As a way of making a crossover operation conceivable, our first step has been to reduce the planning problem to that of planning the sequential order of the facts (in a way that echoes Marcu's approach to some extent). We have done this by making certain restrictions on the RS trees that we are prepared to build. In particular, we make the following assumptions:
The algorithm for building a tree from a sequence essentially makes a
tree that can be processed by a reader with negligable short-term
memory. The tree will be right-branching and if
the reader just remembers the last fact at any point, then they can follow the
connection between the text so far and the next fact
.
Interestingly, Marcu uses ``right skew'' to
disambiguate between alternative trees produced in
rhetorical parsing. Here we are setting it as a much harder constraint.
The only exception is ``joint''
relations, which can join together texts of any size, but since there is
no real relation involved in them there is no memory load in interpreting
them.
The first two assumptions above make fundamental use of the order in which facts will appear in the text. For simplicity, we assume that every relation has a fixed order of nucleus and satellite (though this assumption could be relaxed). The approach is controversial in that it takes into account realisation order in the criterion for a legal tree. It is likely that the above assumptions will not apply equally well to all types of text. Still, they mean that the planning problem can be reduced to that of planning a sequence. The next experiments were an attempt to evaluate this idea.
