Experiments with Discourse Structure

An algorithm was implemented for automatically generating tree-like discourse structures. This was used to generate a variety of discourse trees by using diierent attachment rules. The information used by these rules was essentially referential, and included continuity of theme, temporal reference, and pronominalisation in general. The trees in turn formed the basis for reference resolution, including long-distance pronominalisation. The motivation for this approach to discourse structure was twofold. Firstly, there is the implausibility of assuming that reference resolution has available to it a discourse seg-mentation based on higher-level reasoning processes. Secondly, there is the ready availability of referential information and the associations of it, and theme in particular, with what the discourse is about.