The Need to Move beyond Triples Fabian Suchanek (Vision paper) Amazing! This talk is free of the Corona virus! (about the speaker, we don’t know...)
Cool knowledge‐based applications Apple Siri 2 When was Elvis born? “1935” IBM Watson Discovered 6 kineasis proteins that relate to cancer How long was the Thirty Years’ War? Amazon Echo These applications feed from  knowledge bases .
There are plenty of knowledge bases NELL TextRunner Plus industrial projects at Sponsored message: New version of YAGO at  http://yago-knowledge.org .
What’s in a knowledge base? From YAGO Essentially binary facts (“triples”) in the knowledge format “RDF”: 4
What’s in the real world? In February 1998, Andrew Wakefield published a paper in the medical journal The Lancet, which reported on twelve children with developmental disorders. The parents were said to have linked the start of behavioral symptoms to vaccination. The resulting controversy became the biggest science story of 2002. As a result, vaccination rates dropped sharply. In 2011, the BMJ detailed how Wakefield had faked some of the data behind the 1998 Lancet article. Beliefs Claims Events Reasons Stories Falsifications ...none of which is in a knowledge base! 5
The vision of this paper: “The Need to Move Beyond Triples” If we want tomorrow’s intelligent  applications to be really intelligent, we have to extend their knowledge bases by 6 1) We have to be able to extract complex knowledge from text (“IE”) 2) We have to be able to represent such knowledge and to reason on it Beliefs Claims Events Reasons Stories Falsifications
IE: What is possible already 7 Several cool approaches can extract non‐binary information: - FRED - K-Parser - Document spanners - ClausIE - StuffIE - OpenIE - HighLife - Classical slot fillers Andrew Wakefield published in The Lancet in     1998. Publication_event author venue time (>50% of the vision paper is discussion of related work)
8 Several cool approaches can extract non‐binary information: - FRED - K-Parser - Document spanners - ClausIE Andrew Wakefield published in The Lancet in     1998. Publication_event author venue time IE: What is possible already - StuffIE - OpenIE - HighLife - Classical slot fillers (>50% of the vision paper is discussion of related work)
IE: What we need 9 “Wakefield published a paper that reported on children. Their parents were said to have linked the start of behavioral symptoms to vaccination. The resulting controversy caused vaccination rates to fall. ...” Publication RateChange Wakefield paper Claim symptoms children vaccination Link parents vaccinationRate - caused of direction author pub. content about of by of of of
IE: What we need 10 Publication RateChange Wakefield paper Claim symptoms children vaccination Link parents vaccinationRate - caused of direction author pub. content about of by of of of You know a system that can do (part of) it? Please let me know! Type here: ____________ Cross‐sentence analysis, advanced co‐reference resolution, standardized types of frames, relationships between events, negation, hypothetical stances, storylines, ...
Reasoning: What we have 11 RateChange vaccinationRate - of direction As knowledge representation: - Frames, JSON - complex objects - object-relational databases Publication Wakefield paper caused author pub.
Reasoning: What we have 12 RateChange vaccinationRate - of direction As knowledge representation: - Frames, JSON - complex objects - object-relational databases Publication Wakefield paper caused author pub. great, but do not allow for reasoning - “If X caused Y and Y caused Z, then X caused Z” - “If X did not publish a paper, X is not a scientist” - “If Mary believes what Paul says & Paul says X, then Mary believes X”
Reasoning: What we have 13 RateChange vaccinationRate - of direction For reasoning: - RDFS, OWL DL, SHACL - Description Logic Publication Wakefield paper caused author pub.
Reasoning: What we have 14 RateChange vaccinationRate - of direction For reasoning: - RDFS, OWL DL, SHACL - Description Logic Publication Wakefield paper caused author pub. great, but do not allow for statements about statements - “The paper says that vaccines cause autism” - “Fact A caused Fact B”
Reasoning: What we have 15 RateChange vaccinationRate - direction Annotated Knowledge Representations: - Fact identifiers - RDF* - Reification of Publication Wakefield paper caused author pub.
Reasoning: What we have 16 RateChange vaccinationRate - direction Annotated Knowledge Representations: - Fact identifiers - RDF* - Reification of Publication Wakefield paper caused author pub. cannot deal with hypothetical statements cannot do reasoning - “Mary believes that vaccines cause autism”
Reasoning: What we have 17 Big logic machinery: - Context logics - Modal logics - Epistemic logics
Reasoning: What we have 18 Big logic machinery: - Context logics - Modal logics - Epistemic logics - “All clients believe that the company delivers a good service” - “the loss of value on the stock market happened because the      public learned of a fraudulent activity by the company” (or if they can, they are propositional logics or undecidable) cannot quantify over contexts Formal argumentation has monolithic propositions. Belief revision has monolithic agents. Provenance and annotated logics cannot make claims about annotations. Vagueness, fuzziness, and probability are orthogonal topics.
Reasoning: What we need 19 1) a very simple logic  inside  a context 2) a very simple logic  about  contexts => a moderately simple logic       in combination First‐order logic without   ? OWL EL? Datalog? Horn Rules? Datalog? You have a great idea? Let me know!   (?)                 (?)
Applications 20 •  Analysis of fake news / fact checking:      understand an article about a controversial topic, allow reasoning      (who said what when and why, what is the evidence, ...) •  Analysis of the e-reputation of a company:     extract controversy or beliefs with reasons and supporters,     for companies or their products •  Modeling of controversies:      detect a controversial topic on the Web (in blogs, forums, Twitter),      extract opinions, and model different views >more Understanding the arguments of the other side is a prerequisite for refuting them.
Applications 21 •  Flagging of potentially fraudulent activity:      Detect claims that contradict 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 contract, and derive     what is permitted and what is obligatory for which party.
Our project “NoRDF” 22 We are hiring - PhD students - postdocs - engineers https://suchanek.name/work/research/nordf/ Our project “NoRDF” aims to extract and model complex information from natural language text. We are supported by the French National Research Agency and 4 sponsors:
Backup Slides
Reasoning: What we have 24 RateChange vaccinationRate - direction Annotated Reasoning: - Provenance formalisms - Annotated logics of Publication Wakefield paper caused author pub.
Reasoning: What we have 25 RateChange vaccinationRate - direction Annotated Reasoning: - Provenance formalisms - Annotated logics cannot make statements about annotations of Publication Wakefield paper caused author pub. - “Mary believes that Fact A holds because of Fact B” - “Fact A precedes Fact B”