Explainable Systems: The Inference Web Approach Paulo Pinheiro da Silva Stanford University In collaboration with Deborah L. Mc. Guinness, Richard E. Fikes, Cynthia Chang, Priyendra Deshwal, Dhyanesh Narayanan, Alyssa Glass, Selene Makarios, Jessica Jenkins, Bill Millar, Eric Hsu and many people from IBM, SRI, ISI, IHMC, U. Toronto, U. Trento, U. Fortaleza, U. Texas Austin, Rutgers U. , Maryland U. , Batelle, SAIC, UCSF, MIT W 3 C
Overview 1. What are explainable systems and why should we care about them? 2. Inference Web: Enabling Explainable Systems 3. Explainable Systems in Action 4. Explainable Systems 10 years from now Paulo Pinheiro da Silva
Explanation Need I need to send Paulo a letter but I don’t know his address. I believe Paulo lives in the U. S. So, Stanford, CA, USA. appears to be a possible answer. Google-2. 0, where is Paulo’s office? Google-2. 0, why is Paulo’s address “Manchester, UK”? 1) S 2) tanfor 2) M CA, U d, S anc hes A UK ter, [Betty] Paulo Pinheiro da Silva
Explanation in Action Why should I believe this? Why should I believe these? OK, “Manchester, UK” was Paulo’s address in May, 2002 and we are in 2005 !! I’ll send his letter to Stanford. Paulo At Manchester, UK [Betty] transitivity of At Paulo At University of Manchester At Manchester, UK Source: http: //www. cs. man. ac. uk/~pinheirp Source usage: May/2002 Source: http: //www. cs. man. ac. uk Source usage: May/2002 Paulo Pinheiro da Silva
What are Explainable Systems? question answer expl. 1 explanation request 1 explanation request n question answer explanation request 1 expl. 1 … [Bob] answer … … expl. n answer understanding question explanation request n explanation n Paulo Pinheiro da Silva
Why should we care about explainable systems? p As system users, we often need: n n p To understand system’s response To trust system’s responses Many explanation concerns are the same as in early systems such as n n Shortliffe’s MYCIN [1976] Swartout’s XPLAIN [1983] Paulo Pinheiro da Silva
Why should we care about explainable systems even more now ? p Systems are far more complex than 30 years ago n n p p Hybrid and distributed processing, e. g. , web services, the Grid Large number of heterogeneous, distributed information sources, e. g. , the Web More variation in reliability of information sources, e. g. , information extraction Sophisticated information integration methods, e. g. , SIMS, TSIMMIS Now we have less understanding (and sometimes less trust) of system’s answers and behavior Now we have even more reasons for systems to explain their responses Paulo Pinheiro da Silva
How to Enable Explainable Systems? question answer explanation request 1 expl. 1 … 1 -> ((allof (the played-by of (the instances of Project-Leader)) where (It isa Person)) = (: set *Helen *Jody)) 2 -> (allof (the played-by of (the instances of Project-Leader)) where (It isa Person)) 3 -> (forall (the played-by of (the instances of Project-Leader)) where (It isa Person) It) 4 -> (the played-by of (the instances of Project-Leader)) 5 -> (the instances of Project-Leader) 5 (1) Local value(s): (: set *COGS-Proj. Leader-1 *HI-LITE-Project. Leader-1 *SKIPRProject. Leader-1) 6 -> (: set *COGS-Proj-Leader-1 *HI-LITEProject. Leader-1 *SKIPR-Project. Leader-1) [for (the instances of Project-Leader)] 6 <- (*COGS-Proj-Leader-1 *HI-LITEProject. Leader-1 *SKIPR-Project. Leader-1) [(: set. . . 5 (2) From inheritance: (: set *COGS-Proj -Leader-1 *HI-LITE-Project. Leader-1 *SKIPRProject. Leader-1) explanation request n expl. n Which information do I have to generate an explanation? I may have (or may be able to record) data describing how I manipulate information to produce answers! Paulo Pinheiro da Silva
Explainable System Challenge Explanation Understanding Trust The GAP Information Manipulation Data Paulo Pinheiro da Silva
Overview 1. What are explainable systems and why should we care about them? 2. Inference Web: Enabling Explainable Systems 3. Explainable Systems in Action 4. Explainable Systems 10 years from now Paulo Pinheiro da Silva
Requirements for Explainable Systems p Information Manipulation Traces hybrid, distributed, portable, shareable, combinable encoding of proof fragments supporting multiple justifications n p Presentation n p multiple display formats supporting browsing, visualization, etc. Abstraction n understandable summaries p Interaction n multi-modal mixed initiative options including natural-language and GUI dialogues, adaptive, context-sensitive interaction p Trust n p source and reasoning provenance, automated trust inference [Mc. Guinness & Pinheiro da Silva, ISWC 2003, J. Web Semantics 2004] Paulo Pinheiro da Silva
Explainable System Challenge Explanation Proof Markup Language Information Manipulation Data Paulo Pinheiro da Silva
Proof Markup Language: Node Sets and Inference Steps Direct Assertion From KB 1 A->(A^B) Direct Assertion From Doc 1 Direct Assertion from Doc 2 A B A DAG of PML Node Sets (a collection of justifications) Modus Direct AND Assertion (DA) Ponens Intro (^I) from KB 1 (MP) A^B A->(A^B) A^B A MP A B A^B ^I DA A^B Extracted Proofs for the conclusion A^B Paulo Pinheiro da Silva
Encoding Hybrid and Distributed Proof Fragments p Proof Markup Language has a web-based solution for distribution n n p p Specification written in W 3 C’s OWL Each node set has one URI Node sets can be used to combine proofs generated by multiple agents OMEGA [Siekmann et al. , CADE 2002] has a nice solution for hybrid proofs http: //foo. com/NS. owl#NS 124 http: //bar. com/NS. owl#NS 125 rule: Modus Ponens (MP) has. Engine: JTP conclusion: (and A B) A^B has. Language: KIF http: //foo. com/NS. owl#NS 123 Paulo Pinheiro da Silva
Information Manipulation Traces Proof Markup Language Differences Formal Proofs Information manipulation traces Use of rules Mandatory Optional use or use of ‘unregistered rule’ Sentences Written in some formal language (e. g. , KIF, CL, DIMACS, etc. ) Written in a formal or informal language including natural language Use of multiple representation languages Uncommon Common Proof Markup Language covers the full spectrum of information manipulation traces! [Pinheiro da Silva, Mc. Guinness & Fikes, IS 2005] Paulo Pinheiro da Silva
Explainable System Challenge Explanation Proof Markup Language Information Manipulation Data Provenance Meta-data Paulo Pinheiro da Silva
Infrastructure: IWBase p Meta-data useful for disclosing knowledge provenance and reasoning information such as descriptions of n n n p p p inference engines along with their supported inference rules Information sources such as organizations, publications and ontologies Languages along with their axioms Core IWBase as well as domain IWBases OWL files for interoperability and database for scaling [Mc. Guinness & Pinheiro da Silva, IIWeb 2003] Paulo Pinheiro da Silva
Infrastructure: Core IWBase Statistics for relevant domain independent meta -data: Inference Engines Axioms 38 Method Rules 10 Derived Rules 6 Languages selec 56 Declarative Rules t selec 29 12 t Paulo Pinheiro da Silva
Explainable System Challenge Explanation Presentation Proof Markup Language Information Manipulation Data Provenance Meta-data Paulo Pinheiro da Silva
Browsing Proofs (1/2) Enable the visualization of proofs (and abstracted proofs) p Proofs can be “extracted” and browsed from both local and remote PML node sets and can be combined p Links provide access to proof-related meta-information p selec t Paulo Pinheiro da Silva
Browsing Proofs (2/2) Paulo Pinheiro da Silva
Explainable System Challenge Explanation Presentation Abstraction Proof Markup Language Information Manipulation Data Provenance Meta-data Paulo Pinheiro da Silva
Knowledge Provenance Elicitation Provenance information may be essential for users to trust answers. Data provenance (aka data Google-2. 0 says ‘A^B’ is the answer for my question. “has opinion” lineage) is defined and studied in the database literature. BBC NYT [Buneman et al. , ICDT 2001] [Cui and Widom, VLDB 2001] DA DA DA Knowledge provenance A->(A^B) A B extends data provenance by adding data derivation MP ^I DA provenance information A^B [Pinheiro da Silva, Mc. Guinness & Mc. Cool, A->(A^B) A A B Data Eng. Bulletin, 2003] A^B MP (CNN, BBC) A^B Why should I believe this? CNN ^I (BBC, NYT) Dir. Ass. A^B (CNN) Paulo Pinheiro da Silva
Knowledge Provenance Example er Answ Sourc e Paulo Pinheiro da Silva
Abstracting Proofs p Explanation tactics (a. k. a. rewriting rules) may be used to abstract proofs into more understandable and manageable explanations p Enable the use of axioms as inference rules preventing the presentation of primitive (and potentially less interesting and useful) rules p Eliminate intermediate results from proofs Paulo Pinheiro da Silva
Abstracting Proofs: An Example (1/2) Direct assertion (implies (and (Holds (owner ? person ? object) ? when) (organization ? object)) (Holds* (has. Office Direct assertion ? person ? object) ? when)) (implies (and (Holds* ? f ? t)) (not (Ab ? f ? t)) (Holds ? f ? t)) Direct assertion (Holds (owner Joeseph. Gradgrind. Foods) Apr 1_03) Direct assertion (organization Gradgrind. Foods ) Assumption Generalized Modus Ponens (Holds* (has. Office Joeseph. Gradgrind. Foods) Apr 1_03) Generalized Modus Ponens (Holds ((has. Office Joeseph. Gradgrind. Foods) Apr 1_03) (not (Ab (has. Office Joseph. Gradgri nd ? where) ? when)) Tactic Library Explanation tactic: “Organization Owner Typically Has Office at Organization” (implies (and (Holds (owner ? person ? object) ? when)) (organization ? object)) (Holds* (has. Office ? person (implies ? object) (and ? when)) (Holds* ? f ? t)) (not (Ab ? f ? t)) (Holds ? f ? t)) Direct assertion (Holds ((owner ? person ? object) ? when) Generalized Modus Ponens (Holds* ((has. Office ? person ? object) ? when) Generalized Modus Ponens (Holds ((has. Office ? person ? object) Direct assertion (organization ? object) (not (Ab (has. Office ? person ? object) ? when)) Direct assertion (Holds (owner Joeseph. Gradgrind. Foods) Apr 1_03) Direct assertion (organization Gradgrind. Food s) Organization Owner Typically Has Office at Organization (Holds (has. Office Joeseph. Gradgrind. Foods) Apr 1_03) ABSTRACTED PROOF Abstractor algorithm 1) Match conclusion (key for selecting tactics) 2) Match leaf nodes 3) Unify 4) Propagate conclusion 5) Apply the assertion-level rul 6) Propagate justified nodes Paulo Pinheiro da Silva
Abstracting Proofs: An Example (2/2) Direct assertion (Holds (owner Joeseph. Gradgri nd Gradgrind. Foods ) Apr 1_03) Direct assertion (organization Gradgrind. Foo ds) Organization Owner Typically Has Office at Organization (Holds (has. Office Joeseph. Gradgrind. Foods) Apr 1_03) ABSTRACTED PROOF Assertion-level rules are introduced in [Huang, PRICAI 1996]. Explanation tactics supports multi-level abstraction of proofs A rule says that the owner of an organization typically has an office in an organization Because • Joseph. Grardgrind owned Gradgrind. Foods on April 1 st 2003 • Gradgrind. Food is an organization therefore • Joseph. Gradgrind had an office at Gradgrind. Foods on April 1 st, 2003. ABSTRACTED PROOF IN DISCURSIVE STYLE Maybury describes strategies for rewriting abstracted proofs into English [AAAI 1991, AAAI 1993]. Paulo Pinheiro da Silva
Explainable System Challenge Explanation Understanding Interaction Presentation Abstraction Proof Markup Language Information Manipulation Data Provenance Meta-data Paulo Pinheiro da Silva
Explaining Answers: GUI Explainer Selec t action Users can exit the explainer providing feedback about their satisfiability with explanation(s) Users can ask for alternative explanations Paulo Pinheiro da Silva
Explainable System Challenge Explanation Understanding Interaction Presentation Abstraction Inference Meta-Language Information Inference Provenance Manipulation Rule Meta-data Data Specs Proof Markup Language Paulo Pinheiro da Silva
Inference Meta Language (Inference. ML ) p An inference rule involves pattern of transformations on expressions to produce a conclusion p Inference. ML uses schemas to state such transformations p Inference. ML defines a schema to be a pattern, which is any expression of CL in which: n some lexical items have been replaced by a schematic variable (or meta-variable) Example: nd. UI: '(forall (' N ')' q ')' |- ' (forall (' N - N. i ')' q[t/N. i] ')'; ; (Name N) (Sent q) (Term t) Paulo Pinheiro da Silva
Checking Proofs DA (A) From IWBase DA (implies (A) (and A B)) MP (and A B) MP: x; '(implies ' x y ')' |- y ; ; (Sent x y) (A) ; (implies (A) (and A B)) |- (and A B) binding of expressions to schematic variables: • x binds to (A) • y binds to (and A B) the rule schema instantiates directly to: = (A) ; (implies (A) (and A B)) |- (and A B) Paulo Pinheiro da Silva
Explainable System Challenge Explanation Understanding Trust Interaction Presentation Abstraction Inference Proof Markup Language Meta-Language Information Provenance Manipulation Meta-data Data Inference Rule Specs Paulo Pinheiro da Silva
IWTrust: Trust in Action Trust can be inferred from a Web of Trust. Google-2. 0 says ‘A^B’ is the answer for my question. ++ ? IWTrust provides infrastructure + for building webs of trust. The infrastructure includes a trust component responsible for computing trust values for answers. IWTrust is described in [Zaihrayeu, Pinheiro da Silva & Mc. Guinness, i. Trust 2005] XYZ 0 0 NYT DA DA A->(A^B) A CNN DA B MP ^I DA A^B A->(A^B) A A^B ++ Why should I trust the answer? A MP (CNN, XYZ) + ? B A^B ^I ? (XYZ, NYT) ++ + DA B A^B (CNN) 0 Paulo Pinheiro da Silva
Inference Web and Paulo p p p Paulo is a co-technical leader of the Inference Web project Paulo was the main IW developer during 1 ½ years Paulo has been the manager of the IW development team including members with the following profile: n n n p p p 1 research programmer 3 masters students 1 Ph. D. student Paulo has organized the IW weekly meetings Paulo has been responsible for presenting and demonstrating IW solutions at several DARPA and ARDA PI meetings Paulo has participated of the writing of grant proposals Paulo Pinheiro da Silva
Overview 1. What are explainable systems and why should we care about them? 2. Inference Web: Enabling Explainable Systems 3. Explainable Systems in Action 4. Explainable Systems 10 years from now Paulo Pinheiro da Silva
Application Areas p p Information extraction – IBM (UIMA), Stanford (TAP) Information integration – USC ISI (Prometheus/Mediator); Rutgers University (Prolog/Datalog) Task processing – Theorem proving SRI International (SPARK) n First-Order Theorem Provers –SRI International (SNARK); Stanford (JTP); n SATisfiability Solvers – University of Trento (J-SAT) Expert Systems – University of Fortaleza (JEOPS) n p p p University of Texas, Austin (KM) Service composition – Stanford, University of Toronto, UCSF (SDS) Semantic matching – University of Trento (S-Match) Debugging ontologies – University of Maryland, College Park (SWOOP/Pellet) Problem solving – University of Fortaleza (Expert. Cop) Trust Networks – U. of Trento (IWTrust) No single explanation approach has been used in so many diversified areas as Inference Web! Paulo Pinheiro da Silva
Extraction as Inference Goal: To provide browsable justifications of information extraction p Strategy: Reuse, adapt, and integrate existing technology: p n n p justification technology - Inference Web extraction technology - IBM’s UIMA Requires that systems to describe their processing as logical inferences n Requires a new perspective: IE as Inference n [Murdock, Pinheiro da Silva et al. , AAAI’s SSS 2005] Paulo Pinheiro da Silva
Extraction As Inference: An Example (1/2) Solution: Direct assertion from gradgrind. txt q A taxonomy of extraction tasks expressed as inference rules q Components that record IE justifications using rules in the taxonomy q. We have identified 9 types of extraction inferences: § 6 for analysis, and 3 for integration Joseph Gradgrind is the owner of Gradgrind Foods Entity Recognition IBM EAnnotator Joseph Gradgrind is the owner [organization] of radgrind Foods G Entity Identification IBM Cross -Annotator Coreference Joseph Gradgrind is the owner [organization] of radgrind Foods G [refers to Gradgrind. Foods] Direct assertion from KB 1 (implies (and (Holds (owner ? person ? object) ? when) (organization ? object)) (Holds* (has. Office ? person Direct assertion from ? object) ? when)) (implies KB 1 (and (Holds* ? f ? t)) (not (Ab ? f ? t)) (Holds ? f ? t)) Direct assertion from KB 1 (Holds (owner Joeseph. Gradgrind. Foods) Apr 1_03) Direct assertion from KB 1 Extracted Entity Classification Document Coreference (organization Gradgrind. Foods) ) Assumption Generalized Modus Ponens (has. Office (Holds* Joeseph. Gradgrind. Foods) Apr 1_03) (not (Ab (has. Office Joseph. Gradgri nd ? where) ? when)) Generalized Modus Ponens (Holds ((has. Office Joeseph. Gradgrind. Foods) Apr 1_03) Paulo Pinheiro da Silva
Extraction As Inference: An Example (2/2) Why should I believe this? Why should I believe these? Why should I believe that these documents say that? Paulo At Manchester, UK [Betty] Paulo At University of Manchester transitivity of At Theorem Proving University of Manchester At Manchester, UK http: //www. cs. man. ac. uk/~pinheirp Paulo is a Ph. D student at University of Manchester. http: //www. cs. man. ac. uk Informatio Extraction University of Manchester is located in Manchester, UK. Paulo Pinheiro da Silva
Explaining Tool Responses Requests and Responses Inferences for explaining answers (aka beliefs), and tasks (including actions) Generalization Questions and Answers Inferences for explaining answers (aka beliefs) Explain (v. tr. )1: n “To offer reasons for the actions, beliefs, or remarks of (oneself). ” New perspective: Task processing as inference 1 Dictionary. com Paulo Pinheiro da Silva
NL Explainer: An Example <user>: What are you doing now? <system>: I am trying to get an approval to buy a laptop. <user>: Why? [note: “Why? ” is rephrased to “Why are you trying to get an approval to buy a laptop? ] <system>: I have completed the previous requirement to get quotes so I am now working on get approval. <user>: OK, I am happy with your explanation. Levering explanation dialogues as in [Fiedler, IJCAI 2001] Using natural language support as in [Allen et al. , AAMAS 2002] Paulo Pinheiro da Silva
Overview 1. What are explainable systems and why should we care about them? 2. Inference Web: Enabling Explainable Systems 3. Explainable Systems in Action 4. Explainable Systems 10 years from now Paulo Pinheiro da Silva
Inference Web Contributions 1. Language for encoding hybrid, distributed proof fragments based on web technologies. Support for both formal and informal proofs (information manipulation traces). 2. Support (registry, language, services) for knowledge provenance. 3. Declarative inference rule representation for checking hybrid, distributed proofs. 4. Multiple strategies for proof abstraction, presentation and interaction. 6 Explanation Understanding 5 4 Interaction 4 Presentation 4 Trust Abstraction 3 Inference Proof Markup Language Meta-Language 1 2 Information 2 Inference Provenance Manipulation Rule Meta-data Data Specs 5. End-to-end trust value computation for answers. 6. Comprehensive solution for explainable systems. Paulo Pinheiro da Silva
Open Issues Automated generation of explanation tactics p Performance for abstracting and checking proofs p Use of machine learning and user modeling to support interaction p n n n p Adaptive explanations Explanation contexts Modeling user knowledge Metrics and evaluations for explainable systems Paulo Pinheiro da Silva
Three Years From Now p p p An initial research community working on explainable systems Adaptive explanations based on user modeling IWBase registration of a large set of software systems n p p p Registration of a comprehensive set of primitive rules Established library of explanation tactics First generation of metrics and evaluation methods for explainable systems Inference Web is a solution for the Semantic Web proof and trust layers http: //www. w 3. org/2004/Talks/0412 -RDF-functions/slide 4 -0. html Paulo Pinheiro da Silva
Ten Years From Now p p p An established research community working on explainable systems A theory for explainable systems Established metrics for explainable systems First (or second) generation of industrial explainable systems A standard language for encoding information manipulation traces (probably derived from PML among other proposals). The language will include support for the following: n n probabilistic reasoning inductive reasoning Paulo Pinheiro da Silva
and Inference Web p Immediate connections n Explaining Task Processing p p n Explaining Tool Responses p p Task. Tracer CALO with Intelligent Information Systems team Explaining WYSIWYT – with End Users Shaping Effective Software team Potential connections n Explanation generation p n Explanation-based learning p n Filtering Learning with Learning and Adaptive Systems team Explaining pattern and object recognition from videos and graphs p with Computer Graphics and Vision Paulo Pinheiro da Silva
Скачать презентацию Explainable Systems The Inference Web Approach Paulo Pinheiro
c2185a7830439559fc8393f28efe2c9a.ppt