Hieralign is yet another symmetric word alignment tool simple, fast and unsupervised, it has been implemented with multi-threading in all stages, which makes it run as fast as fast_align.

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

Hao Wang and Yves Lepage. Hierarchical Sub-sentential Alignment with IBM Models for Statistical Phrase-based Machine Translation. Jounral of Natural Language Processing. Vol.24 No.4 2017.9.

1. Pure C++ implementation, included the new features of C++11, no additional library required.

2. multi-threading, It has been implemented with multi-threading in all stages (same as fast_align, using OpenMP) and well-optimized.

3. cross-platform CMake file, easy and one-step compiling and installation.

4. replaced the best-1 version of hierarchical subsentential alignment with beam search version.

5. fast, less than 2 minutes for 300,000 lines.

6. well-formed codes and simple class structure.

The input to Hieralign must be tokenized and aligned into parallel sentences. Each line is a source language sentence and its target language translation, separated by a triple pipe symbol with leading and trailing white space (\t). An example 2-sentence English-Japanese parallel corpus is:

  I have a book . 私 は 本 を 持って いた 。
  I do not know the answer    私 は その 答え を 知ら ない

Building Hieralign requires a modern C++ compiler and the CMake build system. Additionally, the OpenMP library is required here to obtain better performance. Please make sure your complier version (gcc/clang) supports OpenMP before compiling.

To compile, do the following

cmake . 
make

Run Hieralign -h to see a list of command line options. There are 3 ways to use Hieralign:

1. as a normal word aligner: Hieralign generates symmetric alignments (source–target, i.e., left language–right language) alignments is:

   ./Hieralign --train train.txt --model vbibm1 > aligned.hier

2. train the model in advance: The usually and recommended way to train and store a model is to just add the --model_file option:

   ./Hieralign --train train.txt --model vbibm1 --model_file path_to_store_model

3. as an online aligner: Given the trained model, Hieralign can be ran as an online aligner:

   ./Hieralign --input test.txt --model vbibm1 --model_file path_to_store_model > aligned.hier

Hieralign produces outputs in the widely-used i-j “Pharaoh format,” where a pair i-j indicates that the ith word (zero-indexed) of the left language (source) is aligned to the jth word of the right sentence (target) but have been symmetrized. For example, a good alignment for first sentence would be:

  0-0 0-1 1-4 1-5 3-2 4-6
  0-0 1-6 2-6 3-5 4-2 5-3