acid/base conjugate Determine strengths of acids/bases Determine products of reaction and type of reaction Calculate PH (ph.flo) Determine position of equilibrium Describing substances Naming

17. Ontology ontologicalaccess basic chemical operations using predicates like MPhKa and rules acidorder ... ... ... ... ... ... ... ... ... ... calculate PH-value equilibrium acidorder solubility balancing reactions Instances basic facts like elements,... Architecture of KB

18. Summary Architecture KB • chemical operations: independent knowledge chunks - collaborative development of KB - reduce complexity - reduce testing effort • OO – Wrapper: ontological access - eases access - closer to NL

19. Overview • Project • Scenario / Participants • Technical Approach • Performance • Metareasoning / Justifications • Encoding / Knowledge Base • Encoding Method • Architecture of the KB • Challenge • Results • Question encoding, fidelity • Failures, brittleness

20. Results

21. Results

22. Performance

23. Brittleness Classification • (MOD) Knowledge Modeling • (IMP) Knowledge Implementation/Modeling Language • (INF) Inference and Reasoning • (KFL) Knowledge Formation and Learning • (SCL) Scalability: • (MGT) Knowledge Management • (QMN) Query Management • (ANJ) Answer Justification • (QMT) Quality Metrics (MTA) Meta Capabilities

24. burn("CH4"). reacts with oxygen = burn Question encoding - fidelity MC12 When methane, CH4, gas reacts with oxygen, the following changes occur

25. basic operation modeled to determine the pH-value given the Ka-value and not vice versa Brittleness – not expected question type DA18 Ascorbic acid, H2C6H6O6, is a diprotic acid with a Ka1 value of 8.9 x 10-5. The pH of a 0.125 M solution of ascorbic acid is 2.48 and the concentration of C6H6O62- is 1.6 x 10-12 M. Determine the value of Ka2.

26. Results from Halo 1 • controlled experiment • brittleness classification • bottleneck: knowledge acquisition ! • 10000 $ per page • (5000 $ per page OP)

27. Next Steps: Halo 2 development of tools for domain experts to capture knowledge and thus to reduce knowledge acquisition bottleneck

28. Next Steps Scenario: Term acquisition

29. QF V Comp… KF E

30. QF V From your choice of answers it looks as if “weak acid” is a concept. Give me hints Do you have typical examples for “weak acid” in this context? Can you give specializations for “weak acid” in this context? Does “weak acid” refer to a set of elements? Finish· Hints are available about where to add the term to the ontology. Give me the hints· Comp… KF E weak acid

31. Next Steps Scenario: Creating rules

32. KF QF E V Formula: Comp… Rule NL Formula Editor

33. KF QF E V hasSalt Formula: WeakAcid C hasAcid hasSalt hasH-value Salt WeakAcid BufferSolution Comp… Rule NL If BufferSolution hasAcid WeakAcid Acidand BufferSolution hasSaltSalt If BufferSolution hasAcid WeakAcid Acidand BufferSolution hasSaltSaltandSalt isSaltOf the WeakAcid If BufferSolution hasAcid WeakAcid Formula Editor hasSalt hasAcid hasAcid isSaltOf

34. KF QF E V Formula: Salt BufferSolution WeakAcid Comp… Rule NL If BufferSolution hasAcid WeakAcid Acidand BufferSolution hasSaltSaltandSalt isSaltOf the WeakAcid and the Attribute moleSalt of BufferSolution has value [X] and the Attribute moleAcid of BufferSolution has value [HX] and the Attribute hasKa of WeakAcid has value [Ka] then the Attribute hasHValue of BufferSolution is computed to [H] according to the formula [H]=Ka*HX/X. If BufferSolution hasAcid WeakAcid Acidand BufferSolution hasSaltSaltandSalt isSaltOf the WeakAcid If BufferSolution hasAcid WeakAcid Acidand BufferSolution hasSaltSaltandSalt isSaltOf the WeakAcid and the Attribute moleSalt of BufferSolution has value [X] and the Attribute moleAcid of BufferSolution has value [HX] If BufferSolution hasAcid WeakAcid Acidand BufferSolution hasSaltSaltandSalt isSaltOf the WeakAcid and the Attribute moleSalt of BufferSolution has value [X] If BufferSolution hasAcid WeakAcid Acidand BufferSolution hasSaltSaltandSalt isSaltOf the WeakAcid and the Attribute moleSalt of BufferSolution has value [X] and the Attribute moleAcid of BufferSolution has value [HX] and the Attribute hasKa of WeakAcid has value [Ka] Formula Editor HX=hasMoleAcid X=hasMoleSalt hasSalt H=hasHvalue hasAcid isSaltOf Ka=hasKa

35. Next Steps Scenario: Diagram acquisition

36. Example 4.1 A Traffic Light at Rest A traffic light weighing 100 N hangs from a vertical cable tied to two other cables that are fastened to a support, as in Figure Input diagram 4.11a. The upper cables make angles of 37.0˚ and 53.0˚ with the horizontal. Find the tension in each of the three cables. Knowledge Formulation: Diagrams The user decides that the diagram is an integral part of the text, and selects it. The system then opens a new perspective on the diagram and text.

37. A traffic light weighing 100 N hangs from a vertical cable tied to two other cables that are fastened to a support, as in Figure 4.11a. The upper cables make angles of 37.0˚ and 53.0˚ with the horizontal. Knowledge Formulation: Diagrams Based on the text accompanying the diagram, a set of potential glyphs are shown to the user.

38. Set mass A traffic light weighing 100 N hangs from a vertical cable tied to two other cables that are fastened to a support, as in Figure 4.11a. The upper cables make angles of 37.0˚ and 53.0˚ with the horizontal. Find the tension in each of the three cables. Set weight Representing the Initial Scenario • Glyphs are then overlaid on the existing diagram (as for Question Formulation) • As glyphs are added, they may also be linked to the descriptive text. • Dimension values (e.g., weight) and labels can be added by right-clicking on the glyphs and selecting items from a popup menu.

39. Boosted SemanticWeb applications Scientific SemanticWeb based on Halo-2 technology Distributed knowledge KillerApps: SemanticWeb - Editor (from “Frontpage”  DarkMatterStudio) - Browser (from “IE”  DarkMatterQueryInterface) First W3C Standards SemanticWeb Pilot Applications Increasing SemanticWeb Resources Slow maturity process of SemanticWeb applications Halo-2 will boost the SemanticWeb Impact of the Digital Aristotle 2003 2004 2005 2006 2007 RESEARCH COMMERCE

40. CMU: natural language understanding U Brighton: intelligent querying with natural language Team The Team DFKI:usability & intelligent interfaces Ontoprise: reasoning, integration, semantic web Georgia Tech:understanding diagrams and pictures

41. Founded: 1999 (Spin Off Univ. Karlsruhe) Team: 30 Employees Context: “Semantic Europe” (~ 100 R&D) - AIFB Karlsruhe - FZI, Karlsruhe - DERI Galway, Irland - DERI Innsbruck, Austria Technology: • Technology Leader (Gartner Group, Forrester Research) • Vision: SemanticWeb

42. Thank you! Prof. Dr. Jürgen Angele angele@ontoprise.de +49 (0)721 509 809 0