Upload
kenia-orum
View
219
Download
2
Tags:
Embed Size (px)
Citation preview
Exploiting a Thesaurus-Based Semantic Net for Knowledge-Based
Search
Peter ClarkJohn ThompsonLisbeth Duncan
Heather Holmback
Knowledge SystemsBoeing, Mathematics and Computing Technology
Overview• Problem: searching for information
– in particular, for human experts• Approach:
– Search using concepts, not words– Use a thesaurus as the initial ontology– Enhance it using simple AI techniques
• The Application: – Two deployed “Expert Locator” applications
Overall Picture
SearchEngine
Query words
“tube placement”
DatabasesHumanExperts
Web pages Documentrepositories ...
Problems with word searches..• Words have many senses (polysemy)
– e.g. “plane” finds both airplanes and geometry• Many words mean the same thing (synonymy)
– e.g. “tail fin” misses “vertical stabilizer” • Lack of world knowledge
– e.g. “jet engine” misses “propulsion systems”
Goal: organize search around concepts, not words
Need a conceptual vocabulary (“ontology”)
The Ontology Bottleneck• Massive up-front cost to build an ontology
• Use a technical thesaurus, enhanced with AI techniques
• Boeing’s Thesaurus:
– Highly customized to aerospace and Boeing
– Massive knowledge repository • 37,000 concepts, 18,000 synonyms
• 100,000 relationships (3 types)
– Many person-years investment of effort
The Approach
A (tiny) fragment of the ontology...
Jetengines
flameout
combustion
Burningrate
afterburning
Ramjetengines
Hydrogenfuels
enginesPropulsion
systems
thrustliftTurbojetengines
Enginestarters
Flamestability
Combustionstability
Flamepropagation
Pneumaticequipment starting
ignition
sprayJet spray
Converting Words to Concepts
Jetengines
flameout
combustion
Burningrate
afterburning
Ramjetengines
Hydrogenfuels
enginesPropulsion
systems
thrustliftTurbojetengines
Enginestarters
Flamestability
Combustionstability
Flamepropagation
Pneumaticequipment starting
ignition
• Search word: “jet”
sprayJet spray
?
?
?
?
Matching Query and Target Concepts
Jetengines
flameout
combustion
Burningrate
afterburning
Ramjetengines
Hydrogenfuels
enginesPropulsion
systems
thrustliftTurbojetengines
Enginestarters
Flamestability
Combustionstability
Flamepropagation
Pneumaticequipment starting
ignition
• Semantic distance between “ignition” and “jet engines”?
sprayJet spray
• 100,000 links are not enough!– 40% of concepts are “orphans”
• But: Many concept names are phrases– Can add links by analyzing these phrases
Enhancing the Thesaurus:1. Increase connectivity using subsumption
Space Shuttle Main Engine Enginegeneralization
Space Shuttlerelated-to
Subsumption Computation Algorithm
Space Shuttle Main Engine
1. Compute all possible generalizations by “word chopping” and “word generalization”...
Engine
Space Shuttle Engine
Space Engine
Space Vehicle Main Engine
Space Shuttle Main Space Shuttle
Space VehicleSpace
Shuttle
VehicleVehicle Engine
Vehicle Main Engine
Vehicle Main
Space Shuttle Main Engine
Space Shuttle Engine
Space Engine
Space Vehicle Main Engine
Vehicle Main Engine
Space Shuttle Main
Space VehicleSpace
Shuttle
Vehicle Engine
Engine
Space Shuttle
Vehicle
Subsumption Computation Algorithm2. Identify existing Thesaurus concepts and links within these
Vehicle Main
Space Shuttle Engine
Space Engine
Space Vehicle Main Engine
Space Shuttle Main
Space VehicleSpace
Shuttle
Vehicle Engine
Engine
Space Shuttle
Vehicle
Space Shuttle Main Engine
Subsumption Computation Algorithm3. Add missing connections to nearest existing concepts
Vehicle Main Engine
Vehicle Main
MeasuringInstruments
Equipment
OpticalMeasuring
Instruments
DistanceMeasuringEquipment
Range Finders
Optical Range Finders
Halogen Compounds
Fourine Compounds
NitrogenFourine
CompoundsFourides
Nitrogen Flourides
Some Example Inferred Links
• 21,000 generalization/specialization and 37,000 related-to links added
• Number of “orphans” down from 40% to 13%
Metal Tube Metalmade-ofNew:
Enhancing the Thesaurus:2. Use NLP to refine the “related-to” links
Metal Tube Metalrelated toCurrent:
• 27 relationship types chosen (causes, location, …)• heuristic noun-noun rules selects relationship, e.g
For compound “X Y” (e.g. “metal tube”):IF X is a MaterialAND Y is a Physical-ObjectTHEN Y made-of X
• Can use relation type to help compute semantic distance
Definition: “Flap: A movable airfoil attached to an airplane’s wing, and used to increase lift or drag.”
Flap isa: Airfoil attribute: Movable attached-to: Wing part-of: Airplane purpose: Increase
object: Lift, Drag
NLP
Flap
Airfoil
Airplanert
bt
Wing
Lift
DragIncrease
Increase
Movable isaattribute
purpose
purposeobject
object
attached-topart-of
Enhancing the Thesaurus:3. Knowledge from Text
Status and Evaluation• The Applications
– Two “Expert Locators” deployed and in use
– Sustained usage (~20 searches / day)
– Plans to quickly expand them further• more experts
• also cover projects and work groups
• add in attribute filters (years at Boeing, location, …)
• How do the Thesaurus Enhancements Affect Search?
– Study: Expert assessed relevance of “hit” concepts
– Recall increased (44% 75%) with only minimal effect on precision (58% 57%)
Discussion• “Number N of links” “relevance”?
– only for very small N!• The useful bias of a domain-specific Thesaurus:
– only contains relevant concepts• massively reduces errors in Thesaurus enhancement
– only contains relevant links• provides very domain-specific search
• Limitations:– ignored “quality” of expert, social issues, etc.– what if the concept you want isn’t there?
• Generality: Applies to any resource, not just experts
Summary• Search using concepts, not words
• Use of a thesaurus as an initial ontology:
– Can leverage many years of work by librarians
– Made viable using simple AI techniques of• search
• subsumption computation
• language processing
• Domain-specific thesauri provide valuable bias