fast_align_ng is a fast, unsupervised word aligner which can use POS tags to improve performance. It is based on fast_align, and incorporates some additional features described in:

• Simple extensions for a reparameterised IBM Model 2. Douwe Gelling and Trevor Cohn. In Proceedings of ACL Short papers, 2014.

Please cite this paper if you use this software.

fast_align_ng depends on the TCLAP library and a c++11 compiler. You can run make in the base folder to compile, which should give you an executable.

The following command will give you forward alignments, add the -r flag for backwards alignments:

./fast_align_ng -i data.zh-en -P data.pos.zh-en > forward.align

The input file (-i) should have 2 sentences per line, separated by |||, which are translations of each other. Sentences should be tokenized and tokens separated by whitespace. The POS file (-P) is optional, but should have the same format as the input file, and each line should have as many tokens (POS tags) as the corresponding input line.

run fast_align_ng with no arguments for additional options

The source code in this repository is provided under the terms of the Apache License, Version 2.0.