Ce diaporama a bien été signalé.
Nous utilisons votre profil LinkedIn et vos données d’activité pour vous proposer des publicités personnalisées et pertinentes. Vous pouvez changer vos préférences de publicités à tout moment.

The Web We Mix - benevolent AIs for a resilient web

349 vues

Publié le

Keynote of @fabien_gandon at Web Science 2019 Boston #WebSci19

Abstract is here: https://hal.inria.fr/hal-02148911

Publié dans : Sciences
  • Soyez le premier à commenter

The Web We Mix - benevolent AIs for a resilient web

  1. 1. The Web We Mix benevolent AIs for a resilient web Fabien Gandon, http://fabien.info I'm sorry, Dave. I'm afraid I can't do that.
  2. 2. WIMMICS TEAM  Inria  CNRS  University Côte D’Azur (UCA) I3S Web-Instrumented Man-Machine Interactions, Communities and Semantics
  3. 3. CHALLENGE to bridge social semantics and formal semantics on the Web
  4. 4. MULTI-DISCIPLINARY TEAM  40~50 members  ~15 nationalities  1 DR, 3 Professors  3CR, 4 Assistant professors DR/Professors:  Fabien GANDON, Inria, AI, KRR, Semantic Web, Social Web  Nhan LE THANH, UCA, Logics, KR, Emotions, Workflows  Peter SANDER, UCA, Web, Emotions  Andrea TETTAMANZI, UCA, AI, Logics, Agents, CR/Assistant Professors:  Michel BUFFA, UCA, Web, Social Media  Elena CABRIO, UCA, NLP, KR, Linguistics, Q&A, Text Mining  Olivier CORBY, Inria, KR, AI, Sem. Web, Programming, Graphs  Catherine FARON-ZUCKER, UCA, KR, AI, Semantic Web, Graphs  Alain GIBOIN, Inria, Interaction Design, KE, User & Task models  Isabelle MIRBEL, UCA, Requirements, Communities  Serena VILLATA, CNRS, AI, Argumentation, Licenses, Rights Research engineer: Franck MICHEL, CNRS, Linked Data, Integration, DB External:  Andrei Ciortea (University of St. Gallen) Agents, WoT, Sem. Web  Nicolas DELAFORGE (Mnemotix) Sem. Web, KM, Integration  Claude FRASSON (University of Montreal) Emotion, eLearning  Freddy LECUE (Thales, Montreal) AI, Logics, Mining, Big Data, Sem. Web
  5. 5. AI for classical Web tasks
  6. 6. Autonomy AgentWare Web Researcher, 1996 Dogs’ UI metaphor to Machine Learning Sent to Web kennels to work offline
  7. 7. SEARCHING  exploratory search  question-answering DISCOVERYHUB.CO semantic spreading activation new evaluation protocol [Marie, Giboin, Palagi et al.]
  8. 8. SEARCHING  exploratory search  question-answering DISCOVERYHUB.CO semantic spreading activation new evaluation protocol [D:Work], played by [R:Person] [D:Work] stars [R:Person] [D:Work] film stars [R:Person] starring(Work, Person) linguistic relational pattern extraction named entity recognition similarity based SPARQL generation select * where { dbpr:Batman_Begins dbp:starring ?v . OPTIONAL {?v rdfs:label ?l filter(lang(?l)="en")} } [Cabrio et al.] [Marie, Giboin, Palagi et al.] QAKiS.ORG
  9. 9. BROWSING e.g. SMILK plugin [Lopez, Cabrio, et al.]
  10. 10. BROWSING e.g. SMILK plugin [Nooralahzadeh, Cabrio, et al.]
  11. 11. linked open data(sets) cloud on the Web 0 200 400 600 800 1000 1200 1400 01/05/2007 08/10/2007 07/11/2007 10/11/2007 28/02/2008 31/03/2008 18/09/2008 05/03/2009 27/03/2009 14/07/2009 22/09/2010 19/09/2011 30/08/2014 26/01/2017 number of linked open datasets on the Web
  12. 12. CRAWLING  Predict data availability  Select features of URIs [Huang, Gandon 2019]
  13. 13. CRAWLING  Predict data availability  Select features of URIs  Learn crawling selection  Online learning w. crawling [Huang, Gandon 2019]
  14. 14. EDTECH & WEB for e-learning & serious games [Rodriguez-Rocha, Faron-Zucker et al.]
  15. 15. QUIZZES Automated generation of quizzes [Rodriguez-Rocha, Faron-Zucker et al.] rdf:type p rdfs:subClassOf ? ? ? ? RDF Q&A Generation NL Questions Generation Ranking Question SELECTION Knowledge Graph QuizQ&A
  16. 16. but…
  17. 17. over adaptation & filter bubble(c.f. Eli Pariser)
  18. 18. AITOBURSTBUBBLES
  19. 19. bias……
  20. 20. UNCERTAINTY Can uncertainty be formalized on top of the standards of the Semantic Web, to be published on the Web ? [Djebri et al 2019]
  21. 21. UNCERTAINTY Can uncertainty be formalized on top of the standards of the Semantic Web, to be published on the Web ? [Djebri et al 2019] prob:Probability a munc:UncertaintyApproach; munc:hasUncertaintyFeature prob:probabilityValue; munc:hasUncertaintyOperator prob:and. prob:probabilityValue prob:and prob:multiplyProbability. prob:Probability prob:probabilityValue prob:and ex:multiplyProbability munc:hasUncertainty Feature munc:hasUncertainty Operator
  22. 22. UNCERTAINTY LINKED CALCULI [Djebri et al 2019] function prob:multiplyProbability(?s1, ?s2, ?c) { let(?v1 = munc:getMeta(?s1, prob:Probability)){ if(prob:verifyIndependent(?s1, ?s2) == true) ?v2 = munc:getMeta(?s2, prob:Probability, xt:list(?s1, ?c)) return (?v1 * ?v2) } else { ?v2 = munc:getMeta(?s2, prob:Probability) return (?v1 * ?v2) } } } prob:Probability a munc:UncertaintyApproach; munc:hasUncertaintyFeature prob:probabilityValue; munc:hasUncertaintyOperator prob:and. prob:probabilityValue prob:and prob:multiplyProbability. prob:Probability prob:probabilityValue prob:and ex:multiplyProbability munc:hasUncertainty Feature munc:hasUncertainty Operator
  23. 23. UNCERTAINTY Can uncertainty be translated and negotiated? [Djebri et al 2019]
  24. 24. UNCERTAINTY Can uncertainty be translated and negotiated? [Djebri et al 2019] • Specify uncertainty in parameter linked to the format • GET /some/resource HTTP/1.1 Accept: text/turtle;uncertainty="http://example.com/Probability";q=0.8, text/turtle;uncertainty="http://example.com/Possibility";q=0.2; • Use uncertainty as a profile : prof-Conneg • GET /some/resource HTTP/1.1 Accept: text/turtle;q=0.8;profile="prob:Probability", text/turtle;q=0.2;profile="poss:Possibility" • HEAD /some/resource HTTP/1.1 Accept: text/turtle;q=0.9,application/rdf+xml;q=0.5 Link: <http://example.com/Probability>; rel="profile" (RFC 6906) • GET /some/resource HTTP/1.1 Accept: text/turtle Prefer: profile="prob:Probability" (RFC 7240)
  25. 25. AI for classical Web tasks Web for classical AI tasks
  26. 26. QUERY & INFER  graph rules and queries  deontic reasoning  induction CORESE  & G2 H2  & G1 H1 < Gn Hn abstract graph machine STTL [Corby, Faron-Zucker et al.]
  27. 27. QUERY & INFER  graph rules and queries  deontic reasoning  induction CORESE  & G2 H2  & G1 H1 < Gn Hn RATIO4TA predict & explain abstract graph machine STTL [Hasan et al.] [Corby, Faron-Zucker et al.]
  28. 28. QUERY & INFER  graph rules and queries  deontic reasoning  induction CORESE INDUCTION  & G2 H2  & G1 H1 < Gn Hn RATIO4TA predict & explain find missing knowledge abstract graph machine STTL [Hasan et al.] [Tettamanzietal.] [Corby, Faron-Zucker et al.]
  29. 29. QUERY & INFER  graph rules and queries  deontic reasoning  induction CORESE LICENTIA INDUCTION  & G2 H2  & G1 H1 < Gn Hn RATIO4TA predict & explain find missing knowledge deontic reasoning, license compatibility and composition abstract graph machine STTL [Hasan et al.] [Tettamanzietal.] [Villata et al.] [Corby, Faron-Zucker et al.]
  30. 30. the Web: this place where invisible brontobytes graze furiously 1027 bytes
  31. 31. Smarter Cities – IBM Dublin [Lécué, 2015]
  32. 32. Smarter Cities – IBM Dublin [Lécué, 2015]
  33. 33. PREDICT HOSPITALIZATION  Physician’s records classification in order to predict hospitalization [Gazzotti, Faron et al. 2017] Sexe Date Cause CISP2 ... History Observations H 25/04/2012 vaccin-antitétanique A44 ... Appendicite EN CP - Bon état général - auscult pulm libre; bdc rég sans souffle - tympans ok- Element Number Patients Consultations Past medical history Biometric data Semiotics Diagnosis Row of prescribed drugs Symptoms Health care procedures Additional examination Paramedical prescription Observations/notes 55 823 364 684 187 290 293 908 250 669 117 442 847 422 23 488 11 850 871 590 17 222 56 143
  34. 34. PREDICT HOSPITALIZATION  Physician’s records classification in order to predict hospitalization  Augment data with structured knowledge [Gazzotti, Faron et al. 2017] Sexe Date Cause CISP2 ... History Observations H 25/04/2012 vaccin-antitétanique A44 ... Appendicite EN CP - Bon état général - auscult pulm libre; bdc rég sans souffle - tympans ok- Element Number Patients Consultations Past medical history Biometric data Semiotics Diagnosis Row of prescribed drugs Symptoms Health care procedures Additional examination Paramedical prescription Observations/notes 55 823 364 684 187 290 293 908 250 669 117 442 847 422 23 488 11 850 871 590 17 222 56 143 (1)
  35. 35. PREDICT HOSPITALIZATION  Physician’s records classification in order to predict hospitalization  Augment data with structured knowledge and study impact on different prediction methods  Study enrichment and features impact and combinations on different ML methods [Gazzotti, Faron et al. 2017] Sexe Date Cause CISP2 ... History Observations H 25/04/2012 vaccin-antitétanique A44 ... Appendicite EN CP - Bon état général - auscult pulm libre; bdc rég sans souffle - tympans ok- Element Number Patients Consultations Past medical history Biometric data Semiotics Diagnosis Row of prescribed drugs Symptoms Health care procedures Additional examination Paramedical prescription Observations/notes 55 823 364 684 187 290 293 908 250 669 117 442 847 422 23 488 11 850 871 590 17 222 56 143 (1) (2)
  36. 36. MonaLIA  reason & query on RDF metadata to build balanced, unambiguous, labelled training sets. 350 000 images of artworks RDF metadata based on external thesauri Joconde database from French museums [Bobasheva et al. 2017]
  37. 37. MonaLIA  reason & query on RDF metadata to build balanced, unambiguous, labelled training sets.  transfer learning & CNN classifiers on targeted categories (topics, techniques, etc.) 350 000 images of artworks RDF metadata based on external thesauri Joconde database from French museums (1) [Bobasheva et al. 2017]
  38. 38. MonaLIA  reason & query on RDF metadata to build balanced, unambiguous, labelled training sets.  transfer learning & CNN classifiers on targeted categories (topics, techniques, etc.)  reason & query RDF metadata of results to address silence, noise and explain 350 000 images of artworks RDF metadata based on external thesauri Joconde database from French museums (1) (2) [Bobasheva et al. 2017] animal  bird ? painting 
  39. 39. deduce data  model, schemas, ontologies, ... data data 15% progress
  40. 40. learn data embeddings, parameters, configurations, … data data 30% progress
  41. 41. sum intelligence  model, schemas, ontologies, ... embeddings, parameters, configurations, … data data 45% progress
  42. 42. combine intelligence model, schemas, ontologies, ... embeddings, parameters, configurations, … data  60% progress
  43. 43. remotely combine model, schemas, ontologies, … embeddings, parameters, configurations,…  Web 75% progress
  44. 44. deeply combine data, knowledge, model, schemas, ontologies, … data, knowledge, embeddings, parameters, configurations,…  Web 90% progress
  45. 45. combining AIs on the Web data, knowledge, model, schemas, ontologies, … data, knowledge, embeddings, parameters, configurations,…  Web 100% progress
  46. 46. WEB EDGE AI  Edge AI directly in the browser  Web APIs, models, protocols,… [WebML @ W3C]
  47. 47. Web ways applied to AI e.g. copy-paste based reuse and contribution to create Web AIs
  48. 48. but…
  49. 49. automated deduction all birds can fly all penguins are birds so ...
  50. 50. PIPE : 0.9143 automated classification
  51. 51. PROBABILITY Assertions (evidence) Axiom (hypothesis) how to evaluate in an open-world? ?
  52. 52. PROBABILITY POSSIBILITY  Possibilistic Axiom Scoring  Possibility and Necessity Measures [Tettamanzi et al.] Assertions (evidence) Axiom (hypothesis) how to evaluate in an open-world? –1 +1 REJECT ACCEPT ?
  53. 53. correlation
  54. 54. correlation vs. causation http://tylervigen.com/spurious-correlations
  55. 55. AI for classical Web tasks Web for classical AI tasks AI and Web augmentation
  56. 56. “ « a Web-Augmented Interaction (WAI) is a user’s interaction with a system that is improved by allowing the system to access Web resources » [Gandon, Giboin, WebSci17]
  57. 57. AZKAR remotely visit and interact with a museum through a robot and via the Web [Buffa et al.]
  58. 58. ALOOF: Web and Perception [Cabrio, Basile et al.] Semantic Web-Mining and Deep Vision for Lifelong Object Discovery (ICRA 2017) Making Sense of Indoor Spaces using Semantic Web Mining and Situated Robot Perception (AnSWeR 2017)
  59. 59. ALOOF: robots learning by reading on the Web First Object Relation Knowledge Base: 46.212 co-mentions gave 49 tools, 14 rooms, 101 “possible location” relations, Annie cuts the bread in the kitchen with her knife dbp:Knife aloof:Location dbp:Kitchen [Cabrio, Basile et al.]
  60. 60. ALOOF: RDF dataset about objects [Cabrio, Basile et al.]  common sense knowledge about objects: classification, prototypical locations and actions  knowledge extracted from natural language parsing, crowdsourcing, distributional semantics, keyword linking, ...
  61. 61. WASABI Web-augmented music interactions [Buffa et al.]
  62. 62. WASABI Web-augmented music interactions [Buffa et al.] [Fell et al.]
  63. 63. but…
  64. 64. From Eliza to Lenny, talking to machines
  65. 65. complex web applications barely visible evolution of the place of humans in = user = processor = data
  66. 66. You will read this first Then you will read this… Then this… and only now are you reading this top left text… predictable means manipulable
  67. 67. AI for Artificial Intelligence (McCarthy et al., 1955)
  68. 68. AI for Artificial Intelligence (McCarthy et al., 1955) IA for Intelligence Amplification (Ashby, 1956) and Intelligence Augmentation (Engelbart, 1962)
  69. 69. AI for Artificial Intelligence (McCarthy et al., 1955) IA for Intelligence Amplification (Ashby, 1956) and Intelligence Augmentation (Engelbart, 1962) Web as the rendez-vous point for two fields born in the 50s… AI & IA
  70. 70. AI for classical Web tasks Web for classical AI tasks AI and Web augmentation AI and Web societies
  71. 71. MODELING USERS  individual context  social structures PRISSMA prissma:Context 0 48.86034 -2.337599 200 geo:lat geo:lon prissma:radius 1 :museumGeo prissma:Environment 2 { 3, 1, 2, { pr i ssma: poi } } { 4, 0, 3, { pr i ssma: envi r onment } } :atTheMuseum error tolerant graph edit distance context ontology [Costabello et al.]
  72. 72. MODELING USERS  individual context  social structures PRISSMA prissma:Context 0 48.86034 -2.337599 200 geo:lat geo:lon prissma:radius 1 :museumGeo prissma:Environment 2 { 3, 1, 2, { pr i ssma: poi } } { 4, 0, 3, { pr i ssma: envi r onment } } :atTheMuseum error tolerant graph edit distance context ontology OCKTOPUS tag, topic, user distribution tag and folksonomy restructuring with prefix trees [Costabello et al.] [Meng et al.]
  73. 73. MODELING USERS  individual context  social structures PRISSMA prissma:Context 0 48.86034 -2.337599 200 geo:lat geo:lon prissma:radius 1 :museumGeo prissma:Environment 2 { 3, 1, 2, { pr i ssma: poi } } { 4, 0, 3, { pr i ssma: envi r onment } } :atTheMuseum error tolerant graph edit distance context ontology OCKTOPUS tag, topic, user distribution tag and folksonomy restructuring with prefix trees EMOCA&SEEMPAD emotion detection & annotation [Villata, Cabrio et al.] [Costabello et al.] [Meng et al.]
  74. 74. DEBATES & EMOTIONS #IRC [Villata, Cabrio et al.]
  75. 75. DEBATES & EMOTIONS #IRC argument rejection attacks-disgust[Villata, Cabrio et al.]
  76. 76. OPINIONS NLP, ML and arguments to monitor online image [Villata, Cabrio, et al.]
  77. 77. ARGUMENT MINING ON CLINICAL TRIALS  NLP, ML and arguments  assist evidence-based medicine  support doctors and clinicians  identify doc. for certain disease  analyze argumentative content and PICO elements [Mayer, Cabrio, Villata]
  78. 78. ARGUMENT MINING ON POLITICAL SPEECHES  NLP and Machine Learning.  Support historians/social science scholars  Analyze arguments in political speeches  DISPUTool : 39 political debates, last 50 years of US presidential campaigns (1960-2016) [Mayer, Cabrio, Villata]
  79. 79. CYBERBULLYING CREEP EU project: detect and prevent [Corazza, Arslan, Cabrio, Villata]
  80. 80. CYBERBULLYING CREEP EU project: detect and prevent [Corazza, Arslan, Cabrio, Villata]
  81. 81. but…
  82. 82. China: 1 400 millions India: 1 300 millions acebook: 2 400 millions
  83. 83. inverse view of the world China: 1 400 millions India: 1 300 millions
  84. 84. Web Robots
  85. 85. e.g. Wikipedia bots 2 187 bot approved for use on the English Wikipedia to help maintain 45 223 137 pages
  86. 86. web… page social apps data … a.i.
  87. 87. web agents =+
  88. 88. EDITS 0 500000 1000000 1500000 2000000 2500000 3000000 3500000 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 Wikipedia editors, 2012 0 500000 1000000 1500000 2000000 2500000 3000000 3500000 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 Humans
  89. 89. massive interaction design ergonomics in hybrid web communities
  90. 90. tardigrades Web AIs • 1 ° K / −458 °F / −272 °C  • 420 °K / 300 °F / 150 °C  • Vacuum / High pressure  • Ionizing radiation  • Himalayas / Deep sea  Extremophile
  91. 91. perspectives
  92. 92. Web 1.0, …
  93. 93. Web 1.0, 2.0…
  94. 94. price convert? person other sellers? Web 1.0, 2.0, 3.0 …
  95. 95. [TimBL, 94]
  96. 96. Toward a Web of Programs “We have the potential for every HTML document to be a computer — and for it to be programmable. Because the thing about a Turing complete computer is that … anything you can imagine doing, you should be able to program.” (Tim Berners-Lee, 2015)
  97. 97. Toward a Web of Things
  98. 98. the Web as a global blackboard for artificial intelligence
  99. 99. Connected Animals, Animal-computer interaction (ACI) Herdsourcing: monitoring collective animal behavior
  100. 100. Consider a Web of intelligence linking  Web of Animals: connected animals to predict earthquakes  Web of Things, Web of Sensors  Web of AIs: reasoning systems, learning systems, etc.  Web of People: experts from all over the world (i.e. some expert always awake // crowdsourcing & Q&A expert routing) … to form a collective intelligent decision system.
  101. 101. the Web as a global blackboard for artificial intelligence all kinds of
  102. 102. IMAG_NE
  103. 103. IMAG_NE a Web linking all forms of Intelligence
  104. 104. a Web Science research agenda must account for the fact that the long term potential of the Web is to augment and link everything. Peter Steiner, The New Yorker, 1993 Kaamran Hafeez, The New Yorker, 2015 I pass CAPTCHAs, Nobody knows I’m a bot 2019
  105. 105. but…
  106. 106. training common NLP models nearly five times the lifetime emissions of the average American car including the manufacturing of the car itself
  107. 107. Business Insider, US National Archives, George Grantham Bain Collection 1942 ELECTRIC CAR PARIS 70km/h, 100km autonomy 1943 SWITCHBOARD AND LOST JOBS we need historians, sociologists,…
  108. 108. www mmm world wide web massively multidisciplinary method
  109. 109. - Look, we should stay away for a while; you and I are so over- informed that we come out of every conversation exhausted. Sempé
  110. 110. Make the Web AI-friendly content, links, metadata, etc. data, knowledge, etc. AI Web bots: chat bots, recommenders, facilitators, etc. configuration, parameters, embeddings, services, communication, etc.

×