Next Generation Semantic Web Applications

Transcription

1 Next Generation Semantic Web Applications Enrico Motta, Mathieu D Aquin Sofia Angeletou, Claudio Baldassarre, Martin Dzbor, Laurian Gridinoc, Davide Guidi, Ainhoa Llorente, Vanessa Lopez, Marta Sabou Knowledge Media Institute The Open University Milton Keynes, UK

2 Introduction This talk presents a number of projects, which are part of an integrated effort at exploring the possibilities opened by the Semantic Web, viewed as a domain-independent, large scale supplier of formally encoded background knowledge, with respect to enabling intelligent problem solving. We call the resulting applications: Next Generation Semantic Web Applications

3 Organization of the Talk The Semantic Web The Semantic Web in the context of AI research Next Generation Semantic Web Applications What are they? Why are they different from 1st generation SW Applications? Infrastructure needs Examples Ontology Matching Integrating Web2.0 and SW Semantic Web Browsing Question Answering Conclusions

4 The Semantic Web The collection of all formal, machine processable, web accessible, ontologybased statements (semantic metadata) about web resources and other entities in the world, expressed in a knowledge representation language based on an XML syntax (e.g., OWL, DAML, DAML+OIL, RDF, etc )

5 Ontology Metadata <RDF triple> <RDF triple> <RDF triple> <RDF triple> <RDF triple> <RDF triple> <RDF triple> <RDF triple> <RDF triple> <RDF triple> <RDF triple> <RDF triple> <RDF triple> <RDF triple> <RDF triple> <RDF triple> <RDF triple> <RDF triple> <RDF triple> <RDF triple> <RDF triple> <RDF triple> <RDF triple> <RDF triple> <RDF triple> <RDF triple> <RDF triple> <RDF triple> UoD

6

7 Semantic Web Document <foaf:personalprofiledocument rdf:about=" <dc:title>sofia Angeletou&apos;s RDF Description</dc:title> <rdfs:label>sofia Angeletous RDF Description</rdfs:label> <dc:description>rdf description for Sofia Angeletou in machine-read <dc:creator rdf:resource=" <foaf:maker rdf:resource=" <foaf:primarytopic rdf:resource=" </foaf:personalprofiledocument> <foaf:person rdf:about=" <foaf:name>sofia Angeletou</foaf:name> <foaf:firstname>sofia</foaf:firstname> <foaf:surname>angeletou</foaf:surname>

8 Increasing Semantic Content <rdf:rdf> <Featurerdf:about=" <name>shenley Church End</name> <alternatename>shenley</alternatename> <incountry rdf:resource=" w.geonames.org </rdf:rdf>

9 Charting the web

10 Charting the web (2)

11 Distribution of the documents according to their coverage Domain Coverage on the SW Distribution of documents in the 16 top categories of DMOZ Great variety: Some topics are almost not covered (e.g. Adult), while some are over represented (e.g. Society, Computers) As we can expect, a large number of narrow coverage documents and a small number of large coverage ones.

12 Example: Annotating the queen's birthday dinner <RDF triple> <RDF triple> <RDF triple> <RDF triple> <RDF triple> <RDF triple> <RDF triple> <RDF triple> <RDF triple> <RDF triple> <RDF triple> <RDF triple>

13 Knowledge Sparseness Nr.(t1, t2, t3) (t1) (t2) (t3) (t1, t2) (t1, t3) (t2, t3) (t1, t2, t3) 1 (project, article, researcher) (researcher, student, university) (research, publication, author) (adventurer, expedition, photo) (mountain, team, talk) (queen, birthday, dinner) (project, relatedto, researcher) (researcher, workswith, Ontology) (academic, memberof, project) (article, hasauthor, person) (person, trip, photo) (woman, birthday, dinner) (person, memberof, project) (publication, hasauthor, person)

14 Example: Annotating the queen's birthday dinner

15 The Rise of Semantics

16 Thesis #1 The SW today has already reached a level of scale good enough to make it a very useful source of knowledge to support intelligent applications In other words: the Semantic Web is no longer an aspiration but a reality The availability of such large scale amounts of formalised knowledge is unprecedented in the history of AI

17 Thesis #2 The SW may well provide a solution to one of the classic AI challenges: how to acquire and manage large volumes of knowledge to develop truly intelligent problem solvers and address the brittleness of traditional KBS

18 Knowledge Representation Hypothesis in AI Any mechanically embodied intelligent process will be comprised of structural ingredients that we as external observers naturally take to represent a propositional account of the knowledge that the overall process exhibits, and independent of such external semantic attribution, play a formal but causal and essential role in engendering the behaviour that manifests that knowledge Brian Smith, 1982

19 Intelligence as a function of possessing domain knowledge Large Body of Knowledge KA Bottleneck Intelligent Behaviour

20 The Knowledge Acquisition Bottleneck Knowledge Large Body of Knowledge KA Bottleneck Intelligent Behaviour

21

22 SW as Enabler of Intelligent Behaviour Intelligent Behaviour

23 Overall Goal Our research programme is to contribute to the development of this large-scale web of data and develop a new generation of web applications able to exploit it to provide intelligent functionalities

24 First Generation Semantic Web Applications

25

26 Bibliographic Data CS Dept Data AKT Reference Ontology <rdf :Description rdf :abo ut=" ht t p:/ /ww w.ecs. sot on.ac.uk/info/#p erso n-01269"> <ns 0 :family -name>gibbins</ns0:family -name> <ns 0 :full -name>nicholas Gibbins</ns0:full -name> <ns 0 :given-name>n icholas</ns0:g iven-name> <ns 0 :has- - address> nmg@ecs.so t on.ac.u k</ns0:has - - address> <ns 0 :has-affi liation -to -unit rdf: reso urce=" ht t p:// / WEBSITE/GOW /Vie wdepartme nt.aspx?dep art ment =750"/> < / rdf :Descriptio n> </ rdf :RDF> RDF Data

27

28

29 Features of 1st generation SW Applications Typically use a single ontology Usually providing a homogeneous view over heterogeneous data sources. Limited use of existing SW data Closed to semantic resources Hence: current SW applications are more similar to traditional KBS (closed semantic systems) than to 'real' SW applications (open semantic systems)

30 It is still early days

31

32 Next Generation Semantic Web Applications

33 Architecture of NGSW Apps

34 Issue: Semantic Web Infrastructure

35 Current Gateway to the Semantic Web

36 Limitations of Swoogle Limited quality control mechanisms Many ontologies are duplicated Limited Query/Search mechanisms Only keyword search; no distinction between types of elements No support for formal query languages (such as SPARQL) Limited range of ontology ranking mechanisms Swoogle only uses a 'popularity-based' one Limited API No support for ontology modularization and combination

37 A New Gateway to the Semantic Web

38 Sophisticated quality control mechanism Detects duplications Fixes obvious syntax problems E.g., duplicated ontology IDs, namespaces, etc.. Structures ontologies in a network Using relations such as: extends, inconsistentwith, duplicates Provides efficient API Supports formal queries (SPARQL) Variety of ontology ranking mechanisms Modularization/Combination support Plug-ins for Protégé and NeOn Toolkit (under devpt.) Very cool logo!

39 Networked Ontologies M 1 M 2 target source source O 2 relatedwith priorversionof O 1 O 1 priorversionof O 1 dependson extends incompatiblewith O 3 O 4

40

41 Sophisticated quality control mechanism Detects duplications Fixes obvious syntax problems E.g., duplicated ontology IDs, namespaces, etc.. Structures ontologies in a network Using relations such as: extends, inconsistentwith, duplicates Provides efficient API Supports formal queries (SPARQL) Variety of ontology ranking mechanisms Modularization/Combination support Plug-ins for Protégé and NeOn Toolkit (under devpt.) Very cool logo!

42

43

44

45

46

47 Examples of Next Generation Semantic Web Applications

48 Example #1: Ontology Matching

49 Ontology Matching Label similarity methods e.g., Full_Professor = FullProfessor Structure similarity methods Using taxonomic/property related information

50 New paradigm: use of background knowledge Background Knowledge (external source) A R B A R B

51 External Source = One Ontology Aleksovski et al. EKAW 06 Map (anchor) terms into concepts from a richly axiomatized domain ontology Derive a mapping based on the relation of the anchor terms Assumes that a suitable (rich, large) domain ontology (DO) is available.

52 External Source = Web van Hage et al. ISWC 05 rely on Google and an online dictionary in the food domain to extract semantic relations between candidate terms using IR techniques A + OnlineDictionary IR Methods rel B Does not rely on a rich Domain Ont, Precision increases significantly if domain specific sources are used: 50% - Web; 75% - domain texts.

53 External Source = SW Proposal: rely on online ontologies (Semantic Web) to derive mappings ontologies are dynamically discovered and combined Semantic Web Does not rely on any preselected knowledge sources. A rel B M. Sabou, M. d Aquin, E. Motta, Using the Semantic Web as Background Knowledge in Ontology Mapping", Ontology Mapping Workshop, ISWC 06. Best Paper Award

54 Strategy 1 - Definition Find ontologies that contain equivalent classes for A and B and use their relationship in the ontologies to derive the mapping. Semantic Web B 1 A 1 A 2 B 2 A n O 1 O 2 O n B n For each ontology use these rules: A' A' A' A' B' B' => A B' A B' => A A B B B B A rel B These rules can be extended to take into account indirect relations between A and B, e.g., between parents of A and B : A' C C B' A' B'

55 Strategy 1- Examples Semantic Web AcademicStaff Researcher ka2.rdf Semantic Web Food MeatOrPoultry RedMeat Beef Tap Researcher ISWC AcademicStaff SWRC Beef SR-16 Food FAO_Agrovoc

56 Strategy 2 - Definition Principle: If no ontologies are found that contain the two terms then combine information from multiple ontologies to find a mapping. Semantic Web A C rel C rel B B Details: (1) Select all ontologies containing A equiv. with A (2) For each ontology containing A : A' C A' C ( r1) A' C C (a) if find relation between C and B. (b) if find relation between C and B. B A B ( r2) A' C C B A B ( r3) A' C C B A B A rel B ( r4) A' C C ( r5) A' C C B A B B A B

57 Strategy 2 - Examples Ex1: Chicken Vs. Food Chicken Poultry Poultry Food (midlevel-onto) (Tap) (r1) Chicken Food (Same results for Duck, Goose, Turkey) Ex2: Ham Vs. Food Ham Meat Meat Food (pizza-to-go) (SUMO) (r1) Ham Food Ex3: Ham Vs. Seafood Ham Meat Meat Seafood (pizza-to-go) (wine.owl) (r3) Ham Seafood

58 Large Scale Evaluation Matching AGROVOC (16k terms) and NALT(41k terms) (derived from 180 different ontologies) Evaluation: 1600 mappings, two teams Overall performance comparable to best in class M. Sabou, M. d Aquin, W.R. van Hage, E. Motta, Improving Ontology Matching by Dynamically Exploring Online Knowledge. In Press

59 Chart 2 Ontologies (180) used to derive mappings. TAP CPE Mid-level-ontology.daml SUMO.daml Economy.daml

60 Thesis #3 Using the SW to provide dynamically background knowledge to tackle the Agrovoc/NALT mapping problem provides the first ever test case in which the SW, viewed as a large scale heterogeneous resource, has been successfully used to address a realworld problem

61

62 Thesis #4 The claim that the information on the SW is of poor quality and therefore not useful to support intelligent problem solving is a myth not supported by concrete experience: Our experience in the NALT/Agrovoc ontology matching benchmark problem shows that without any particularly intelligent filter, the info available on the SW already allows a 85% theoretical precision for our algorithm, well beyond the performance of any other ontology matching algorithm

63 Example #2: Integrating SW and Web2.0

64 Features of Web2.0 sites Tagging as opposed to rigid classification Dynamic vocabulary does not require much annotation effort and evolves easily Shared vocabulary emerge over time certain tags become particularly popular

65 Limitations of tagging Different granularity of tagging rome vs colosseum vs roman monument Flower vs tulip Etc.. Multilinguality Spelling errors, different terminology, plural vs singular, etc This has a number of negative implications for the effective use of tagged resources e.g., Search exhibits very poor recall

66 Giving meaning to tags

67 What does it mean to add semantics to tags? 1. Mapping a tag to a SW element "japan" <akt:country Japan> 2. Linking two "SW tags" using semantic relations {japan, asia} <japan subregionof asia>

68 Applications of the approach To improve recall in keyword search To support annotation by dynamically suggesting relevant tags or visualizing the structure of relevant tags To enable formal queries over a space of tags Hence, going beyond keyword search To support new forms of intelligent navigation i.e., using the 'semantic layer' to support navigation

69 Pre-processing Tags Clean tags Group similar tags Filter infrequent tags Folksonomy Analyze co-occurrence of tags Co-occurence matrix Cluster tags Clustering Concise tags Cluster 1 Cluster 2 Cluster n Yes Remaining tags? No END 2 related tags Find mappings & relation for pair of tags <concept, relation, concept> SW search engine Wikipedia Google Concept and relation identification

70 Pre-processing Scope: Subsets of Flickr and del.icio.us tags. Total Distinct # entries # tags # users # resources # tags del.icio.us 19,605 89,978 7,164 14,211 11,960 Flickr 49, ,130 6,140 49,087 17,956 Pre-processing (thresholds): To be similar, Levenshtein >= 0.83; A tag has to occur at least 10 times. Total Distinct # entries # tags # users # resources # tags del.icio.us 18,882 70,194 7,090 13,579 1,265 Flickr 44, ,098 5,321 44,032 2,696

71 Clustering 1. Each pair of similar tags, as determined by a cooccurrence analysis (e.g., audio and mp3), is a seed constituting an initial cluster; 2. The cluster is enlarged by including tags that are similar to both the initial tags; 3. Repeat procedure recursively for all tags: each new candidate tag for a cluster must be similar to the whole (possibly enlarged) set of tags in that cluster. 4. If there are no more candidates for the cluster, go to step 1 with a new seed (e.g., audio and music).

72 Clustering audio semantic -web adult apple chat 1 mp3 rdf girls mac aim 2 music ontology nude macintosh messenger 3 playlist owl babes tiger gtalk 4 streaming semweb pics osx msn 5 radio daml sex macosx icq? Fruit

73 Combining Clusters Smoothing heuristics are applied to avoid having a number of very similar clusters originated from distinct seeds that are similar amongst each other. For every two clusters: 1. If one cluster contains the other, i.e., if the larger cluster contains all the tags of the smaller, remove the smaller cluster; 2. If clusters differ within a small margin, i.e., the number of different tags in the smaller cluster represents less than a percentage of the number of tags in the smaller and larger clusters, add the distinct words from the smaller to the larger cluster and remove the smaller.

74 Extracting relations For each pair of tags for which the search engine retrieved information, investigate the possible relationships: 1. A tag can be an ancestor of the other. For example, in the FOOD ontology, apple is a subclass of fruit. 2. A tag is the range or the value of a property of another tag. E.g., Class Zinfandel has a property hascolor, with value red 3. Both tags have the same direct parent: apple and pear are subclasses of fruit 4. Both tags have the same ancestors: assembly has as ancestors building (1 st level) and construction (2 nd ), while formation has fabrication (1 st ) and construction (2 nd ) in WordNet.

75 Examples Cluster_1: {admin application archive collection component control developer dom example form innovation interface layout planning program repository resource sourcecode} Information Object archive has-mention-of event in-event resource participant creator developer application participatesin innovation activity planning typerange example user admin component interface

76 Examples Cluster_2: {college commerce corporate course education high instructing learn learning lms school student} activities 4 education learning 4 teaching 4 training 1,4 qualification school 2 corporate 1 institution postsecondary School 2 student 3 takescourse studiesat course 3 university 2,3 college 2 offerscourse

77 Faceted Ontology Ontology creation and maintenance is automated Ontology evolution is driven by task features and by user changes Large scale integration of ontology elements from massively distributed online ontologies Very different from traditional top-downdesigned ontologies

78 Lessons Learnt Approach proven to be feasible and promising. However Assumptions in initial experiments (e.g., single ontology coverage for pairs of tags; focus on classes, clusteringbased approach, etc..) too restrictive Swoogle is too limited to support a fully automated approach we are now using Watson for the current experiments Integration with SW-enabled ontology matching algorithm is essential to improve term matching

79 Example #3: Semantics-Enhanced Web Browsing

80

81 Magpie Architecture Web Page Ontology cache (Lexicon) Enriched Web Page Magpie Hub Problem Domain & Resources Jabber Server Ontology based Proxy Server (found-item localhost #u" e/motta/" john-domingue john-domingue) (found-item localhost Semantic Log

82

83

84

85

86 PowerMagpie Architecture

87

88

89

90

91 Example #4: Question Answering on the Semantic Web

92 Aqualog: QA for Corporate Semantic Webs Which KMi researchers work on the Semantic Web? <akt:person rdf:about="akt:peterscott"> <rdfs:label>peter Scott</rdfs:label> <akt:hasaffiliation rdf:resource="akt:theopenuniversity"/> <akt:hasjobtitle>kmi deputy director</akt:hasjobtitle> <akt:worksinunit rdf:resource="akt:knowledgemediainstitute"/> <akt:hasgivenname>peter</akt:hasgivenname> <akt:hasfamilyname>scott</akt:hasfamilyname> <akt:hasprettyname>peter Scott</akt:hasPrettyName> <akt:haspostaladdress rdf:resource="akt:kmipostaladdress"/> <akt:hashomepage rdf:resource=" </akt:person> Answer

93 An Ontology-Modular System Which premiership footballers have played for Leeds and Chelsea? <ftb:footballer rdf:about= ftb:waynerooney"> <rdfs:label>wayne Rooney</rdfs:label> <ftb:playsfor ftb:manunited> <ftb:hasposition ftb:forward> <ftb:haspreviousclub ftb:everton> </ftb:footballer> <ftb:footballer rdf:about= ftb:davidbeckham"> <rdfs:label>david Beckham</rdfs:label> <ftb:playsfor ftb:realmadrid> <ftb:hasposition ftb:rightmidfield> <ftb:haspreviousclub ftb:manunited> </ftb:footballer> AquaLog Answer

94 Coarse-grained Architecture LINGUISTIC & QUERY CLASSIFICATION RELATION SIMILARITY SERVICE INFERENCE ENGINE NL SENTENCE INPUT ANSWER QUERY TRIPLES ONTOLOGY COMPATIBLE TRIPLES Linguistic Component obtains intermediate representation from the input query Relation Similarity Service maps the intermediate representation to the ontology/kb

95 Relation Similarity Service Which are the KMi researchers in the semantic web area? Translated query KMi researchers (person/organization, semantic web area) Ontological structures Has-research-interest (kmi-research-staff-member, Semantic-web-area) MECHANISMS: Ontology relationships and taxonomy, String algorithms, WordNet, Learning Mechanism, User s feedback

96 Learning Mechanism Which academics work in Akt? academic project User Lexicon Ontology Concepts Which academics work in Akt? project User Disambiguation Has-project-member academic Mapping work (inverse-of) has-project-member

97 Learning Mechanism - Key : the CONTEXT of a mapping is given by its particular place within the ontology, namely, since we work with triples, by the two arguments that the relation connects. work academic Has-project-member project - How can we generalize what we have learned? we take the highest concepts in the ontology, which can handle the same relation work person Has-project-member project Database academic

98 Interpretation Mechanisms: Ontology structure, String algorithms, WordNet, Machine Learning, User s feedback

99 User Feedback

100

101 AquaLog --> PowerAqua NL Query <akt:person rdf:about="akt:peterscott"> <rdfs:label>peter Scott</rdfs:label> <akt:hasaffiliation rdf:resource="akt:theopenuniversity"/> </akt:person> <ftb:footballer rdf:about= ftb:waynerooney"> <rdfs:label>wayne Rooney</rdfs:label> <ftb:playsfor ftb:manunited> <ftb:hasposition ftb:forward> <ftb:haspreviousclub ftb:everton> </ftb:footballer> PowerAqua <ptb:builder rdf:about= ptb:bob> <rdfs:label>bob the Builder</rdfs:label> <ptb:playedby > </ptb:builder> Answer

102 PowerAqua vs AquaLog Challenges when consulting and aggregating (dynamically mapping) information derived from multiple heterogeneous ontologies: Locating the right ontologies Intra-ontology semantic relevance analysis Filtering the right mappings Intg. heterogeneous information to provide an answer This reduces to deciding whether two instances specified according to different ontologies denote the same entity

103

104 Conclusions SW provides an unprecedented opportunity to build a new generation of intelligent systems, able to exploit large scale background knowledge The large scale background knowledge provided by the SW may address one of the fundamental premises (and holy grails) of AI The SW is not an aspiration: it is a concrete technology that is already in place today and is steadily becoming larger and more robust The new class of systems enabled by the SW is fundamentally different in many respects both from traditional KBS and even from early SW applications The examples shown in this talk provide an initial taste of the new generation of applications which will be made possible by the emerging Semantic Web

105 References Ontology Mapping Lopez, V., Sabou, M., Motta, E. (2006). "Mapping the real semantic web on the fly". ISWC 2006 Sabou, M., D'Aquin, M., Motta, E. (2006). "Using the semantic web as background knowledge for ontology mapping". ISWC 2006 Workshop on Ontology Mapping. Integration of Web2.0 and Semantic Web L.Specia, E. Motta, "Integrating Folksonomies with the Semantic Web", ESWC Angeletou, S., Sabou, M., Specia, L., and Motta, E., (2007). Bridging the Gap Between Folksonomies and the Semantic Web: An Experience Report. ESWC 2007 Workshop on Bridging the Gap between Semantic Web and Web 2.0. Watson d Aquin, M., Sabou, M., Dzbor, M., Baldassarre, C., Gridinoc, L., Angeletou, S. and Motta, E.: "WATSON: A Gateway for the Semantic Web". Poster Session at ESWC 2007

106 'Vision' Papers Motta, E., Sabou, M. (2006). "Next Generation Semantic Web Applications". 1st Asian Semantic Web Conference, Beijing. Motta, E., Sabou, M. (2006). "Language Technologies and the Evolution of the Semantic Web". LREC 2006, Genoa, Italy. Motta, E. (2006). "Knowledge Publishing and Access on the Semantic Web: A Socio-Technological Analysis". IEEE Intelligent Systems, Vol.21, 3, (88-90).

107