(carmel includes EM and gibbs-sampled (pseudo-Bayesian) training)

(see carmel/LICENSE - free for research/non-commercial)

(see carmel/README and carmel/carmel-tutorial).

mkdir build
cd build
cmake ..
make
make install

optional cmake parameters:

• -DCMAKE_INSTALL_PREFIX=/custom/install/path
• -DBOOST_ROOT=/path/to/boost (if it is not installed in standard location)
• -DOPENFST_ROOT=/path/to/openfst (if desired, and it is not installed in standard location)

prerequisites:

• cmake 3.1 or higher
• a C++11 compiler
• Boost. Tested on versions between 1.53 and 1.59, but others should work, too
• Optionally, openfst (http://www.openfst.org)

• carmel: finite state transducer toolkit with EM and gibbs-sampled (pseudo-Bayesian) training

• forest-em: derivation forests EM and gibbs (dirichlet prior bayesian) training

• graehl/shared: utility C++/Make libraries used by carmel and forest-em

• gextract: some python bayesian syntax MT rule inference

• sblm: some simple pcfg (e.g. penn treebank parses, but preferably binarized)

• clm: some class-based LM feature? I forget.

• cipher: some word-class discovery and unsupervised decoding of simple probabilistic substitution cipher (uses carmel, but look to the tutorial in carmel/ first)

• util: misc shell/perl scripts