IKRAFT: Interactive Knowledge Representation and Acquisition from Text

1
IKRAFTInteractive Knowledge Representation and
Acquisition from Text

• Yolanda Gil
• Varun Ratnakar
• www.isi.edu/expect/projects/trellis
• trellis.semanticweb.org
• USC/Information Sciences Institute
• gil_at_isi.edu

2
MotivationHow KBs Are Built Today
Domain Expert
Read/ask /study/listen...
Knowledge Engineer
analyze/group/index...
Knowledge Acquisition Tools
structure/relate/fit...
KB
reason/deduce/solve
3
MotivationThe Aftermath of Knowledge Base
Development
Domain Expert
Knowledge Engineer
TRASH
Knowledge Acquisition Tools
KB
reason/deduce/solve
4
MotivationCapturing the Design of Knowledge
Bases
Richer representations More ambiguous More
versatile
Introductory texts, expert hints, explanations,
dialogues, comments, examples, exceptions,...
Info. extraction templates, dialogue segments and
pegs, filled-out forms, high-level
connections,...
Knowledge Base
Descriptions augmented with prototypical
examples exceptions, problem-solving steps
and substeps, ...
More formal More concrete More introspectible
Alternative formalizations (KIF, MELD, RDF,),
alternative views of the same notion (e.g.,
what is a threat)
((( )) ())))
(defconcept bridge ()))
5
Claims

• Knowledge can be reused at any level of
  (in)formality
• Knowledge can be extended more easily
• Addtl documents and semi-formal structures
  readily available
• Knowledge can be translated and integrated at any
  level to facilitate interoperability
• KR languages can be a straitjacket for some kinds
  of knowledge
• Intelligent systems will provide better
  justifications
• Many users want to know where axioms came from
  before they trust systems reasoning
• Content providers will not need to be
  sophisticated programmers/knowledge engineers
• May be easier for end users to organize knowledge
  rather than formalize it
• Good symbiosis of sophisticated and
  unsophisticated users

6
An ExampleBuilding a Knowledge Base from a
Textbook(DARPA Rapid Knowledge Formation -- RKF)

• The first step a cell takes in reading out part
  of its genetic instructions is to copy the
  required portion of the nucleotide sequence of
  DNA the gene into a nucleotide sequence of
  RNA. The process is called transcription because
  the information, though copied into another
  chemical form, is still written in essentially
  the same language the language of nucleotides.
  Like DNA, RNA is a linear polymer made of four
  different types of nucleotides subunits linked
  together by phosphodiester bonds. It differs from
  DNA chemically in two respects (1) the
  nucleotides in RNA are ribonucleotides that is,
  they contain the sugar ribose (hence the name
  ribonucleic acid) rather than deoxyribose (2)
  although, like DNA, RNA contains the bases
  adenine (A), guanine (G), and cytosine (C), it
  contains uracil (U) instead of the thymine (T) in
  DNA. Since U, like T, can base-pair by
  hydrogen-bonding with A, the base-pairing
  properties described for DNA also apply to RNA
• --
  Essential Cell Biology, Alberts et al. 1992

7
Protein Synthesis in RKFs SHAKEN Authored by a
Biologist Chaudri et al 2001
8
Step 1 Selecting Relevant Knowledge Fragments

• The first step a cell takes in reading out part
  of its genetic instructions is to copy the
  required portion of the nucleotide sequence of
  DNA the gene into a nucleotide sequence of
  RNA. The process is called transcription because
  the information, though copied into another
  chemical form, is still written in essentially
  the same language the language of nucleotides.
  Like DNA, RNA is a linear polymer made of four
  different types of nucleotides subunits linked
  together by phosphodiester bonds. It differs from
  DNA chemically in two respects (1) the
  nucleotides in RNA are ribonucleotides that is,
  they contain the sugar ribose (hence the name
  ribonucleic acid) rather than deoxyribose (2)
  although, like DNA, RNA contains the bases
  adenine (A), guanine (G), and cytosine (C), it
  contains uracil (U) instead of the thymine (T) in
  DNA. Since U, like T, can base-pair by
  hydrogen-bonding with A, the base-pairing
  properties described for DNA also apply to RNA
• --
  Essential Cell Biology, Alberts et al. 1992

9
Step 2Composing Stylized Knowledge Fragments

• - ribose
• - it is a kind of sugar, like deoxyribose
• - it is contained in the nucleotides of
  RNA
• - uracil
• - it is a kind of nucleotide, like adenine
  and guanine
• - it can base-pair with adenine
• - RNA
• - it is a kind of nucleic acid, like DNA
• - it contains uracil instead of thymine
• - it is single-stranded
• - it folds in complex 3-D shapes
• - nucleotides are linked with
  phospohodiester bonds, like DNA
• - there are many types of RNA
• - RNA is the template for synthesizing
  protein
• - its nucleotides contain the sugar ribose
  (DNA has deoxyribose)
• - gene
• - subsequence of DNA that can be used as a
  template to create protein
• - protein synthesis
• - non-destructive creation process RNA and
  protein created from DNA

10
Step 3Creating Knowledge Base Items

• (defconcept uracil is-primitive nucleotide
• constraints (the base-pair adenine))
• (defconcept RNA
• is (and nucleic-acid
• (some contains uracil)))

11
IKRAFT Interactive Knowledge Representation and
Acquisition from Text

• User starts with documents, extracts a small
  amount of information from them
• Text contains significant portions for
  context/reference/recall
• IKRAFT allows users to annotate text with
  statements, expressed in natural language
• Highlight portions of original text, annotate
  statement
• Statements tend to be stylized
• Statements are parsed, system generates summary
  of
• Objects
• Events/actions

12
IKRAFT Annotating Manual Information Extraction
13
IKRAFT Extracting Statements from
Complementary/Contradictory Text Sources
14
IKRAFT Documenting Seismic Hazard in Southern
California
15
Seismic Hazard Analysis (SHA) for Southern
California Earthquake Center (SCEC)
16
DOCKER Scientist Publishes SHA Models

• User specifies
• Types of model parameters
• Format of input messages
• Documentation
• Constraints

Web Browser
AS97
DOCKER
Model Specification
User Interface
AS97
docs
types
msg
constrs
Wrapper Generation (WSDL, PWL)
Constraint Acquisition
AS97 ontology
SCEC ontologies
17
Documenting the Model with IKRAFT
18
Documenting Each Constraint
19
Formalizing Simple Constraints
20
Documentation of Constraints (Some Are
Formalized, Some Are Not)
21
DOCKER Engineer Uses SHA Model

• User can
• Browse through SHA models
• Invoke SHA models
• Get help in selecting appropriate model

AS97
Web Browser
DOCKER
AS97
docs
constrs
Model Reasoning
User Interface
msg
types
AS97 ontology
Pathway Elicitation
Constraint Reasoning
Shared ontologies
KRR (Powerloom)
22
DOCKER Detects Constraint Violations
23
Should Engineer Override Constraint Specified by
Model Developer?
24
Engineer Brings Up IKRAFT to Find Reasons for the
Constraint
25
Engineer Can Check Additional Model Constraints
(Not Formalized)
26
Constraints Grounded on Model Documentation
27
Engineers Makes an Informed Decision on Whether
to Override the Constraint
28
Discussion

• Overhead in capturing the rationale?
• Related to motivation and payoff
• Rationale here is captured in a very simple
  process
• Related Work
• Documenting design rationale Shum 96
• Methodologies for knowledge base development
  Schreiber et al 00
• Higher-level languages, e.g., KARL Fensel et al
  98

29
Conclusions and Future Work

• IKRAFT helps users document formal expressions
• Each formal expression is back up by a concise NL
  statement that is linked back to one or more
  sources
• Users can understand justification for systems
  reasoning (e.g., SHA)
• Future work
• NLP techniques to extract terms from users
  concise statements
• Controlled grammar for formulation of statements
• Other documentation e.g., tables, forms,
  exceptions
• High payoff in capturing the rationale of
  knowledge bases

30
Speculation Will the (Semantic) Web End Up
Looking Like This?
Richer representations More ambiguous More
versatile
Introductory texts, expert hints, explanations,
dialogues, comments, examples, exceptions,...
Info. extraction templates, dialogue segments and
pegs, filled-out forms, high-level
connections,...
Descriptions augmented with prototypical
examples exceptions, problem-solving steps
and substeps, ...
More formal More concrete More introspectible
Alternative formalizations (KIF, MELD, RDF,),
alternative views of the same notion (e.g.,
what is a threat)
((( )) ())))
(defconcept bridge ()))