1 Instructions for running the compound splitter, which is a reimplementation
2 and extension (more features, larger non-word list) of the model described in
4 C. Dyer. (2009) Using a maximum entropy model to build segmentation
5 lattices for MT. In Proceedings of NAACL HLT 2009,
6 Boulder, Colorado, June 2009
8 If you use this software, please cite this paper.
11 GENERATING 1-BEST SEGMENTATIONS AND LATTICES
12 ------------------------------------------------------------------------------
14 Here are some sample invokations:
16 ./compound-split.pl --output 1best < infile.txt > out.1best.txt
17 Segment infile.txt according to the 1-best segmentation file.
19 ./compound-split.pl --output plf < infile.txt > out.plf
21 ./compound-split.pl --output plf --beam 3.5 < infile.txt > out.plf
22 This generates denser lattices than usual (the default beam threshold
23 is 2.2, higher numbers do less pruning)
26 MODEL TRAINING (only for the adventuresome)
27 ------------------------------------------------------------------------------
29 I've included some training data for training a German language lattice
30 segmentation model, and if you want to explore, you can or change the data.
31 If you're especially adventuresome, you can add features to cdec (the current
32 feature functions are found in ff_csplit.cc). The training/references are
37 The format is the unsegmented form on the right and the reference lattice on
38 the left, separated by a triple pipe ( ||| ). Note that the segmentation
39 model inserts a # as the first word, so your segmentation references must
42 To retrain the model (using MAP estimation of a conditional model), do the
48 Note, the optimization objective is supposed to be non-convex, but i haven't
49 found much of an effect of where I initialize things. But I haven't looked
50 very hard- this might be something to explore.