key-sun choi and yeun-bae kim* kaist korterm, nhk strl* [email protected]
DESCRIPTION
Terminology, Interactive Media and Standards - A Scenario toward Personalized Interactive Knowledge Service s -. Key-Sun Choi and Yeun-Bae Kim* KAIST Korterm, NHK STRL* [email protected] http://www.korterm.org/. Outline. Integrating the multiple number of lexical knowledge base - PowerPoint PPT PresentationTRANSCRIPT
Korea Terminology Research Center for Language and Knowledge Engineering
Terminology, Interactive Media and Standards - A Scenario toward Personalized Interactive K
nowledge Services -
Key-Sun Choi and Yeun-Bae Kim*KAIST Korterm, NHK STRL*
http://www.korterm.org/
Korea Terminology Research Center for Language and Knowledge Engineering
Outline
Integrating the multiple number of lexical knowledge base
Question-answering for what-, and why-type question
By causality probingIntegration of Video clipping of answer
segments
Korea Terminology Research Center for Language and Knowledge Engineering
Steps to Give Appropriate Answering to Each Customer Who Asks about BSE
BSE ?
Since it was identified in the mid-1980s in Britain, mad cow disease, or BSE, has resulted in the slaughter of millions of cattle -- and the deaths of dozens of people from the related brain-wasting disease known as...
Term: BSEBT: diseasesymptom: Sponge-like change of a brain partcause: infectious-prionsynonym: mad cow
(5) template: term “BSE”consists of characteristics for BT, symptom and casuse.
diseaseBSEcow disease
mad cow
disease
noun
infectious
Prionnervous ..
BT
synony
mPO
Scause
sympto
m
mad
related
(7) representation
…a fatal disease of cattle affecting the nervous system, resembling or identical with scrapie of sheep and goats, and probably caused by a prion transmitted by infected tissue -- an infectious protein particle similar tovirus …
(9) definition for semi-expert
a brain disease of cows that causes death, and some people who eat them…
(9’) definition for child
(8) (2) unit of
(3)
no item in dictionary when
the first BSE
BSE ?
Since it was identified in the mid-1980s in Britain, mad cow disease, or BSE, has resulted in the slaughter of millions of cattle -- and the deaths of dozens of people from the related brain-wasting disease known as...
(1) Query
Term: BSEBT: diseasesymptom: Sponge-like change of a brain partcause: infectious-prionsynonym: mad cow disease
(4) Acquisition(5) Templateterm “BSE”consists of characteristics for BT, symptom and casuse.
diseaseBSEcow disease
mad cow
disease
noun
infectious
Prionnervous ..
BT
synony
mPO
Scause
sympto
m
mad
related
diseaseBSEcow disease
mad cow
disease
noun
infectious
Prionnervous ..
BT
synony
mPO
Scause
sympto
m
mad
related
(7) Representation
…a fatal disease of cattle affecting the nervous system, resembling or identical with scrapie of sheep and goats, and probably caused by a prion transmitted by infected tissue -- an infectious protein particle similar tovirus …
a brain disease of cows that causes death, and some people who eat them…
(9’) Definition for Child
(6) Integration
(8) Presentation(2) Unit of Database
(3) Wrapping
no item in dictionary when
the first BSE
(9)Definition for Semi-expert
Korea Terminology Research Center for Language and Knowledge Engineering
Cow disease database
D1D2
world space (a1)
(b1) cow disease relates to human disease
(b4) mad cow disease causes human brain disease
(b3) human disease causes cow disease
D4D3
disease
cow diseas
e
knowledge space (a2)
human brain disease
(a3) justification
Referent mapping
(a4)
BSE
disease food
disease beef
disease vegetable
disease meat
(b2) cow disease causes human disease
disease cow
disease
animal
disease plant
edible (d2)
Disease living thing
justification cause (d2)
disease
human
infected Referent
mapping (d2)
(d1) justification instance
human disease
(b) hypotheses
(c1) ontology for disease
(c3) ontology for living thing
Human disease
database
Justification process goes from the world space to the knowledge space
Korea Terminology Research Center for Language and Knowledge Engineering
Does mad cow disease cause human brain disease?an enlarged information space
H2: cause(mad cow disease,human brain disease)
H3: cause(mad cow disease,human brain and neural disease)
H4: cause(mad cow disease,human brain and non-neural disease)
…
…
media on H3 article on H2 media on H1article on H4 …
Worldspace
Knowledgespace
justification
opposition
specification
H1: relate(cow disease,human disease)
hypothesisuniversal hypothesis
specification
generalization
absurd hypothesisreferent
Korea Terminology Research Center for Language and Knowledge Engineering
Construction of knowledge space
How to construct the knowledge space and
its linkage to the associated justification world space By
Experienced humans or
(Semi-)automatic machines Knowledge Discovery from Text
Korea Terminology Research Center for Language and Knowledge Engineering
Tell me whether a mad cow disease will cause a human brain disease.
Where is relevant resources in response to a given knowledge request?
How to meta-data catalogs by categorizing the information content in each repository
Information Seeking Problem
Korea Terminology Research Center for Language and Knowledge Engineering
Does the “mad cow disease” cause the “human brain disease”?
What is the correlation between “mad cow disease” and “human disease”?
Where is the related repository? the human disease repository, and
The mad cow disease repository.
Knowledge Seeking Problem
Korea Terminology Research Center for Language and Knowledge Engineering
Tell me whether a mad cow disease will cause a human brain disease.
How to correlate concepts- Find possible relationships between animal
disease and human disease Link a animal disease record with the corresponding
a human disease report. What is shared ontology?
Disease, virus, … What is related contextual ontology?
Food chain, time period How to standardize that?
ISO/TC37 MPEG7
Korea Terminology Research Center for Language and Knowledge Engineering
Knowledge Seeking Model
When you have a pneumonia, you should be rest and take an antifebrile if you have a high fever ….
When you have a pneumonia, you should be rest and take an antifebrile if you have a high fever ….
The new corona virus is the leading hypothesis for the cause of SARS . The primary way that SARS appears to spread is by close person-to-person contact.
The new corona virus is the leading hypothesis for the cause of SARS . The primary way that SARS appears to spread is by close person-to-person contact.
I feel cold and suffer a high fever.
I feel cold and suffer a high fever.
What is the cause of SARS?What is the cause of SARS?No specific treatment recommendations can be made at this time. Empiric therapy should include coverage for organisms associated with any community-acquired pneumonia of unclear etiology, …..
No specific treatment recommendations can be made at this time. Empiric therapy should include coverage for organisms associated with any community-acquired pneumonia of unclear etiology, …..
Would you please tell me an effective remedy?
Would you please tell me an effective remedy?
OntologyOntology
nutritionblood Waste mattersecretion
bacteria physiologyblood/secretion/waste matter
Health caredrink/eat
sense
rest Physical disability
condition/abnormalityimperfect
disease injury
animal
animate
object
inanimate
drug
Human activityaction
life
Natural phenomenawork<abstract>
abstract
concrete/abstract
concrete
High feverOBJ
feelSUBJ
Cold
cause symptom remedy sufferlipappear
blister
ADVSUBJ
understand
pneumonia
Knowledge Seeking
Korea Terminology Research Center for Language and Knowledge Engineering
Knowledge Seeking
Necessity of Knowledge SeekingQuestions requiring -- "why", "what" , "how“ type questions
Problems of Knowledge SeekingJustification Probing on knowledge How to utilize various Knowledge resources
GoalBuilding an algorithm for
searching some topical paths in order to find causal explanations for questions
E.g., “Why do patients pay money to doctors?”
Korea Terminology Research Center for Language and Knowledge Engineering
Knowledge Representation
Two types
Tree ( Hierarchical )Graph
Korea Terminology Research Center for Language and Knowledge Engineering
Knowledge Representation
Implication
doctor human, *cure, #occupation, medical Instantiation
*cure {doctor, medical worker, …} Event Relation
cause(cure,suffer-from)cure.patient = suffer-from.experiencer
cure.content = cure.content
Converse/Antonymy
Converse(take,give)
Korea Terminology Research Center for Language and Knowledge Engineering
Knowledge Structure Knowledge structure : Multi-facet Structure
Knowledge space for each knowledge feature
Knowledge feature : e.g. isa(x,y)
give take
pay
*cure*cure
支払い
givegive take
Alter-possession
patient
doctor occupationoccupation money
cure
earn $earn$earn
#occupation#occupation
entity
event
earn
人間
why
role
question
cause
dictionaryConceptfacets
pay
human
converse
agent=possession=target=
agent=possession=source=
Korea Terminology Research Center for Language and Knowledge Engineering
Justification by Knowledge Seeking
(1) Doctor cures the patient and earn the money as occupation(2) ‘Pay’ is ‘give’. Converse of ‘give’ is ‘take’. ‘Patient’ is the agent (*) of ‘give’
and the source ($) of ‘take’. ‘Take’ is the hypernym of ‘earn’. “The patient pay X to the doctor.” is “The doctor earn X.”
(3) The hypernym of ‘pay’ is ‘take’. From “pay money”, the possession of ‘take’ is ‘money’.
patient doctor occupation money
$cure *cure earn $earn
#occupation
*pay $pay
give takeconverse
agent=patientpossession=target=
agent=possession=source=patient
entity
alter-position event
conversecrossing
(1)
(2)
(3)
doctor
doctormoneymoney
Korea Terminology Research Center for Language and Knowledge Engineering
Video indexing scheme based on natural language content description
T3
Vid
eo
Material
segm
entatio
n
T1
T2
T3
T1
T2
T3
“doctor cures patient”
“doctor earns”
parsin
g
SDS-type indexScript-type index
doctorcure patient
earn
subject
subject
object(from T1 to T2)
(from T2 to T3)timeT1 T2
doctorcure
patientearn
descriptivecomponents
Attach
ing
d
escriptio
n
Korea Terminology Research Center for Language and Knowledge Engineering
Gaps in possible syllogism Syllogism
If pay(human2,money,to-human1) is earn(human1,money,from-human2), and earn(doctor,money,from-human2),Then pay(human2,money,to-doctor).
Axiom-1Converse (antonymy) event role inter-relation
pay(agent=human2,content=X,target=human1) iff earn(agent=human1,content=X, source=human2)
Axiom-2If cure(doctor, human2) and occupation(doctor), then earn(doctor,money,source=human2)
Extended syllogismIf Axiom-2 and cure(doctor,human2) and occupation(doctor), then earn(doctor,money,from-human2).
Korea Terminology Research Center for Language and Knowledge Engineering
Gaps in Linguistic Knowledge Base Query:
Why doctor cure patient?Why doctor earn money?Why patient pay money to doctor?
Extended syllogismAxiom-2
If cure(doctor, human2) and occupation(doctor), then earn(doctor,money,source=human2)
If Axiom-2 and cure(doctor,human2) and occupation(doctor), then earn(doctor,money,from-human2).
Gaps in Linguistic Knowledge BaseNo axiom
Motivation of Crossover algorithm: causal linkingto link the gaps in linguistic knowledge baseto find possible axioms
Korea Terminology Research Center for Language and Knowledge Engineering
Motivation of Crossover algorithm
Motivation of Crossover algorithm: causal linking
to link the gaps in linguistic knowledge base
to find possible axioms
Korea Terminology Research Center for Language and Knowledge Engineering
Similarity: mission
Mission
for two nodes, check the connectability between two nodes
if sense ambiguities for a word,select the best sense.
if there are multiple expansion possibilities in the next step,
select the best node.
Final goal
to find a causal linkage
Korea Terminology Research Center for Language and Knowledge Engineering
Similarity: motivated example
For query: “Why does doctor cures patient?”similarity guides to find all kinds of possible situation between doctor and patient before/during/after “doctor cures patient”.
for example, Patient suffers from a disease. Doctor cures the patient. Doctor is an occupation. Occupation is to earn money for living. All properties of doctor (e.g., cure) is relevant to the
occupation. Curing is to earn the money. Doctor earns the money. Patient pays the money to the doctor.
Korea Terminology Research Center for Language and Knowledge Engineering
Similarity for causal connectability
How to find the relevance from linguistic knowledge base incl. role relation for:
patient ~ disease
suffer from ~ cure
earn ~ cure
pay ~ patient
patient ~ disease
role relation suffer from ~ cure
cause interrelation earn ~ cure
subject of event
relevance of occupation
pay ~ patient
converse event relationpay ~ earn
Korea Terminology Research Center for Language and Knowledge Engineering
Similarity Measure patient ~ disease
role relation suffer from ~ cure
cause interrelation earn ~ cure
subject of eventrelevance
of occupation pay ~ patient
path up and hypernym’s general properties’ inheritance
pay < actpatient < human < *act
patient ~ diseaseto find the same event feature.patient human $cure *sufferFromdisease medical $cure undesired
suffer from ~ cureto expand thru event interrelation.sufferFrom(exp=X,cont=Y) causes cure(agent=A,patient=X,cont=Y)
earn ~ cure*earn ~ *cure, *earn ~ $cure
to compare of role entities of events.doctor *cure #occupation
*cure ~#occupation if they are inside of the definition of ‘doctor’.
occupation affairs earn alive pay ~ patient
pay << act (hypernym)human *actpatient << human
path-up, to find role entity of event, path-down
Korea Terminology Research Center for Language and Knowledge Engineering
Rationale of Similar, but exhausted list must be checked.
1. feature similarsimilar(1)
2. expand similarsearch “rolerelation”
for event interrelation3. inverse and sister similar
a. expand inverse similarsimilar if their role entities share these events.(expanded version) similar if their inverse concepts share these two concepts.
b. expand sister and feature similarIf they are in sisters, they are similar, and plus if the sister (e.g., occupation) contains it as a feature. (crossover(1))
4. Algorithm: path-up + inverse + path-down
• path similarSimilarity between two hypernyms
1. patient ~ diseaseto find the same event feature.patient human $cure *sufferFromdisease medica $cure undesired
2. suffer from ~ cureto expand thru event interrelation.sufferFrom(exp=X,cont=Y) causes cure(agent=A,patient=X,cont=Y)
3. earn ~ curea. *earn ~ *cure, *earn ~ $cure
to compare of role entities of events.b. doctor *cure #occupation
*cure ~#occupation if they are inside of the definition of ‘doctor’.
occupation affairs earn alive4. pay ~ patient
pay << act (hypernym)human *actpatient << human
path-up, to find role entity of event, path-down
Link
Korea Terminology Research Center for Language and Knowledge Engineering
similar(Level)
Goalto merge all of similarity measures
path similar
feature similar
crossover similar (not yet fixed)
to give Level of SimilarityNotation: similar(Level)
Korea Terminology Research Center for Language and Knowledge Engineering
Level of Similarity
Motivation:
to be able to compare the similarity levels
to be dynamically adaptable to the knowledge levels
Levels of details
Level 0: based on itself and its top path= path similar
Level 1: to expand to the features=feature similar
Level 2:
Korea Terminology Research Center for Language and Knowledge Engineering
human*sufferFrom$cure
doctor cure patientwhy
human#occupation*curemedical
agentpatientcontentmedical
medicine*disease&medical#
cure>cause>sufferFrom{patient=experiencer,content=content}
cure>possible consequence>beRecovered{patient=experiencer}
similar level 0
commercial$earn*buy#sell$setAside
patient pay moneywhy
human*sufferFrom$cure
agentcontentsource
payer*money advanced$
give>consequence>lose{agent=possessor,possession=possession}
give<implication<receive{possessor=target,possession=possession}
doctor
give
hypernym
take
converse
hypernym
occupation affirsearn
human#occupation*curemedical
Virtual Knowledge Base: causal linking
Korea Terminology Research Center for Language and Knowledge Engineering
A Snapshot
Korea Terminology Research Center for Language and Knowledge Engineering
Experimentation
SourceConcept
Reached concept’s path size
path = 1 path = 2 path = 3
Cure 275 593 24,854
Eat 268 605 24,903
Study 276 358 23,172
Food 532 650 18,066
Human 6,713 3,686 51,171
Money 328 1,312 19,827
Q2: Why does researcher read textbook?Path: researcher → #knowledge → #information ← readings ← textbook
Interpretation: Researcher is related to the knowledge and the knowledge is related to the information. Textbook is the object of readings and readings is related to information. So researcher read textbook.
Connected concepts
Q1: Why do patients pay money to doctors? Path: patient → $cure → doctor → #occupation ← $earn ← money
Interpretation: Patients cured by doctor. The doctor is related to the occupation and the occupation is that to earn the money. So patients pay money to doctors.
Korea Terminology Research Center for Language and Knowledge Engineering
Future Works Resources
EDR, WordNet, CycHowNet based
Algorithm and RepresentationLexical ChainInterpretation as AbductionBayesian Belief Net
CausalityStopping ConditionWhat is “Causality” and “Explanation”?
Automatic Video Synchronization with Causal Justification Path
Korea Terminology Research Center for Language and Knowledge Engineering
Conclusion
How to link the already existing linguistic knowledge base To be tested for the definition and the causality link. To be adapted for the user knowledge level
to find more causality link. How to link the video archive to linguistic causal path
Dependency structure
MPEG7