ANEMONE: An Effective Minimal Ontology Negotiation Environment AAMAS 11 May 2006 Utrecht University, the Netherlands Jurriaan van Diggelen, Robbert-Jan Beun, Frank Dignum, Rogier M. van Eijk, John-Jules Meyer
Problem statement Approach Protocols Case study Conclusion problem Layout of presentation
Inform • … • content:NBA news(#93) • … • ontology:sports • … problem Common ontologies News Agent 1 News Agent 2 Knowledge base Knowledge base Ontology : sports Ontology : sports NBA news Boxing news NBA news Boxing news ABox ABox NBA news(#93) Boxing news(#185) NBA news(#93)
Inform • … • content:NBA news(#93) • … • ontology:sports • … problem What if the receiver doesn’t know the ontology? ? News Agent 1 News Agent 3 Knowledge base Knowledge base Ontology : sports Ontology : news NBA news Boxing news Basketball article Science article ABox ABox Boxing news(#185) NBA news(#93) Currently, this is a burning problem in heterogeneous systems such as the semantic web and open multi agent systems
problem Ontology reconciliation Standardization: Make all agents use the same standardized ontology Pre-defined mappings: Define all ontology mappings before the agents start interacting Ontology negotiation: Enable the agents to resolve ontology problems themselves. Add ontology alignment layer to the agent communication protocol. How do we solve the babylonian confusion of tongues?
ANEMONE approach AN E ffective M inimal O ntology N egotiation E nvironment
QuestionANEMONE’s answer What? Effective & Minimal ontology When? Lazy alignment Where? Decentralized alignment approach Design objectives of ANEMONE • Ontology negotiation serves to build up a • shared ontology that enables agents to reach the • desired level of coordination -
approach Ontologies in ontology negotiation Ag-1 Ag-2 T native T i shared native c n f n c f acquired h d j m l shared acquired g m e k d T T • Acquired concepts are not used for KR and reasoning. • Shared concepts are concepts that are common knowledge
protocols Layered communication protocol NCP Normal Communication Protocol CDP Concept Definition Protocol CEP Concept Explication Protocol NCP deals with social interaction that agents normally exhibit in semantically integrated systems. CDP deals with concept learning using concept definitions. CEP deals with concept learning using pointing to instances.
+ + How to compose and interpret a message? or How to be effective with a minimal shared ontology How to recognize overgeneralization? or How to establish lazy ontology alignment protocols Overview Sender Receiver Send message Interpret message Otherwise NCP Message overgeneralized? Message not understood? Send definition Interpret definition CDP Definition understood? Definition inadequate? CEP Send explication Interpret explication
native shared native acquired shared acquired d k d protocols How to be effective with a minimal ontology? (1) Ag-1 intends to convey d T Ag-2 sends d T i c n <Inform d> f f n receives d c h g interprets as k j m m l e <OK> T T Note: Receiver translates to native concept
native shared native acquired c shared acquired d e k d protocols How to be effective with a minimal ontology? (2) Ag-1 intends to convey e T Ag-2 sends d T i c n <Inform d> f f n receives d h g interprets as k j m m l <OK> T T Note: Sender becomes more general in message composition. This is required to obtain effective communication using a minimal ontology. How to prevent overgeneralization?
native T shared native f acquired f shared acquired g g d e k d protocols How to establish lazy ontology alignment? (1) intends to convey g Ag-1 sends f T Ag-2 <Inform f> i c n receives f n interprets as T c h checks its knowledge m m l <ReqSpec> T <Inform g> T <StartCEP> <Explicate g> interprets as g as l <OK> Note: Receiver recognizes overgeneralization.
c c j protocols How to establish lazy ontology alignment? (2) intends to convey j Ag-2 Ag-1 T T sends c i <Inform c> n f n receives c f h d interprets as c m l g checks its knowledge m e k d derives implicature T T <OK> native shared native Discussion: Receiver correctly recognizes lossless communication acquired shared acquired
case study Case study: RSS news feeds Heterogeneous RSS news providers proliferate on today’s internet.
Agents are wrappers around RSS aggregators Every agent represents one news provider Agents exchange news articles with each other. In CEP, the speaker sends positive and negative examples of the unknown concept. The receiver classifies these and derives the relations with its native concept. Agent’s concept classifiers are implemented using text classification techniques based on support vector machines. case study Application
case study Example: Initial situation Ag-M Moreover Ag-B BBC Article Article Boxing Tennis Business UK Science/Nature Basketball Reuters Ag-R Ag-Y Yahoo Article Article Sports News Science News Sports Science Business News NBA WNBA Every agent collected news for 2 months
Science News case study Example: Situation after 9 dialogues • The agents made use of: • Non-equivalence mappings • Generalization in message composition • Detecting information loss • Speaking in unknowingly shared concepts • Speaking in acquired concepts • Exchanging concept definitions Moreover BBC Ag-B Ag-M Article Article Science News Tennis Business UK Sports Boxing Basketball Science/Nature NBA Ag-R Yahoo Ag-Y Reuters Article Article Business News Sports Science Sports News Sports Science News Business News NBA WNBA
ANEMONE’s features are useful in a realistic domain. CEP needs improvements to apply ANEMONE to more complex domains. When larger systems are involved, macro issues become important. Strategies required to make ontologies globally shared. conclusion Conclusion