Probabilistic Dialogue Models for Dynamic Ontology Mapping

Introduction Agents communicate to perform tasks that they cannot accomplish alone. To communicate means to exchange messages that convey meanings encoded into signs for transmission. To understand a message, a receiver should be able to map the signs in the messages to meanings aligned with those intended by the transmitter. Agents should agree on the terminology used to describe the domain of the interaction: ontologies specify the terminology Having a shared ontology can be a strong assumption in an open environment: Introduction In this sort of environment, communication implies ontology mapping. In an open environment, it is impossible to know which agents will take part in the interactions. Agents have to map ontologies dynamically when needed. However, agents may meet infrequently and only for interactions on specic topics. A full ontology mapping would be a waste of resources: only the terms that are needed for the interaction should be mapped Ontology Mapping and interactions Usually, a term t j ∈ O 1 is mapped to another term w i ∈ O 2 comparing the term t j with all the terms in O 2 However, during an interaction it is often possible to avoid all these comparisons. A term in a received message is unlikely to refer to an entity completely unrelated with the context of the dialogue. Intuitively, the type of interaction, the specic topic and the messages already exchanged bind the content of a message to a set of possible expected entities. Probabilities are often used in ontologies to express the uncertainty of static relation between entities In this work, probability is used to predict the content of received messages in specic interactions. The repetition of similar interactions provide the information needed to compute the probability distribution of the entities. The predictions are used as suggestions for an Ontology Mapping system that must match foreign terms with local ones. The use of suggestions: Reduce the number of comparisons to perform, improving its eciency Reduce the ambiguities, excluding terms that are unrelated to the interaction. Introduction General concepts Learning the assertions Using the assertions Computing the probabilities of terms Conclusion General Structure The predictor learns the probabilities of terms received as feedback from the Mapper The predictor provides suggestions to the Mapper Introduction General concepts Learning the assertions Using the assertions Computing the probabilities of terms Conclusion Predicting the terms in messages Suppose an agent receives a message m …