Allo-media data selection tool
This code implements the data selection method and algorithms proposed in Axelrod's paper CYNICAL SELECTION OF LANGUAGE MODEL TRAINING DATA, based on the paper's explanations and the Perl implementation Axelrod proposed on github
Comments in code and details on usage to come, but it's pretty simple right now.
Say you have a (small) representative corpus (task.txt) and a (big) general one (unadapted.txt) and you want to select sentences from the big corpus that look like the small corpus ones.
Usage would be:
./cynical-selection.py --task task.txt --unadapted unadapted.txt
This will produce a
.jaded file containing the selected sentences using the following tab-separated format:
sentence score (penalty + gain)
sentence id (in the selection)
sentence id (in the unadapted corpus)
See header of the script for available options, here is the two most important:
batch: essential with big corpora, allows to select more than one sentence at a time, see Axelrod's paper
iterate: iterate selection runs till no more than 10% of original size can be removed