Where Does It Break?


Published on

Keynote at the European Semantic Web Conference (ESWC 2006). The talk tries to figure out what the main scientific challenges are in Semantic Web research.
This talk was also recorded on video, and is available on-line at http://videolectures.net/eswc06_harmelen_wswnj/

Published in: Technology, Education
  • Be the first to comment

  • Be the first to like this

No Downloads
Total views
On SlideShare
From Embeds
Number of Embeds
Embeds 0
No embeds

No notes for slide

Where Does It Break?

  1. 1. Where does it break? or: Why Semantic Web research is not just “Computer Science as usual” Frank van Harmelen AI Department Vrije Universiteit Amsterdam
  2. 2. But first: “the Semantic Web forces us to rethink the foundations of many subfields of Computer Science” if you can “the challenge of the Semantic Webscontinues to , read thishared break many often silently held andst be you mu research” assumptions underlying decades of ids on stero “I will try to identify silently held assumptions which are no longer true on the Semantic Web, prompting a radical rethink of many past results” 2
  3. 3. Oh no, not more “vision”… our plan is to invent the visionary sales are dropping some sort of leadership work is like a rock doohicky that done. How long will everyone wants to your part take? buy. Don’t worry, there will be lots of technical content 3
  4. 4. Grand Topics…  what are the science challenges in SW?  Which implicit traditional assumptions break?  Illustrated with 4 such “traditional assumptions” and also:  “Which Semantic Web” ? 4
  5. 5. Before we go on: Which Semantic Web are we talking about?
  6. 6. Typical SemWeb slide 1: General idea of Semantic Web  Make current web more machine accessible (currently all the intelligence is in the user)  Motivating use-cases q search q personalisation q semantic linking q data integration q web services q ... 6
  7. 7. Typical SemWeb slide 2: General idea of Semantic Web Do this by: 2. Making data and meta-data available on the Web in machine-understandable form (formalised) These are non-trivial design decisions. 3. Structure the data and meta-data in Alternative would be: ontologies 7
  8. 8. Which Semantic Web?  Version 1: "Semantic Web as Web of Data" (TBL)  recipe: expose databases on the web, use RDF, integrate  meta-data from: q expressing DB schema semantics in machine interpretable ways  enable integration and unexpected re-use 8
  9. 9. Which Semantic Web?  Version 2: “Enrichment of the current Web”  recipe: Annotate, classify, index  meta-data from: q automatically producing markup: named-entity recognition, concept extraction, tagging, etc.  enable personalisation, search, browse,.. 9
  10. 10. Which Semantic Web?  Version 1: “Semantic Web as Web of Data”  Version 2: “Enrichment of the current Web”  Different use-cases  Different techniques  Different users 10
  11. 11. Before we go on: The current state of the Semantic Web?
  12. 12. What’s up in the Semantic Web? The 4 hard questions: Q1: "where does the meta-data come from?”  NL technology is delivering on concept-extraction Q2: “where do the ontologies come from?”  many handcrafted ontologies  ontology learning remains hard  relation extraction remains hard Q3: “what to do with many ontologies?”  ontology mapping/aligning remains VERY hard Q4: “where’s the ‘Web’ in the Semantic Web?”  more attention to social aspects (P2P, FOAF)  non-textual media remains hard 12
  13. 13. What’s up in the Semantic Web? The 4 hard questions:  healthy uptake in some areas:  knowledge management / intranets  data-integration (Boeing)  life-sciences (e-Science)  convergence with Semantic Grid  cultural heritage  emerging applications in search & browse  Elsevier, Ilse, MagPie, KIM  very few applications in  personalisation  mobility/context awareness  Most applications for companies, few applications for the public 13
  14. 14. Semantic Web: Science or technology?
  15. 15. Semantic Web as Technology  better search & browse  personalisation  semantic linking  semantic web services  ... Semantic Web as Science 15
  16. 16. 4 examples of “where does it break?” d assumptions that no longer hold, d approaches that no longer work
  17. 17. 4 examples of “where does it break?”  Traditional complexity measures
  18. 18. Who cares about decidability?  Decidability ≈ completeness guarantee to find an answer, or tell you it doesn’t exist, given enough run-time & memory  Sources of incompleteness: q incompleteness of the input data q insufficient run-time to wait for the answer Completeness is unachievable in practice anyway, regardless of the completeness of the algorithm 18
  19. 19. Who cares about undecidability?  Undecidability ≠ always guaranteed not to find an answer  Undecidability = not always guaranteed to find an answer  Undecidability may be harmless in many cases; in all cases that matter 19
  20. 20. Who cares about complexity?  worst-case: may be exponentially rare  asymptotic  ignores constants 20
  21. 21. What to do instead?  Practical observations on RDF Schema: 6 9 q Compute full closure of O(105) statements  Practical observations on OWL: q NEXPTIME  but fine on many practical cases  Do more experimental performance profiles with realistic data  Think hard about “average case” complexity…. 21
  22. 22. 4 examples of “where does it break?”  Traditional complexity measures  Hard in theory, easy in practice
  23. 23. Example: Reasoning with Inconsistent Knowledge This work with Zhisheng Huang & Annette ten Teije
  24. 24. Knowledge will be inconsistent Because of:  mistreatment of defaults  homonyms  migration from another formalism  integration of multiple sources 24
  25. 25. New formal notions are needed  New notions: q Accepted: q Rejected: q Overdetermined: q Undetermined:  Soundness: (only classically justified results) 25
  26. 26. Basic Idea – Start from the query – Incrementally select larger parts of the ontology that are “relevant” to the query, until: • you have an ontology subpart that is Selection function small enough to be consistent and large enough to answer the query or • the selected subpart is already inconsistent before it can answer the query 26
  27. 27. General Framework s(T,φ,0) s(T,φ,2) s(T,φ,1) 27
  28. 28. More precisely: Use selection function s(T,φ,k), with s(T,φ,k) µ s(T,φ,k+1)  Start with k=0: s(T,φ,0) j¼ φ or s(T,φ,0) j¼ :φ ?  Increase k, until s(T,φ,k) j¼ φ or s(T,φ,k) j¼ :φ  Abort when q undetermined at maximal k 28
  29. 29. Nice general framework, but...  which selection function s(T,φ,k) to use?  Simple option: syntactic distance q put all formulae in clausal form: a1 Ç a2 Ç … Ç an q distance k=1 if some clausal letters overlap a1 Ç X Ç … Ç an, b1 Ç … X Ç bn q distance k if chain of k overlapping clauses are needed a1 Ç X Ç … X1 Ç an 29 b ÇX Ç…X Çb,
  30. 30. Evaluation Ontologies:  Transport: 450 concepts Communication: 200 concepts Madcow: 55 concepts Selection functions:  symbol-relevance = axioms overlap by ≥1 symbol  concept-relevance ≈ axioms overlap by ≥1 concept Query a random set of subsumption queries: Concept1 ⊆ Concept2 ? 30
  31. 31. Evaluation: Lessons this makes concept-relevance a high quality sound approximation (> 90% recall, 100% precision) 31
  32. 32. Works surprisingly well On our benchmarks, allmost all answers are “intuitive”  Not well understood why  Theory doesn’t predict that this is easy q paraconsistent logic, q relevance logic q multi-valued logic  Hypothesis: due to “local structure of knowledge”? 32
  33. 33. 4 examples of “where does it break?”  Traditional complexity measures  Hard in theory, easy in practice  context-specific nature of knowledge
  34. 34. Opinion poll left right meaning of a sentence meaning of sentence is only determined is not only determined by the sentence itself, by the sentence itself, and not influenced by but is also influenced by the surrounding by the surrounding sentences, sentences, and not by the situation and also by the situation in which the sentence in which the sentence is used is used 34
  35. 35. Opinion poll left right don’t you see what I mean? 35
  36. 36. Example: Ontology mapping with community support This work with Zharko Aleksovski & Michel Klein
  37. 37. The general idea background knowledge anchoring anchoring inference source target mapping 37
  38. 38. Example 1 38
  39. 39. Example 2 39
  40. 40. Results Example matchings discovered q OLVG: Acute respiratory failure cardiale AMC: Asthma cardiale q OLVG: Aspergillus fumigatus cause AMC: Aspergilloom q OLVG: duodenum perforation abnormality AMC: Gut perforation q OLVG: HIV cause AMC: AIDS q OLVG: Aorta thoracalis dissectie type B location AMC: Dissection of artery 40
  41. 41. Experimental results  Source & target = flat lists of ±1400 ICU terms each  Anchoring = substring + simple germanic morphology  Background = DICE (2300 concepts in DL) 41
  42. 42. New results:  more background knowledge makes mappings better q DICE (2300 concepts) q MeSH (22000 concepts) q ICD-10 (11000 concepts)  Monotonic improvement of quality  Linear increase of cost 100 90 80 70 60 50 40 30 20 10 0 1 2 3 42 4
  43. 43. Distributed/P2P setting background knowledge anchoring anchoring inference source target mapping 43
  44. 44. So…  The OLVG & AMC terms get their meaning from the context in which they are being used.  Different background knowledge would have resulted in different mappings  Their semantics is not context-free  See also: S-MATCH by Trento 44
  45. 45. 4 examples of “where does it break?”  Traditional complexity measures  Hard in theory, easy in practice  context-specific nature of knowledge  logic vs. statistics
  46. 46. Logic vs. statistics  DB schema’s & integration is only logic, no statistics  AI is both logic and statistics, but completely disjoint  Find combinations of the two worlds? q Statistics in the logic? q Statistics to control the logic? q Statistics to define the semantics of the logic? 46
  47. 47. Statistics in the logic? Fuzzy DL  (TalksByFrank v InterestingTalks) ¸ 0.7  (Turkey:EuropeanCountry) · 0.2 1  youngPerson = Person u9age.Young Young(x) = 0 10yr 30yr 1  veryYoungPerson = Person u9age.very(Young) 0 10yr 30yr 47 Umberto Straccia
  48. 48. Statistics to control the logic?  query: A v B ?  B = B1 u B2 u B3 A v B1, A v B2, A v B3 ? B2 B1 A B3 48
  49. 49. Statistics to control the logic?  Use “Google distance” to decide which ones are reasonable to focus on  Google distance ≈symmetric conditional probability of co-occurrence ≈estimate of semantic distance ≈ estimate of “contribution” to A v B1 u B2 u B3 B2 B1 A B3 This work by Riste Gligorov 49
  50. 50. Statistics to define semantics?  Many peers have many mappings on many terms to many other peers  Mapping is good if results of “whispering game” are truthful  Punish mappings that contribute to bad whispering results  Network will converge to set of good mappings (or at least: consistent) This work by Karl Aberer 50
  51. 51. Statistics to define semantics?  Meaning of terms = relations to other terms  Determined by stochastic process  Meaning ≈ stable state of self-organising system  statistics = getting a system to a meaning-defining stable state  logic = description of such a stable state  Note: meaning is still binary, classical truth-value  Note: same system may have multiple stable states… 51
  52. 52. 4 examples of “where does it break?”  Traditional complexity that no longer hold,  old assumptions measures don’t work completeness, decidability, complexity Sometimes “hard in theory, no longer work  old approaches that easy in practice” Q/A over inconsistent ontologies is easy, but why?  Meaning dependent on context meaning determined by background knowledge  Logic versus statistics statistics in the logic statistics to control the logic statistics to determine semantics
  53. 53. Final comments  These 4 “broken assumptions/old methods” were just examples. There are many more. (e.g. Hayes, Halpin on identity, equality and reference)  Notice that they are interlinked, e.g  hard theory/easy practice &  complexity  meaning in context &  logic/statistics  Working on these will not be SemWeb work per se, but q they will be inspired by SemWeb challenges q they will help the SemWeb effort (either V1 or V2) 53
  54. 54. Have fun with the puzzles! 54