
files
------

README: this file
src: source code of the HMM and fertility HMM model
run-word-align-example: details on running the word alignment models.
ini: some ini files used by the scripts in run-word-align-example.

run
----

step 1: prepare train test and afile text (examples available at fert/data)
============================================================================

1. train: tab seperated bitext sentences
2. test: same as train, usually small
3. afile: human labeled alignments


step 2: compile source code
=============================

> cd src
> make 

binaries are built in src/bin


step 3: run the models
=======================

Run on toy example included with release:

> cd run-word-align-example
> ./run.sh example

Details on changing run.sh and run-model.sh, read run-word-align-example/README

step 4: creating giza++ style alignments
========================================

fert/bin/format2giza <bitext-file> <afile> <reverse-or-not>
Note(Sam): it is possible that the word alignment decoder fails to create alignments for a few lines in the bitext file.
Currently, we can manually examine these two directions to make sure that MOSES is "happy".
This can be done easily because format2giza output the line numbers.
I will change the code slightly so as to take two afiles (two directions), and automatically generate files that MOSIS likes.

Sept 22, 2010
Sam Shaojun Zhao
