An Optimizing Method for Structuring Inferentially Linked Discourse

In recent times, there has been an increase in the number of Natural Language Generation systems that take into consideration a user's inferences. The statements generated by these systems are typically connected by inferential links, which are opportunistic in nature. In this paper, we describe a discourse structuring mechanism which organizes inferentially linked statements as well as statements connected by certain prescriptive links. Our mechanism rst extracts relations and constraints from the output of a discourse planner. It then uses this information to build a directed graph whose nodes are rhetorical devices, and whose links are the relations between these devices. The mechanism then applies a search procedure to optimize the traversal through the graph. This process generates an ordered set of linear discourse sequences, where the elements of each sequence are maximally connected. Our mechanism has been implemented as the discourse organization component of a system called WISHFUL which generates concept explanations.