🚀 go-pugleaf

AI Seminar:  2PM, Tuesday March 29, Sandford Fleming 1105


               PLANNING COHERENT MULTISENTENTIAL TEXT 


                            Eduard Hovy 

                Information Sciences Institute of USC 
                       4676 Admiralty Way 
                   Marina del Rey, CA 90282-6695 


Generating multisentential text is hard.  Though most text generators 
are capable of simply stringing together more than one sentence, they 
cannot determine coherent order.  Very few programs attempt to plan 
out the structure of multisentential paragraphs. 

Clearly, the key notion is coherence.  The reason some paragraphs are 
coherent is that the information in successive sentences follows some 
pattern of inference or of knowledge with which the hearer is familiar, 
so that the hearer is able to relate each part to the whole.  To signal 
such inferences, people usually link successive blocks of text in one 
of a fixed set of ways.  The inferential nature of such linkage was 
noted by Hobbs in 1978.  In 1982, McKeown built schemas (scripts) for 
constructing some paragraphs with stereotypical structure.  Around 
the same time, after a wide-ranging linguistic study, Mann proposed 
a relatively small number of intersentential relations that suffice to 
bind together coherently most of the things people tend to speak about. 

The talk will describe a prototype text structurer that is based on the 
inferential ideas of Hobbs, uses Mann's relations, and is more general 
than the schema applier built by McKeown. The structurer takes the form 
of a standard hierarchical expansion planner, in which the relations 
act as plans and their constraints on relation fillers (represented 
in a formalism similar to Cohen and Levesque's work) as subgoals in the 
expansion. The structurer is conceived as part of a general text planner, 
but currently functions on its own. It is being tested in two domains: 
database output and expert system explanantion.