In designing the parameterized action representation (PAR), several
issues from the natural language perspective guided our choices. The
action representation scheme is an intermediate representation language
between the animation constructs, i.e. PaT-Nets, and natural language
instructions. It allows a hierarchy of actions to be represented
explicitly (by listing subactions for an action). This hierarchy is
needed in order to generate natural language at the correct description
level. PaT-Nets can also form a hierarchy of structures. In fact, PAR
shares enough in common with PaT-Nets that the conversion between PAR
and PaT-Nets is relatively straight-forward. However, one aspect in
which they significantly differ, in terms of representing actions, is
that PaT-Nets require instantaneous transitions from one node in a
network to another whereas the corresponding concept in our
representation, i.e. culmination conditions, does not have this
restriction. Culmination conditions are represented as queries to
sensors, which as described in Section 
can have
significant duration, even to the point of having subactions of their
own. Thus, the culmination conditions in our representation might be
translated into PaT-Nets that need to be executed in order to check
whether the culmination conditions hold. In generating instructions, we
will not usually generate text corresponding to the internal structure
of the sensing process, so we do not need it in our action
representation and can leave the details of expanding the sensor queries
to a lower level.
The overall picture of how PaT-Nets, our action representation, and
natural language all fit together is shown in Figure . The
control algorithms for the (a) and (d) transitions are described in
Section and Figure in Section
, respectively. (The (b) and (c) transitions are not
currently relevant, but they are certainly kept in mind.)
Figure: The action representation as an intermediary between PaT-Nets
and natural language.
Task simulations from which to generate natural language will be restricted to be in PAR form by providing an authoring system that enforces the specification requirements of the action representation. Through the conversion of PARs to animation, animations of the task simulations can be done while preserving the structure that is suitable for generating natural language.