The NoRDF Project Fabian Suchanek Amazing! This talk is free of - PowerPoint PPT Presentation

The NoRDF Project Fabian Suchanek Amazing! This talk is free of the Corona virus! (about the speaker, we don’t know...)

Knowledge Bases P erson subclassOf Singer type born 1935 For us, a knowledge base (KB) is a graph, where the nodes are en ti ties and the edges are relations. (We do not distinguish T -Box and A-Box.) 2

Cool knowledge‐based applications Discovered 6 kineasis How long was the proteins that relate When was Thirty Y ears’ War? to cancer Elvis born? “1935” Apple Siri Amazon Echo IBM Watson These applications feed from knowledge bases. 3

There are plen ty of knowledge bases TextRunner NELL Plus industrial projects at Sponsored message: New version of YAGO at h ttp://yago-knowledge.org .

What’s in a knowledge base? Essen tially binary facts (“triples”) in the knowledge format “RDF”: 5 From YAGO

What’s in the real world? In February 1998, Andrew Wake field published a paper in the medical journal The Lance t, which reported on twelve children wi th developmen tal disorders. The paren ts were said to have linked the start of behavioral symptoms to vaccination. The resul ting con troversy became the biggest science story of 2002. As a resul t, vaccination rates dropped sharply. In 2011, the BMJ de tailed how Wake field had faked some of the data behind the 1998 Lance t article. Even ts Stories Belie fs Reasons Claims Falsifications ...none of which is in a knowledge base! 6

The NoRDF Project: Go Beyond Triples If we wan t tomorrow’s in telligen t applications to be really in telligen t, we have to extend their knowledge bases by Even ts Stories Belie fs Reasons Claims Falsifications 1) We have to be able to extract complex knowledge from text (“IE”) 2) We have to be able to represen t such knowledge and to reason on i t 7

IE: What is possible already Several cool approaches can extract non‐binary in formation: - FRED - StuffIE - K-P arser - OpenIE - Documen t spanners - HighLife - ClausIE - Classical slot fillers published in The Lance t in 1998. Andrew Wake field venue author time Publication_even t 8

IE: What we need “Wake field published a paper that reported on children. Their paren ts were said to have linked the start of behavioral symptoms to vaccination. The resul ting con troversy caused vaccination rates to fall. ...” caused Publication RateChange author direction about pub. con ten t of paper vaccinationRate - Wake field Claim children by of of Link paren ts of of symptoms vaccination 9

IE: What we need Cross‐sen tence analysis, advanced co‐re ference resolution, standardized types of frames, relationships be tween even ts, negation, hypothe tical stances, storylines, ... caused Publication RateChange author direction about pub. con ten t of paper vaccinationRate - Wake field Claim children by of of Y ou know a system Link paren ts that can do (part of) i t? of of Please le t me know! symptoms vaccination T ype here: _ _ _ _ _ _ _ _ _ _ _ _ 10

Reasoning: What we have As knowledge represen tation: - Frames, JSON - complex objects - object-relational databases caused Publication RateChange author direction pub. of paper vaccinationRate - Wake field 11

Reasoning: What we have As knowledge represen tation: For reasoning: - Frames, JSON - RDFS, OWL DL, SHACL - complex objects - Description Logic - object-relational databases - Con text logics - Fact iden tifiers - Modal logics - RDF* - Epistemic logics - Reification - Formal argumen tation - Belie f revision - Provenance and annotated logics Cannot represen t - “All clien ts believe that the company delivers a good service” - “the loss of value on the stock marke t happened because the public learned of a fraudulen t activi ty by the company” - “Mary believes everything P aul says, P aul says Mary believes ” ... or if they can, they are undecidable 12

Reasoning: What we need 1) a very simple logic inside a con text First‐order logic wi thout ? (?) OWL EL? Datalog? 2) a very simple logic about con texts Horn Rules? Datalog? (?) => a moderately simple logic in combination Y ou have a great idea? Le t me know! V agueness, fuzziness, and probabili ty: orthogonal topics 13

Applications • Analysis of fake news / fact checking: understand an article about a con troversial topic, allow reasoning (who said what when and why, what is the evidence, ...) • Analysis of the e-reputation of a company: extract con troversy or belie fs wi th reasons and supporters, for companies or their products • Modeling of con troversies: de tect a con troversial topic on the Web (in blogs, forums, T wi tter), extract opinions, and model differen t views Understanding the argumen ts of the other side is a prerequisi te for re futing them. >more 14

Applications • Flagging of poten tially fraudulen t activi ty: De tect claims that con tradict knowledge, or violate rules. • Modeling of processes: Model sequences of actions, causal relationships, and suggestions. • Smarter chatbots: Allow dialogues that go beyond single-shot questions. • Legal text understanding: Analyze a law, a regulation, or a con tract, and derive what is permi tted and what is obligatory for which party. 15

Our project “NoRDF” Our project “NoRDF” aims to extract and model complex in formation from natural language text. We are supported by the French National Research Agency, T élécom P aris, and 4 sponsors: 16

Our project “NoRDF”: Who’s there? Fabian Suchanek Professor at T élécom P aris Knowledge Bases, Reasoning, NLP Chloé Clavel Professor at T élécom P aris Affective Computing, Sen timen t Analysis We hired • Pierre-Henri P aris (CNAM) as a postdoc • Chadi Helwe (American Univ. of Beirut) as PhD studen t • Sanaz Hasanzadeh (AUT T ehran) as PhD studen t 17

And we are still hiring! We are hiring PhD studen ts, postdocs, and engineers, for the project or anything that has to do wi th NLP , knowledge bases, and reasoning! Join our team! h ttps:// suchanek.name -> NoRDF 18

Backup Slides

Reasoning: What we have As knowledge represen tation: - Frames, JSON - complex objects - object-relational databases caused Publication RateChange author direction pub. of paper vaccinationRate - Wake field 20

Reasoning: What we have As knowledge represen tation: - Frames, JSON great, but do not allow for reasoning - complex objects - object-relational databases caused Publication RateChange author direction pub. of paper vaccinationRate - Wake field - “If X caused Y and Y caused Z, then X caused Z” - “If X did not publish a paper, X is not a scien tist” - “If Mary believes what P aul says & P aul says X, then Mary believes X” 21

Reasoning: What we have For reasoning: - RDFS, OWL DL, SHACL - Description Logic caused Publication RateChange author direction pub. of paper vaccinationRate - Wake field 22

Reasoning: What we have For reasoning: - RDFS, OWL DL, SHACL great, but do not allow for statemen ts - Description Logic about statemen ts caused Publication RateChange author direction pub. of paper vaccinationRate - Wake field - “The paper says that vaccines cause autism” - “Fact A caused Fact B” 23

Reasoning: What we have Annotated Knowledge Represen tations: - Fact iden tifiers - RDF* - Reification caused Publication RateChange author direction pub. of paper vaccinationRate - Wake field 24

Reasoning: What we have Annotated Knowledge Represen tations: - Fact iden tifiers cannot deal wi th hypothe tical statemen ts - RDF* cannot do reasoning - Reification caused Publication RateChange author direction pub. of paper vaccinationRate - Wake field - “Mary believes that vaccines cause autism” 25

Reasoning: What we have Big logic machinery: - Con text logics - Modal logics - Epistemic logics 26

Reasoning: What we have Big logic machinery: - Con text logics - Modal logics cannot quan tify over con texts (or if they can, they are proposi tional logics or undecidable) - Epistemic logics - “All clien ts believe that the company delivers a good service” - “the loss of value on the stock marke t happened because the public learned of a fraudulen t activi ty by the company” Formal argumen tation has monoli thic proposi tions. Belie f revision has monoli thic agen ts. Provenance and annotated logics cannot make claims about annotations. V agueness, fuzziness, and probabili ty are orthogonal topics. 27