summaryrefslogtreecommitdiff
path: root/compound-split/README
diff options
context:
space:
mode:
authorChris Dyer <cdyer@allegro.clab.cs.cmu.edu>2012-11-14 20:33:51 -0500
committerChris Dyer <cdyer@allegro.clab.cs.cmu.edu>2012-11-14 20:33:51 -0500
commitf8d9ff4aaeb1d1f773bacfe9ee75d1d1778ec26b (patch)
treecfd9cd1e19e3fa33888626c204a4e0b73ca2edc4 /compound-split/README
parentdf5b25f73c12ef03482bd902ee0155a56789e6b9 (diff)
major mert clean up, stuff for simple system demo
Diffstat (limited to 'compound-split/README')
-rw-r--r--compound-split/README51
1 files changed, 0 insertions, 51 deletions
diff --git a/compound-split/README b/compound-split/README
deleted file mode 100644
index b7491007..00000000
--- a/compound-split/README
+++ /dev/null
@@ -1,51 +0,0 @@
-Instructions for running the compound splitter, which is a reimplementation
-and extension (more features, larger non-word list) of the model described in
-
- C. Dyer. (2009) Using a maximum entropy model to build segmentation
- lattices for MT. In Proceedings of NAACL HLT 2009,
- Boulder, Colorado, June 2009
-
-If you use this software, please cite this paper.
-
-
-GENERATING 1-BEST SEGMENTATIONS AND LATTICES
-------------------------------------------------------------------------------
-
-Here are some sample invokations:
-
- ./compound-split.pl --output 1best < infile.txt > out.1best.txt
- Segment infile.txt according to the 1-best segmentation file.
-
- ./compound-split.pl --output plf < infile.txt > out.plf
-
- ./compound-split.pl --output plf --beam 3.5 < infile.txt > out.plf
- This generates denser lattices than usual (the default beam threshold
- is 2.2, higher numbers do less pruning)
-
-
-MODEL TRAINING (only for the adventuresome)
-------------------------------------------------------------------------------
-
-I've included some training data for training a German language lattice
-segmentation model, and if you want to explore, you can or change the data.
-If you're especially adventuresome, you can add features to cdec (the current
-feature functions are found in ff_csplit.cc). The training/references are
-in the file:
-
- dev.in-ref
-
-The format is the unsegmented form on the right and the reference lattice on
-the left, separated by a triple pipe ( ||| ). Note that the segmentation
-model inserts a # as the first word, so your segmentation references must
-include this.
-
-To retrain the model (using MAP estimation of a conditional model), do the
-following:
-
- cd de
- ./TRAIN
-
-Note, the optimization objective is supposed to be non-convex, but i haven't
-found much of an effect of where I initialize things. But I haven't looked
-very hard- this might be something to explore.
-