NLG covers a wide variety of tasks and levels of representation but can generally be
seen as performing mappings between representations.

To some extent, the NLG community can try to reuse existing shared tasks, for
example those for parsing (e.g. the Penn treebank [Langkilde, INLG-2002]) or
information extraction [Varges&Mellish, NAACL-2001]. This should be pursued but
there are limits due to the level of detail required as input to generation: IE
tasks, for example, tend to extract much less information than is needed to
regenerate the original texts.

Our proposal is to take the diversity of NLG tasks as an opportunity to do
things differently from the shared tasks of NLU: let us collect a large number
of small tasks, in the way they are created in typical research projects
(including PhDs). We should not force anyone to stick to predefined,
`standardized' levels of representation but use corpora of aligned
representations more or less as they arise in those projects (and publish them
in a neutral, XML-based format).

Let us evaluate our systems not only on how well they perform on individual
tasks but also on how well they can

1. switch between tasks, possibly while the system is running,
2. learn from small corpora / limited resources (if learning is involved),
3. connect several smaller tasks into larger ones.

We should focus on flexibility and adaptivity rather than on monolithic,
large-scale tasks. (However, a purely machine-learning based approach may still
see all shared tasks just as input-output relations that can be learned
together.) Item (3) above is the most advanced scenario: can we combine a
content selection model derived for a restaurant selection task with a sentence
planner derived from a museum explorer and a realizer for newspaper texts? What
additional mappings between intermediate representations would be required?