You are on page 1of 30

Explaining Complex Systems

Deborah L. McGuinness

Acting Director Knowledge Systems, AI Lab,


Stanford University
Tetherless World Chair, Rensselaer Polytechnic
Institute (RPI)
Increasing Explanation Motivations
 Systems are getting more complex
 Multiple heterogeneous distributed information sources

 Highly variable reliability of information sources

 Interest in reuse of information systems (many times for purposes other


than those originally planned for)
 Hybrid and distributed processing

 Multiple types of components, including multiple learners (e.g., calo, gila),


multiple text components (e.g., uima, kani, …)
 Less transparency of system computation and reasoning

 Systems are taking more autonomous control


 Guide/assist user actions

 Perform autonomous actions on behalf of user

 “reason, learn from experience, be told what to do, explain what they are
doing, reflect on their experience, and respond robustly to surprise” *

* DARPA PAL program: http://www.darpa.mil/ipto/programs/pal/

Deborah L. McGuinness Semantic e-Science Vancouver July 23, 2007 2


Motivation
Support explanations of provenance, information manipulation
trace, and trust using an interoperable, transparent, and
user-friendly knowledge provenance infrastructure.
 Explanation
 Provenance – if users (humans and agents) are to use and integrate data from
unknown, unreliable, or evolving sources, they need provenance metadata for
evaluation
 Information manipulation trace – if information has been manipulated (i.e., by
sound deduction or by heuristic processes), information manipulation trace
information should be available
 Trust – if some sources are more trustworthy than others, representations
should be available to encode, propagate, combine, and (appropriately) display
trust values
 Interoperability – as systems use varied sources and multiple information
manipulation engines, they benefit more from encodings that are shareable
& interoperable
 Transparency –explanations can be used to provide transparency and
accountability for systems by allowing (authorized) users to see what the
system has done..
 Usability – varied users need rich representation options and a broad range
of tool support to provide context- and user-appropriate presentations.
Deborah L. McGuinness Semantic e-Science Vancouver July 23, 2007 3
Inference Web Explanation
Infrastructure
WWW Toolkit
IWTrust Trust computation
SDS OWL-S/BPEL
Trace of web service discovery
Proof Markup IW Explainer/ End-user friendly
CWM N3 Language (PML) Abstractor
visualization
Trace of rule application
Expert friendly
JTP KIF Trust IWBrowser Visualization
Trace of theorem prover

Justification search engine


SPARK SPARK-L IWSearch based publishing
Trace of task execution
Provenance provenance
UIMA Text Analytics IWBase registration
Trace of information extraction

 Semantic Web based infrastructure


 PML is an explanation interlingua
 Represent knowledge provenance (who, where, when…)
 Represent justifications and workflow traces across system boundaries
 Inference Web provides a toolkit for data management and visualization
Deborah L. McGuinness Semantic e-Science Vancouver July 23, 2007 4
Explaining Information Manipulation in PML
(behind the scenes)
Question - foo:question1
“when and where does Ramazi have an office”
Query - foo:query1
“(Holds (|hasOffice| |Ramazi| ?where) ?when) ” isQueryFor
hasAnswer IWRegistry
(Provenance Metadata)
NodeSet - foo:ns1 {answer to query} hasLanguage
“(Holds (|hasOffice| |Ramazi| |SelectGourmetFoods|) April_01_2003)”
isConsequentOf Language - reg:KIF
fromQuery
hasInferencEngine
InferenceStep InferenceEngine - reg:JTP
hasInferenceRule
hasAntecendent hasVariableMapping InferenceRule - reg:GMP
Mapping
From: “?f” To: “(|hasOffice| |Ramazi| ?where)”
More mappings …
Source – reg:TypicalityOnto
More NodeSets…

NodeSet - foo:ns2 {direct assertion}


“(<= (or (Ab ?f ?t) (Holds ?f ?t)) (Holds* ?f ?t))”
isConsequentOf
fromAnswer
hasSourceUsage hasSource
InferenceStep SourceUsage

Deborah L. McGuinness Semantic e-Science Vancouver July 23, 2007 5


Filtered View
Views of Explanation

filtered focused global

Explanation abstraction
(in PML)
discourse
trust
provenance

 Show Highlights
 Query
 Answer
 Supporting
assertions
 Sources

Deborah L. McGuinness Semantic e-Science Vancouver July 23, 2007 6


Focused View
Views of Explanation

filtered focused global

Explanation abstraction
(in PML)
discourse
trust
provenance

 One step of justification


 Original query
 Conclusion
 Direct antecedents
 Inference rule

 Contextually appropriate
follow-up questions
 Sources
 Ground Assertions
 Assumptions
 Full trace
 Question answerers used
 …

Deborah L. McGuinness Semantic e-Science Vancouver July 23, 2007 7


Global View and More
Views of Explanation

filtered focused global

Explanation abstraction
(in PML)
discourse
trust
provenance

 Explanation as a graph
 Customizable browser options
 Proof style
 Sentence format
 Lens magnitude
 Lens width

 More information
 Provenance metadata
 Source PML
 Proof statistics
 Variable bindings
 Link to tabulator

Deborah L. McGuinness

Semantic e-Science Vancouver July 23, 2007 8
Abstraction View
 Rewrite rules used to hide part(s) of the sub-graph

Deborah L. McGuinness Semantic e-Science Vancouver July 23, 2007 9


Discourse View
Views of Explanation
 (Limited) natural language interface
 Mixed initiative dialogue filtered focused global
 Exemplified in CALO domain abstraction
Explanation
 Explains task execution component (in PML)
powered by learned and human discourse
trust
generated procedures provenance

Deborah L. McGuinness Semantic e-Science Vancouver July 23, 2007 10


Provenance View
Views of Explanation
 Source metadata: name, description, …
 Source-Usage metadata: which fragment of filtered focused global
a source has been used when
Explanation abstraction
(in PML)
discourse
trust
provenance

Deborah L. McGuinness Semantic e-Science Vancouver July 23, 2007 11


Trust View
Views of Explanation

filtered focused global

Trust Tab Explanation abstraction


Detailed trust
(in PML)
explanation
discourse
trust
provenance

 (preliminary) simple
trust representation
 Provides colored
(mouseable) view
based on trust values
 Enables sharing and
collaborative
computation and
propagation of trust
Fragment values
colored by
trust value

Deborah L. McGuinness Semantic e-Science Vancouver July 23, 2007 12


Inference Web Data Access Interface
 Browse
 Organized by class-hierarchy
 Customized entry
 Summary and audit views (in report
listing)

 Search (with filters)


 NodeSet
 root NodeSet
 Query
 Conclusion
 …

Deborah L. McGuinness Semantic e-Science Vancouver July 23, 2007 13


The Use-Ask-Understand-Update Cycle

Use Ask

Understand /
Update
Accept

Deborah L. McGuinness Semantic e-Science Vancouver July 23, 2007 14


Selected IW and PML Applications
 Portable proofs across reasoners: JTP (with temporal
and context reasoners (Stanford); CWM (W3C),
SNARK(SRI), …
 Explaining web service composition and discovery
(SNRC)
 Explaining information extraction (more emphasis on
provenance – KANI, UIMA)
 Explaining intelligence analysts’ tools (NIMD/KANI)
 Explaining tasks processing (SPARK / CALO)
 Explaining learned procedures (TAILOR, LAPDOG, /
CALO)
 Explaining privacy policy law validation (TAMI)
 Explaining decision making and machine learning
(GILA)
 Explaining trust in social collaborative networks
(TrustTab)
 Registered knowledge provenance: IW Registrar
(Explainable
Deborah L. McGuinness
Knowledge Aggregation)
Semantic e-Science Vancouver July 23, 2007 15
Trend: Semantically Enabling Applications
leveraging the Semantic Web

focus

Deborah L. McGuinness Semantic e-Science Vancouver July 23, 2007 16


Semantic Web Methodology and
Technology Development Process
 Establish and improve a well-defined methodology vision for Semantic
Technology-based application development

Adopt
Leverage Technology Expert Review &
Rapid Technology Approach Iteration
Open World: Prototype Infrastructure
Evolve, Iterate,
Redesign,
Redeploy

Use Tools
Analysis
Use Case

Develop
Small Team, model/
mixed skills ontology

Joint with P. Fox


Deborah L. McGuinness Semantic e-Science Vancouver July 23, 2007 17
Scientific Environment Goal

Scientists should be able to access a global,


distributed knowledge base of scientific data that:
• appears to be integrated
• appears to be locally available

But… data is obtained by multiple instruments,


using various protocols, in differing vocabularies,
using (sometimes unstated) assumptions, with
inconsistent (or non-existent) meta-data. It may
be inconsistent, incomplete, evolving, and
distributed.

Deborah L. McGuinness Semantic e-Science Vancouver July 23, 2007 18


Virtual Observatory

 Workshop: A Virtual Observatory (VO) is a suite of


software applications on a set of computers that
allows users to uniformly find, access, and use
resources (data, software, document, and image
products and services using these) from a
collection of distributed product repositories and
service providers. A VO is a service that unites
services and/or multiple repositories.
lwsde.gsfc.nasa.gov/VO_Framework_7_Jan_05.doc
 VxOs - x is one discipline
 Trend: VxyO – multi-discipline virtual
observatories
Deborah L. McGuinness Semantic e-Science Vancouver July 23, 2007 19
Selected VxyO Motivation: Mt. Spurr, AK. 8/18/1992
eruption, USGS

http://www.avo.alaska.edu/image.php?id=319

Deborah L. McGuinness Semantic e-Science Vancouver July 23, 2007 20


Eruption cloud movement from Mt.Spurr, AK,1992

Deborah L. McGuinness Semantic e-Science VancouverUSGS


July 23, 2007 21
Tropopause

http://aerosols.larc.nasa.gov/volcano2.swf
22
Atmosphere Use Case
 Determine the statistical signatures of both
volcanic and solar forcings on the height of
the tropopause
From paleoclimate researcher – Caspar Ammann – Climate and Global
Dynamics Division of NCAR - CGD/NCAR

Layperson perspective:
- look for indicators of acid rain in the part of the atmosphere we
experience…
(look at measurements of sulfur dioxide in relation to sulfuric acid after
volcanic eruptions at the boundary of the troposphere and the
stratosphere)

Nasa funded effort with Fox - NCAR, Sinha - Va. Tech, Raskin - JPL

Deborah L. McGuinness Semantic e-Science Vancouver July 23, 2007 23


Use Case detail: A volcano erupts

 Preferentially it’s a tropical mountain (+/- 30 degrees of the equator) with ‘acidic’ magma; more
SiO2, and it erupts with great intensity so that material and large amounts of gas are injected into
the stratosphere.
 The SO2 gas converts to H2SO4 (Sulfuric Acid) + H2O (75% H2SO4 + 25% H2O). The half life of
SO2 is about 30 - 40 days.
 The sulfuric acid condensates to little super-cooled liquid droplets. These are the volcanic
aerosol that will linger around for a year or two.
 Brewer Dobson Circulation of the stratosphere will transport aerosol to higher latitudes. The
particles generate great sunsets, most commonly first seen in fall of the respective hemisphere. The
sunlight gets partially reflected, some part gets scattered in the forward direction.
 Result is that the direct solar beam is reduced, yet diffuse skylight increases. The scattering is
responsible for the colorful sunsets as more and more of the blue wavelength are scattered away.in
mid-latitudes the volcanic aerosol starts to settle, but most efficient removal from the stratosphere
is through tropopause folds in the vicinity of the storm tracks.
 If particles get over the pole, which happens in spring of the respective hemisphere, then they will
settle down and fall onto polar ice caps. Its from these ice caps that we recover annual records of
sulfate flux or deposit.
 We get ice cores that show continuous deposition information. Nowadays we measure sulfate or
SO4(2-). Earlier measurements were indirect, putting an electric current through the ice and
measuring the delay. With acids present, the electric flow would be faster.
 What we are looking for are pulse like events with a build up over a few months (mostly in summer,
when the vortex is gone), and then a decay of the peak of about 1/e in 12 months.
 The distribution of these pulses was found to follow an extreme value distribution (Frechet) with a
heavy tail.

Deborah L. McGuinness Semantic e-Science Vancouver July 23, 2007 24


Use Case detail: … climate

 So reflection reduces the total amount of energy, forward scattering just changes the
beam, path length, but that's it.
 The dry fogs in the sky (even after thunderstorm) still up there, thus stratosphere not
troposphere.
 The tropical reservoir will keep delivering aerosol for about two years after the
eruption.
 The particles are excellent scatterers in short wavelength. They do absorb in NIR
and in IR. Because of absorption, there is a local temperature change in the lower
stratosphere.
 This temperature change will cause some convective motion to further spread the
aerosol, and second: Its good factual stuff. Once it warms up, it will generate a
temperature gradient. Horizontal temperature gradients increase the baroclinicity
and thus storms, and they speedup the local zonal winds. This change in zonal wind in
high latitudes is particularly large in winter. This increased zonal wind (Westerly) will
remove all cold air that tries to buildup over winter in high arctic.
 Therefore, the temperature anomaly in winter time is actually quite okay.
 Impact of volcanoes is to cool the surface through scattering of radiation.
 In winter time over the continents there might be some warming. In the stratosphere,
the aerosol warm.
 The amount of GHG emitted is comparably small to the reservoir in the air.
 The hydrologic cycle responds to a volcanic eruption.

Deborah L. McGuinness Semantic e-Science Vancouver July 23, 2007 25


Atmosphere (portions from SWEET)

Deborah L. McGuinness Semantic e-Science Vancouver July 23, 2007 26


Atmosphere II

Deborah L. McGuinness Semantic e-Science Vancouver July 23, 2007 27


Interdisciplinary VOs, Semantics, and
Explanation
 Background ontologies can be used to help access and
integrate distributed data sources
 Semantically-enabled VOs are starting to go into service (e.g.,
VSTO – talk later today on services and in IAAI on tues –
deployed application track)
 Provenance issues become more critical in such systems –
where did the data come from? How was it collected? Who
collected it? What are their credentials? etc.
 Annotations and explanations may be the key to increasing trust
in answers
 Annotations may simultaneously be a key to increasing
contributions as users become confident that they will get
appropriate credit
 An explanation interlingua (such as PML) may be a critical
component to semantic integration, sharing, and acceptance
 An explanation infrastructure (such as Inference Web) may
provide a foundation on which to build such applications

Deborah L. McGuinness Semantic e-Science Vancouver July 23, 2007 28


References
 IW and PML
 Deborah L. McGuinness and Paulo Pinheiro da Silva. Explaining Answers from the Semantic Web: The Inference
Web Approach. Journal of Web Semantics. Vol.1 No.4, 2004
 Deborah L. McGuinness. Knowledge Representation for Question Answering. In Proceedings of the American
Association for Artificial Intelligence Spring Symposium Workshop on New Directions for Question Answering.
Stanford University, Stanford, CA. pages 75-77, AAAI Press, March 2003.
 Paulo Pinheiro da Silva, Deborah L. McGuinness and Richard Fikes. A Proof Markup Language for Semantic Web
Services. Information Systems. Volume 31, Issues 4-5, 2006 (New Version – PML2 in Explanation Aware
Computing Workshop at AAAI 2007.
 Trust
 Deborah L. McGuinness, Honglei Zeng, Paulo Pinheiro da Silva, Li Ding, Dhyanesh Narayanan, and Mayukh
Bhaowal. Investigations into Trust for Collaborative Information Repositories: A Wikipedia Case Study.
WWW2006 Workshop on the Models of Trust for the Web (MTW'06)
 Ilya Zaihrayeu, Paulo Pinheiro da Silva and Deborah L. McGuinness. IWTrust: Improving User Trust in Answers
from the Web. Proceedings of 3rd International Conference on Trust Management (iTrust2005)
 H. Zeng, M. Alhossaini, L. Ding, R. Fikes, and D. McGuinness. Computing Trust from Revision History. The 2006
International Conference on Privacy, Security and Trust (PST 2006)
 Honglei Zeng, Maher Alhossaini, Richard Fikes, and Deborah L. McGuinness. Mining Revision History to Assess
Trustworthiness of Article Fragments. The 2nd International Conference on Collaborative Computing: Networking,
Applications and Worksharing
 Some particular aspects of explanation:
 Text Analytics: J. William Murdock, Deborah L. McGuinness, Paulo Pinheiro da Silva, Christopher Welty and David
Ferrucci. Explaining Conclusions from Diverse Knowledge Sources. The 5th International Semantic Web
Conference(ISWC2006)
 Learning Task Procedures: Deborah L. McGuinness, Alyssa Glass, Michael Wolverton and Paulo Pinheiro da Silva.
Explaining Task Processing in Cognitive Assistants That Learn. FLAIRS 2007.
 Explaining Data Usage: Daniel J. Weitzner, Hal Abelson, Tim Berners-Lee, Chris P. Hanson, Jim Hendler, Lalana
Kagal, Deborah L. McGuinness, Gerald J. Sussman, K. Krasnow Waterman. Transparent Accountable Inferencing
for Privacy Risk Management. Proceedings of AAAI Spring Symposium on The Semantic Web meets
eGovernment. AAAI Press, Stanford University, USA 2006. Also available as Stanford KSL Technical Report KSL-
06-03 and MIT CSAIL Technical Report-2006-007.
 User needs: Andrew. J. Cowell, Deborah L. McGuinness, Carrie F. Varley, and David A. Thurman. Knowledge-
Worker Requirements for Next Generation Query Answering and Explanation Systems. In the Proceedings of
the Workshop on Intelligent User Interfaces for Intelligence Analysis, International Conference on Intelligent User
Interfaces (IUIL.2006),
Deborah Sydney, Australia.
McGuinness Semantic e-Science Vancouver July 23, 2007 29
Extra

Deborah L. McGuinness Semantic e-Science Vancouver July 23, 2007 30