
The complete set of examples discussed in Section 5 of the paper
appear in the files bleu_wins.txt and bleu_losses.txt.

In the files, annotations with a plus (+) indicate cases where the SVM
ranker improved the realization choice, while annotations with a minus
(-) indicate cases where the ranker made the choice worse, and
annotations with an equals sign (=) indicate cases where the change
from the baseline perceptron model choice is neither for the better or
worse. 

The plus, minus and equals annotations are combined with the following
overall indicators:

adq - adequacy
flu - fluency
ok  - neither better nor worse (only used with equals)

There's also an annotation, !bad, indicating that neither choice is
acceptable. 

For each change, there are also linguistic categories assigned, as
follows:

advmod	- adverbial modification 
agr 	- agreement
apos	- apostrophe
aux	- auxiliary verb
contrac	- contraction
init	- initial constituent
ne	- named entity
nn	- noun-noun compound
npcoord - np coordination
npord 	- np order
ppord	- pp order
punct	- punctuation
sbjinv	- subject inversion
thatcmp	- 'that'-complementizer
vpord	- vp order
