Enabling Explanations The Inference Web and PML Approach

Enabling Explanations: The Inference Web and PML Approach Deborah Mc. Guinness, Paulo Pinheiro da Silva, Li Ding Knowledge Systems Laboratory Stanford University info: [dlm | pp ] @ ksl. stanford. edu Inference Web and PML is joint work with Fikes, Chang, Deshwal, Ding, Narayanan, Glass, Makarios, Jenkins, Millar, and has particularly benefited from input from Batelle, IBM, SRI, MIT, UMBC, and others More info: http: //iw. stanford. edu/documents_papers. html

Outline n n n n Motivation Background - Inference Web PML Building a PML document by example Browsing PML Directions and plans Discussion 2

Motivation – Trust through. Transparency If users (humans and agents) are to use, reuse, and integrate system answers, they must trust them. System transparency supports understanding and trust. Even simple “lookup” systems benefit from providing information about their sources. Systems that manipulate information (with sound deduction or potentially unsound heuristics) benefit from providing information about their manipulations. Goal: Provide interoperable infrastructure that supports explanations of sources, assumptions, and answers as an enabler for trust. 3

Inference Web Concerns Information Manipulation Traces n hybrid, distributed, portable, shareable, combinable encoding of proof fragments supporting multiple justifications Presentation n multiple display formats supporting browsing, visualization, summaries, … Abstraction n understandable summaries Interaction n multi-modal mixed initiative options including naturallanguage and GUI dialogues, adaptive, contextsensitive interaction Trust n source and reasoning provenance, automated trust inference [Mc. Guinness & Pinheiro da Silva, ISWC 2003, J. Journal of Web Semantics 2004] 4

Inference Web Framework for explaining question answering tasks by abstracting, storing, exchanging, combining, annotating, filtering, comparing, and rendering justifications from question answerers n n n n IW’s Proof Markup Language (PML) is an interlingua for proof interchange. Represented in OWL IWBase is a distributed repository of meta-information IW Registration and Validation services provide support for PML generation, validation, and checking* IW Browser provides display capabilities for PML documents IW Abstractor provides rewriting capabilities enabling more understandable presentations IW Explainer provides multi-modal dialogue options including alternative strategies for presenting explanations, visualizations, and summaries IW Trust provides methods for propagating trust values IW Search (enhanced SWOOGLE for PML documents) 5

How to achieve transparency using IW n n n Question Answering system gets registered with Inference Web (using registration services) Question Answering system encodes justifications of its answers in PML (using PML generation and validator services) Application provides access to explanations via IW’s explainer, browser, and search features 6

One Proof Browser 7

8

9

IW Explanation Application Areas Theorem proving First-Order Theorem Provers – Stanford (JTP (KIF/OWL/…)); SRI (SNARK); SATisfiability Solvers – University of Trento (JSAT) Information extraction – IBM (UIMA), Stanford (TAP) Information integration/aggregation – USC ISI (Prometheus, Mediator -> Fetch); Rutgers , Stanford (TAP) Task processing – SRI International (SPARK) Service composition – Stanford, U. of Toronto, UCSF (SDS) Semantic matching – University of Trento (S-MATCH) Problem solving – University of Fortaleza (Expert. Cop) Trust Networks – U. of Trento (IWTrust) New collaboration with W 3 C / MIT Govt. Program Usage: ARDA NIMD, DARPA PAL, SAPIENT, Explainable Knowl. Agg, UIMA … Company Usage: IBM, SRI, Sandpiper, … 10

Proof Markup Language n n n PML is a proof interlingua It can be used to represent justification of information manipulation steps done by theorem provers, extractors, etc. The main components concern inference representation and provenance issues such as author, source, etc. 11

Structure 12

Querying the Web n John asks Google 2. 0 the following question: What happened in The Night Club on March 1 st, 1997? (=>S^F) n n An Answer: There was smoke and fire (S^F) Sources: n “The whole place got tons of black smoke. ” (S) http: //www. cnn. com/1997/03/03/US/XYZ. html n “fire took hold with devastating speed” (F) http: //news. bbc. co. uk/1/hi/world/279 w 1119. stm n “Where there's smoke there's fire” (S->F) http: //www. giga-usa. com/quotes/topics/proverbs_t 374. htm n “the fire quickly spread, filling the building with (S^F) thick, black smoke” http: //www. redcross. org/article/947, 00. html 13

Proof Markup Language Top-Level Concepts Source. Usage Inference. Step Model. Element Mapping Proof. Element Node. Set Question Query Provenance. Element Person Ontology … Organization 14

Proof Markup Language: Node Sets and Inference Steps (1/2) A trivial justification: (S^F) has been asserted from the Red Cross (RC) website Encoding this justification in PML: http: //foo. com/Example. owl#Smoke. Fire iw: Node. Set iw: is. Consequence. Of iw: Inference. Step iw: has. Rule: Direct Assertion (DA) iw: has. Source. Usage: 824, 1058 on RC iw: has. Engine: CWM iw: has. Conclusion: (S^F) iw: has. Language: N 3 15

Proof Markup Language: Node Sets and Inference Steps (2/2) And this is the same Node. Set in XML: <iw: Node. Set rdf: about="http: //foo. com/Example. owl#Smoke. Fire"> <iw: has. Conclusion>(S^F)</iw: has. Conclusion> <iw: has. Language rdf: resource="http: //inferenceweb. stanford. edu/registry/LG/N 3. owl#N 3" /> <iw: is. Consequent. Of> <iw: Inference. Step> <iw: has. Index rdf: datatype="http: //www. w 3. org/2001/XMLSchema#int">0</iw: has. Index> <iw: has. Rule rdf: resource="http: //inferenceweb. stanford. edu/registry/DPR/Told. owl#Told"/> <iw: has. Inference. Engine rdf: resource="http: //inferenceweb. stanford. edu/registry/IE/CWM. owl#CWM"/> <iw: has. Source. Usage> <iw: span. From. Byte rdf: datatype="http: //www. w 3. org/2001/XMLSchema#int">824</iw: span. From. Byte> <iw: span. To. Byte rdf: datatype="http: //www. w 3. org/2001/XMLSchema#int">1058</iw: span. To. Byte> <iw: has. Source rdf: resource="http: //inferenceweb. stanford. edu/registry/PUB/RC. owl#RC"/> </iw: Source. Usage> </iw: has. Source. Usage> </iw: Inference. Step> </iw: is. Consequent. Of> </iw: Node. Set> 16

Justification Collections and Proofs Direct Assertion From BBC S DA DA F S^F F A DAG of PML Node Sets (a collection of justifications) Direct AND Assertion (DA) Intro (^I) from RC S^F iw: has. Antecedent S Direct Assertion from CNN ^I DA S^F Extracted Proofs for the conclusion S^F 17

The Combine Operation Direct Assertion From BBC Direct Assertion From Prov S S->F Direct Assertion From BBC Modus Ponens (MP) F F F #A Direct AND Assertion (DA) Intro (^I) from RC S^F S->F Modus Direct Ponens Assertion (MP) from CNN Direct Assertion from CNN S Direct Assertion From Prov S #B Direct Assertion From BBC Combine(#A, #B)= #C Direct AND Assertion (DA) Intro (^I) from RC #C S^F 18

A Result of Combining Justifications Direct Assertion From BBC Direct Assertion From Prov S S->F {BBC, Prov} S Modus Direct Ponens Assertion (MP) from CNN S DA DA S->F MP F ^I S^F {BBC, CNN} F Direct AND Assertion (DA) Intro (^I) from RC DA #C S DA F S^F DA ^I {RC} DA S^F 19

SWOOP Class View 20

Protege 21

SWOOP Property View Swoop 2. 2. 1 loaded with: iw. stanford. edu/2004/07/iw. owl 22

Proof Elements 23

Provenance Elements 24

Potential Usage in TAMI n n PML as a justification interlingua Inference Web tools for browsing, abstracting, summarizations, etc. IW Registry for components IW Search for finding proofs Next steps: n Review PML for representational adequacy ( chris’ issue of markers for ignoring terms, …) n Register question answerers (CWM, Pychyncho? , TMS) n Augment question answerers to generate PML n Review output leveraging existing strategies and tactics n Generate new strategies and tactics as needed 25

Discussion http: //iw. stanford. edu/2004/07/iw. owl 26