An Adaptive and Distributed Framework for Advanced IR

It has been often noticed that modern IR ((Gregory, 1991), (Alan, 1991)) should exhibit capabilities that are sensitive to the document content, integrate interactivity, multimodality and multilinguality over a large scale and support the very dynamic nature of the current needs for information access (so to be adaptable to chanes of the sources, language and content/style). This paper discuss the architectural design aspects of TREVI (Text Retrieval and Enrichment for Vital Information - ESPRIT project EP23311), a distributed Object-Oriented Java/CORBA driven system for NLP-driven news classification, enrichment and delivery. The advanced features of TREVI include the extensive use of a well defined model ((Mazzucchelli, 1999)) based on a typed mechanism for static/dynamic control of the distributed process and on a principled representation of linguistic types into computational OO data structures and the adaptivity of the employed linguistic p rocessors (namely, the robust and lexically-driven parser). The original aspect of TREVI is its novel combination of a systematic design approach with the contribution of advanced and adaptive NLP processors for content-driven text classification. A full toolkit system was developed within the operational scenarios related to three different users (i.e. three different information providers), in two languages (English and Spanish) and its good performances (classification accuracy and usability) are basic evidences of the success of this approach.