I have a quick look at the classification, but I dont
take time to re-lable the sentences from the view of 
target generation: I think if you catch the grammatical
segment of the sentence, it is possible to generate the
target language by writing the "reverse" grammar(of course,
chinese grammar is different from english grammar, but I 
think most of them could be done.) the problem here is that
we need to keep the generation grammar as general as possible.

As to the partially parsed sentences, I think we should make
an agreement on the annotations of the semantic roles of the 
chunks, for example, MAIN-CLAUSE, SUB-CLAUSE, TIME, REASON, 
LOCATION, or simply CLAUSE1, CLAUSE2, ets. Is it possible to
give such information during parsing? If so, the chunks could
be mapped into target sentence according to the feature of 
the target language. Another problem is how we deal with the 
omitted words(if there are some) during generation, omit them
or put them into the target sentence based on some strategy.

Ambiguity is another problem, especially word sense ambiguity
in generation, I encounter this problem when started looking into
Wei's previous work, I think we should do some disambiguity in
our system. the question is: disambiguity in parsing or generation,
or both. Maybe some words like "plane(aircraft, surface)" could
be sovled in parsing(again the same question, is it possible for 
you), and some words like "for(FOR a long time, FOR me)", 
"cover(COVER the surface, cover the fact)" could be sovled in 
generation, that is a problem of lexicon selection.

Dr. Li reduced the 3-step generation to 2-step generation when 
he constructed the Chinese sentence generation system, I don't
know whether this will hurt to disambiguity. He mentioned lexicon
selection for disambiguity in his report, but I don't know whether
he implement it in his work, I mean, the work related to 
Caterpillar domain. Maybe Eric could tell me more.

Statistical method will work in disambiguity, but it requires a 
large amount of data. In our case, most of the previous work is
done by KB method, so I think it is better to introduce a word
sense hierarchy(semantic hierarchy) to our system, at least to
the generation system. The hierarchy can be simple(but complete)
now, so it is possible fo manual work in a demo system. In the
long run, if a practical Chinese generation system will be
constructed, it is worth to do so.

Sorry, Krzysztof, many words here are written to Eric, I just want 
to refer to your opinion, anyway, the parsing and generation are not 
entirely seperated.

In your classification of the "Marke Wrap" file, at least half 
sentences are labled TM or KB, do you get some interlingua of them,
if so I can take them to have some tests.