My research deals with the problem of Relevance Formalization. Considering a particular situation (defined by a goal, an environment and a knowledge) the problem is to select the pieces of information which are relevant to this situation. The problem of relevance formalization is ubiquitous in Artificial Intelligence. Thus, a lot of problems in this research domain can be described as selection problems. Of course, the Search Engines on the web which aim at proposing documents relevant to a query come to mind. But a lot of other problems like assigning a class to an object, using the appropriate sense of a word in an automatic translation or choosing a move in a chess game, also address the same problem.
My research aims at designing a system sufficiently general to be applied to different selection problems. Taking inspiration from Cognitive Models, I have developed a system called Echo, coherent with relevance formalization proposed in Information Retrieval. Echo, in a Machine Learning step, builds links between differents pieces of information (links between words and documents, between features and classes, between words and senses, between positions and moves). Echo, In a selection step, uses these links using a Spreading Activation method and computes a kind of « echo » in order to select the relevant pieces of information in a particular situation. This system which is based on simple neural mechanisms has been successfully applied to different types of problems especially in a text processing context.
Examples of selection problems in a text processing context
My research follows an Enactive approach in which relevance is defined relatively to the identity of the judging system which itself depends on its historical coupling with its environment. The notion of echo which underlies the selection process (in the Echo system) is very closely related to the interaction loop between the system and its environment.