Efficient Low-Memory Aligner

This is a word alignment tool based on efmaral, with the following main differences:

• More compact data structures are used, so memory requirements are much lower (by orders of magnitude).
• The estimation of alignment variable marginals is done one sentence at a time, which also saves a lot of memory at no detectable cost in accuracy.
• New: User-specified Dirichlet priors, which can be generated by the makepriors.py script to allow models to be saved. See below under Input data format and Generating priors.

Technical details relevant to both efmaral and eflomal can be found in the following article:

• Östling and Tiedemann (2016) (BibTeX).

To compile and install the C binary and the Python bindings:

make
sudo make install
python3 setup.py install

Change the INSTALLDIR parameter in the install step if you want to install somewhere other than the default /usr/local/bin (e.g. make install -e INSTALLDIR=~/bin). Note that the align.py script now uses the eflomal executable in the same directory as align.py, rather than in $PATH.

There are three main ways of using eflomal:

1. Directly call the eflomal binary. Note that this requires some preprocessing.
2. Use the align.py command-line interface, which is partly compatible with that of efmaral. Run python3 align.py --help for instructions.
3. Use the Cython module to call the eflomal binary, this takes care of the preprocessing and file conversions necessary. See the docstrings in eflomal.pyx for documentation.

In addition, there are convenience scripts for aligning and symmetrizing (with the atools program from fast_align) as well as evaluating with data from the WPT shared task datasets. These work the same way as in efmaral, please see its README for details.

When used with the -s and -t options for separate source/target files, the align.py interface expects one sentence per line with space-separated tokens, similar to most word alignment software.

The -i option assumes a fast_text style joint source/target file of the format

source sentence ||| target sentence
another source sentence ||| another target sentence
...

The --priors option expects a file generated by makepriors.py (see below). This file contains user-specified lexical, HMM and/or fertility distribution priors. Since the algorithm is asymmetric, HMM and fertility priors can be stored for both the forward and reverse directions. makepriors.py handles this automatically, see examples below.

Note that the default value of the Dirichlet priors (defined in eflomal.c as LEX_ALPHA, JUMP_ALPHA and FERT_ALPHA) will be added to whatever is specified in the priors file. This means that integer counts for whatever word forms you have data on are fine in the priors file.

It s possible to use the special <NULL> token in the priors file, in case you want to encourage certain word forms to remain unaligned. Currently the makepriors.py script does not generate these, and this feature has not been tested yet.

If you have a large file that you want to use as "training data", en-sv, and a small file that you later want to align quickly, en-sv.small, start by aligning the large file as usual, e.g.:

./align.py -i en-sv --model 3 -f en-sv.fwd -r en-sv.rev

Now you can generate priors based on this large aligned file, stored in en-sv.priors:

./makepriors.py -i en-sv -f en-sv.fwd -r en-sv.rev --priors en-sv.priors

Alternatively, you can symmetrize en-sv.fwd and en-sv.rev into en-sv.sym and pass the same file to both -f and -r:

atools -c grow-diag-final-and -i en-sv.fwd -j en-sv.rev >en-sv.sym
./makepriors.py -i en-sv -f en-sv.sym -r en-sv.sym --priors en-sv.priors

Now, if you have another file to align, en-sv.small, simply use e.g.:

./align.py -i en-sv.small --priors en-sv.priors --model 3 \
    -f en-sv.small.fwd -r en-sv.small.rev

This will be much faster than merging en-sv and en-sv.small and aligning them jointly, while nearly as accurate (assuming en-sv.small is much smaller than en-sv).

The alignment output contains the same number of lines as the input files, where each line contains pairs of indexes. For instance, if the source input contains the following:

a black cat

and the target input is the following:

kuro neko

the correct output would be:

1-0 2-1

That is, 1-0 indicates token 1 of the source (black) is aligned to token 0 of the target (kuro), and 2-1 that token 2 of the source (cat) is aligned to token 1 of the target (neko). NULL alignments are not present in the output.

Note that the forward and reverse alignments both use source-target order, so the output can be fed directly to atools (see scripts/align_symmetrize.sh for an example).

In case you made a mistake with the direction, you can fix it afterwards with scripts/reverse_moses.py.

This is a comparison between eflomal, efmaral and fast_align.

The difference between efmaral and eflomal is in part due to different default parameters, in particular the number of iterations and the number of independent samplers.

Note that all timing figures below include alignments in both directions (run in parallel) and symmetrization.