1 / 40

L ogics for D ata and K nowledge R epresentation

L ogics for D ata and K nowledge R epresentation. Semantic Matching. Outline. Introduction Matching Problems Graphs Matching Syntactic vs. Semantic Semantic Relations Semantic Matching in ClassL. Introduction.

evita
Download Presentation

L ogics for D ata and K nowledge R epresentation

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Logics for Data and KnowledgeRepresentation Semantic Matching

  2. Outline • Introduction • Matching Problems • Graphs Matching • Syntactic vs. Semantic • Semantic Relations • Semantic Matching in ClassL

  3. Introduction • An important problem on graphs that we will apply to richer representations like concept hierarchies, classifications, schemas and ontologies is “matching.” • This problem is also popular in semantic networks, where one may want to check whether a particular concept is present. • E.g.

  4. Date: a matching for marriage?

  5. Introduction (cont’) • Some popular situations that can be modeled as a matching problem are: - Concept matching in semantic networks. - Schema matching in distributed databases. - Ontology matching (ontology “alignment”) in the Semantic Web. - ...

  6. ? ? ? The Semantic Matching Problem (Example)

  7. Outline • Introduction • Matching Problems • Graphs Matching • Syntactic vs. Semantic • Semantic Relations • Classification matching • L.O. and ClassL • Semantic Matching in ClassL

  8. Matching Problems • There are two kinds of matching: • Syntactic: matching of nodes as objects or strings (so, as such, without meaning). • Semantic: matching of nodes as concepts. • A Matching Problem (syntactic or semantic) is a problem on graphs summarized as:Given two finite graphs, is there a matching between the (nodes of the) two graphs?

  9. Matching Problems • A problem of matching can be decomposed in two steps: • Extract the graphs from the conceptual models under consideration; • Match the resulting graphs. • Below we show some examples of step 1. (We follow [Giunchiglia & Shvaiko, 2007].)

  10. Relational DB Schemas • Let us consider the following relational database (RDB) model, say “BANK”: (Giunchiglia & Shvaiko, 2007)

  11. Relational DB Schemas: Representation 1 • We can represent the RDB model “BANK” as a graph (a tree) with root “BANK”: • The RDB model is first partitioned into relations, then attributes and data instances. (Giunchiglia & Shvaiko, 2007)

  12. Relational DB Schemas: Representation 2 • We can represent the RDB model “BANK” as a graph (a tree) with root “BANK”: • The model is partitioned into relations, then into tuples, attributes and data instances. (Giunchiglia & Shvaiko, 2007)

  13. Relational DB Schemas: NOTEs • Which of the two representations is more preferable depends on the concrete task? • It is always possible to transform one representation into the other. • In contrast to the example of RDB “BANK”, DB schemas are seldom trees. More often, DB schemas are translated into Directed Acyclic Graphs (DAG’s).

  14. OODB Schemas • Let us consider the RDB “BANK” in terms of an object-oriented DB (OODB) schema: BRANCH (Street, City, Zip) PERSON (F_Name, L_Name) STAFF : PERSON (Position, Salary, Manager) • The resulting graph is: (Giunchiglia & Shvaiko, 2007)

  15. OODB Schemas: NOTEs • OODB schemas capture more semantics than the relational DBs. • In particular, an OODB schema: • explicitlyexpresses subsumption relations between elements; • admits special types of arcs for part/whole relationships in terms of aggregation and composition.

  16. RDB vs. OODB

  17. Semi-structured Data • Neither RDBs nor OODBs capture all the features of semi-structured or unstructured data (Buneman, 1997): • semi-structured data do not possess a regular structure (schemaless); • the “structure” of semi-structured data could be partial or even implicit. • Typical examples are: HTML and XML.

  18. XML Schemas • XML schemas can be represented as DAGs. • The graph from the RDB “BANK” could also be obtained from an XML schema. (Giunchiglia & Shvaiko, 2007)

  19. XML Schemas: NOTEs • Often XML schemas represent hierarchical data models. • In this case the only relationships between the elements are {is-a}. • Attributes in XML are used to represent extra information about data. There are no strict rules telling us when data should be represented as elements, or as attributes.

  20. Concept Hierarchies • Def. A concept hierarchy is asemi-formal conceptualization of an application domain in terms of concepts and relationships. (Giunchiglia & Shvaiko, 2007)

  21. Concept Hierarchies: NOTEs • Examples are classification hierarchies, e.g., and directories (catalogs). • Classification hierarchies / Web directories are sometimes referred to as lightweight ontologies (Uschold & Gruninger, 2004). However: • They are not ontologies, as they lack of a formal semantics (semi-formal vs formal.) • They don’t formalize class instances.

  22. Semantic Matching • Introduction • Matching Problems • Graphs Matching • Syntactic vs. Semantic • Semantic Relations • Semantic Matching via SAT in ClassL

  23. How to deal with these heterogeneity? • Matching: given two graph-like structures (e.g., concept hierarchies or ontologies), produce a mapping between the nodes of the graphs that semantically correspond to each other Matching

  24. Matching Syntactic Matching Semantic Matching Syntactic vs. Semantic Relations are computed between concepts at nodes R = { =, ⊑, ⊒, , ⊓} Relations are computed between labels at nodes R = {x[0,1]} Note:all previous systems are syntactic… Note:needed for proper labeling of context mappings

  25. Semantic Matching Mapping elementis a 4-tuple < IDij, n1i, n2j, R >, where • IDij is a unique identifier of the given mapping element; • n1i is the i-th node of the first graph; • n2jis the j-th node of the second graph; • R specifies a semantic relation between the concepts at the given nodes ComputedR’s, listed in the decreasing binding strength order: equivalence { = }; more general/specific {⊒,⊑}; mismatch {  }; overlapping {⊓}. Semantic Matching: Given two graphs G1and G2, for any node n1iG1,find the strongest semantic relation R’ holding with node n2jG2

  26. Images Europe ? 1 1 Europe = ? Pictures Wine and Cheese 2 2 3 ? 5 4 4 3 Austria Italy Italy Austria < ID22, 2, 2, = > < ID22, 2, 2, = >  < ID21, 2, 1, ⊑ > < ID24, 2, 4, ⊒ > Example Step 4

  27. S-Match Algorithm Four Macro Steps Given two labeled trees T1 and T2, do: • For all labels in T1 and T2 compute concepts at labels • For all nodes in T1 and T2 compute concepts at nodes • For all pairs of labels in T1 and T2 compute relations between concepts at labels • For all pairs of nodes in T1 and T2 compute relations between concepts at nodes Steps 1 and 2 constitute the preprocessing phase, and are executed once and each time after the schema/ontology is changed (OFF- LINE part) Steps 3 and 4 constitute the matching phase, and are executed every time the two schemas/ontologies are to be matched (ON - LINE part)

  28. Step 1: compute concepts at labels • The idea: • Translate natural language expressions into internal formal language • Compute concepts based on possible senses of words in a label and their interrelations • Preprocessing: • Tokenization. Labels (according to punctuation, spaces, etc.) are parsed into tokens. E.g., Wine and Cheese <Wine, and, Cheese>; • Lemmatization. Tokens are morphologically analyzed in order to find all their possible basic forms. E.g., ImagesImage; • Building atomic concepts. An oracle (WordNet) is used to extract senses of lemmatized tokens. E.g., Image has 8 senses, 7 as a noun and 1 as a verb; • Building complex concepts. Prepositions, conjunctions, etc. are translated into logical connectives and used to build complex conceptsout of the atomic concepts E.g.,CWine and Cheese = <Wine, U(WNWine)> ⊔ <Cheese, U(WNCheese)>, where ⋃ is a union of the senses that WordNet attaches to lemmatized tokens

  29. Europe 1 Pictures Wine and Cheese 2 3 4 5 Austria Italy Step 2: compute concepts at nodes • The idea: extend concepts at labels by capturing the knowledge residing in a structure of a graph in order to define a context in which the given concept at a label occurs • Computation (basic case): Concept at a node for some node n is computed as an intersection of concepts at labels located above the given node, including the node itself C4= Ceurope⊓CPictures⊓CItaly

  30. T1 T2 Europe Images 1 1 Wine and Cheese Pictures Europe 2 2 3 T1 T2 CEurope CPictures CItaly CAustria Italy CWine CCheese 4 5 Austria Austria Italy 3 4 CImages = CEurope = CAustria  = CItaly =  Step 3: compute relations between concepts at labels • The idea:Exploit a priori knowledge, e.g., lexical, domain knowledge with the help of element level semantic matchers • Results of step 3:

  31. Step 4? Wait! • What we have done in Step 3? Find semantic relations. • What for? To build the semantic relation bases for further matching. • What is a ‘relation base’ in the logic sense? TBox! We are building the TBox!

  32. Step 4: compute relations between concepts at nodes • The idea: Reduce the matching problem to a validity problem • Context: the relations between concepts at labels(from step 3) • Wffrel(C1i, C2j): relation to be proved (with: C1i in Tree1 and C2j in Tree2) • Translate into propositional logic: • C1i = C2jis translated into C1iC2j • C1i subsumes C2jis translated into C1iC2j • C1i C2jis translated into ¬ (C1i  C2j) • Prove Context Wffrel (C1i, C2j) is valid with Rel={=,⊑,⊒,⊥,⊓} • A propositional formula is valid iff its negation is unsatisfiable (SAT deciders are sound and complete…)

  33. Pseudo code of Step4 CHANGE LINE15 • 1. i, j, N1, N2: int; • 2. context, goal: wff; • 3. n1, n2: node; • 4. T1, T2: tree of (node); • 5. relation = {=, ⊑ , ⊒ ,  }; • 6. ClabMatrix(N1, N2), CnodMatrix(N1, N2), relation: relation • 7. function mkCnodMatrix(T1, T2, ClabMatrix) { • 8. for (i= 0; i < N1; i++) do • 9. for (j= 0; j < N2; j++) do • 10. CnodMatrix(i, j):=NodeMatch(T1(i),T2(j), ClabMatrix)} • 11. function NodeMatch(n1, n2, ClabMatrix) { • 12. context:=mkcontext(n1, n2, ClabMatrix, context); • 13. foreach (relation in < =, ⊑ , ⊒ ,  >) do { • 14. goal:= w2r(mkwff(relation, GetCnod(n1), GetCnod(n2)); • 15. if VALID(mkwff(, context, goal)) • 16. return relation;} • 17. return IDK;} Validity Reasoning Contex|=goal?

  34. Context Goal T1 T2 C21 C22 C23 C24 C25 C11 C12 = C13  C14 =  Example • Example. Suppose we want to check if C12 = C22 (C1Images  C2Pictures)  (C1Europe  C2Europe)  (C12  C22 )

  35. T1 T1 T1 T1 T2 T2 T2 T2 Europe Europe Europe Europe Images Images Images Images 1 1 1 1 1 1 1 1 Wine and Cheese Wine and Cheese Wine and Cheese Wine and Cheese Pictures Pictures Pictures Pictures Europe Europe Europe Europe 2 2 2 2 2 2 2 2 3 3 3 3 Italy Italy Italy Italy 4 4 4 4 5 5 5 5 Austria Austria Austria Austria Austria Austria Austria Austria Italy Italy Italy Italy 3 3 3 3 4 4 4 4 Examples = 

  36. Another Example: Concept Hierarchies (Serafini et. al., 2003)

  37. Example (cont’) • Suppose we want to discover the relation R between Chat and Forum in the Google directory (left) and Chat and Forum in the Yahoo directory (right):

  38. Example (cont’) • Step 1 : transformation of nodes n1 = Chat and Forum and n2 = Chat and Forum to propositions, P and Q; selection of the portion T of knowledge (background theory) relevant to the application of a SAT solver. • WordNet is used at this step to build T. • Step 2 : select the relation R between n1 and n2 among the semantic relations of interest.

  39. Example (cont’) • Step 3 : The question “Is Chat and Forum less general than Chat and Forum?” becomes the SAT problem “Is T |= P ⊑ Q?” where:P = (art#1 ⊓ literature#2 ⊓(chat#1 ⊔ forum#1)),Q = (art#1⊔humanities#1)⊓humanities#1⊓(chat#1⊔ forum#1), T = {art#1⊑ humanities#1, humanities#1⊒ literature#2}

  40. References( Preprints available at http://dit.unitn.it/~ldkr/#Biblio ) • F. Giunchiglia, P. Shvaiko, “Semantic matching.” Knowledge Engineering Review, 18(3):265-280, 2003. • L. Serafini, P. Bouquet, B. Magnini, S. Zanobini, “An Algorithm for Matching Contextualized Schemas via SAT.” IRST Technical Report 0301-06, ITC, January 2003. • F. Giunchiglia, M. Marchese, I. Zaihrayeu. “Encoding Classifications into Lightweight Ontologies.” J. of Data Semantics VIII, Springer-Verlag LNCS 4380, pp 57-81, 2007.

More Related