TEXT UNDERSTANDING FROM DISCOURSE MODEL AND INFERENTIAL PROCESSES

Text Understanding (hence TU) is viewed here as a linguistically-based process which activates Knowledge of the World only when needed, by triggering inferential processes and tapping encyclopaedic information. For a system based on linguistic principles, TU is a process requiring a pipeline of interleaved independent modules which deliver to the Semantic Processing Modules the preferred structural representation according to parsing strategies. These strategies can in turn access semantic information limited though to lexical and dictionary lookup. Semantic interpretation requires an intermediate module for Discourse processing, which takes as input the DAG structure produced by the parser, and the Discourse Model as asserted by the system, with the task of choosing the most relevant entity and make it the object of referential processes. A Centring-like algorithm is used to generate a stack of possible Topics of Discourse and then choose the one which fits best the current sentence. Semantic processes vary according to the referential nature of nominal expressions. We use a children story and other texts taken from newspapers to show how the TU system works.