ppt

Ontology Learning Tools: A Survey of Existing Tools

Patrick Cash

Outline• Ontology learning• Ontology learning from text

– Learning from text alone– Learning from text and other resources

• Ontology learning from structured data– Learning from a machine readable dictionary– Learning from existing ontologies

• Conclusion

Ontology Learning• Ontology

– An explicit formal specification of a shared conceptualization of a domain

• Facilitates knowledge sharing and machine understanding of knowledge

– Manually building ontologies is a tedious task that becomes a bottleneck to knowledge acquisition

• Ontology learning techniques were created to address this problem

Ontology Learning• Uses machine learning and other AI techniques

to learn ontology structure from input data• Fully automatic ontology learning remains in the

distant future– Most tools are semi-automatic and require human

(expert) intervention using cooperative learning approaches for ontology building

Ontology Learning from Text• Learning from text alone

– Natural language processing of the input text to find lexical and syntactic structure

– Machine learning algorithms used to derive ontology structure out of this structure

• Clustering – using user supplied similarity measure• Rule/template base knowledge extraction• Workbench tools allow the use of multiple algorithms

Ontology Learning from Text• Problems with these tools

– Requires large amounts of user interaction in validation

• Cooperative learning approach• Not practical for large scale ontologies

– Requires hand created rules and operators– Requires large amount of “high quality input” for

domain coverage and concept learning

Ontology Learning from Text• Learning from text and other resources

– Natural language processing of the input text to find structure and extract domain keywords

– Machine learning algorithms used to derive ontology structure out of the lexical and syntactic structure

• Clustering, Rule/template base techniques– Uses structure in input ontologies like WordNet

• Uses collocation and other attributes from the input ontology to form the new ontology

Ontology Learning from Text• Problems with these tools

– Many of the systems are rule based and require many hand created rules

• Not practical for large ontologies– Depends on enough of the ontology domain being

represented in the input ontology, taxonomy or knowledge base

• Assumption will often not hold for technical or specialized domains

Ontology Learning from Structured Data• Learning from a machine readable dictionary

– Takes as input a machine readable dictionary and a set of domain keywords

– Uses structure in machine readable dictionary• Creates initial ontology using information from machine

readable dictionary• Adds domain keywords to the initial ontology• Trims initial ontology to create a domain specific ontology

Ontology Learning from Structured Data• Problems with these tools

– Depend on enough of the ontology domain being represented in the machine readable dictionary

• Assumption will often not hold for technical or specialized domains

• Matching to dictionary is based only on simple text or regular expression matching

Ontology Learning from Structured Data• Learning from existing ontologies

– Takes as input an existing ontology and enhances it• Uses classification techniques of the input data to determine where it

is added to the input ontology– Supervised learning techniques with labeled training data

– Takes as input two or more ontologies and translates or merges them by mapping between them

• Simple text matching or mapping rules• Iterative labeling using statistical machine learning• Uses classifiers to map instances from one ontology into the other

Ontology Learning from Structured Data• Problems with these tools

– Makes several assumptions about the structure of the input data

• Several of the tools are based on extraction of knowledge from HTML pages with specific structures (ex. Forms)

– Requires large amount of upfront work done by information provider

• Translation into a tool specific representation• Explicit mapping of input data’s structure using schemas

and other hooks into the input knowledge base

Conclusion• Selecting an ontology learning tool

– Types of input available– Types of knowledge learned

• Combined workbench/framework – The most robust tools are workbench or framework

based using a modular architecture so that different learning techniques can be used in different use cases

– A common representation will be needed for tools to work together

Conclusion• Ontology Validation

– Approaches using in current research• Validate against a gold standard ontology created by an expert

– Not practical for large ontologies• Using machine learning validation techniques to validate the

learned ontology– Precision: ratio of relevant terms retrieved over the entire

number of terms in the ontology – Recall: ratio of relevant terms retrieved over the entire number of

relevant terms

Conclusion• Ontology learning techniques are still a long way

from a fully automatic solution• Ontology learning techniques are necessary to

make wide spread ontology use and the possibilities that entails practical– Tools that implement these techniques in a user friendly

way are necessary for making these techniques available to non-expert users creating ontologies

Questions ?