SimAlign: Similarity Based Word Aligner

SimAlign is a high-quality word alignment tool that uses static and contextualized embeddings and does not require parallel training data.

The following table shows how it compares to popular statistical alignment models:

Shown is F1, maximum across subword and word level. For more details see the Paper.

Installation and Usage

Tested with Python 3.7, Transformers 3.1.0, Torch 1.5.0. Networkx 2.4 is optional (only required for Match algorithm). For full list of dependencies see setup.py. For installation of transformers see their repo.

Download the repo for use or alternatively install with PyPi

pip install simalign

or directly with pip from GitHub

pip install --upgrade git+https://github.com/cisnlp/simalign.git#egg=simalign

An example for using our code:

from simalign import SentenceAligner

# making an instance of our model.
# You can specify the embedding model and all alignment settings in the constructor.
myaligner = SentenceAligner(model="bert", token_type="bpe", matching_methods="mai")

# The source and target sentences should be tokenized to words.
src_sentence = ["This", "is", "a", "test", "."]
trg_sentence = ["Das", "ist", "ein", "Test", "."]

# The output is a dictionary with different matching methods.
# Each method has a list of pairs indicating the indexes of aligned words (The alignments are zero-indexed).
alignments = myaligner.get_word_aligns(src_sentence, trg_sentence)

for matching_method in alignments:
    print(matching_method, ":", alignments[matching_method])

# Expected output:
# mwmf (Match): [(0, 0), (1, 1), (2, 2), (3, 3), (4, 4)]
# inter (ArgMax): [(0, 0), (1, 1), (2, 2), (3, 3), (4, 4)]
# itermax (IterMax): [(0, 0), (1, 1), (2, 2), (3, 3), (4, 4)]

For more examples of how to use our code see scripts/align_example.py.

Demo

An online demo is available here.

Gold Standards

Links to the gold standars used in the paper are here:

Evaluation Script

For evaluating the output alignments use scripts/calc_align_score.py.

The gold alignment file should have the same format as SimAlign outputs. Sure alignment edges in the gold standard have a '-' between the source and the target indices and the possible edges have a 'p' between indices. For sample parallel sentences and their gold alignments from ENG-DEU, see samples.

Publication

If you use the code, please cite

@inproceedings{jalili-sabet-etal-2020-simalign,
    title = "{S}im{A}lign: High Quality Word Alignments without Parallel Training Data using Static and Contextualized Embeddings",
    author = {Jalili Sabet, Masoud  and
      Dufter, Philipp  and
      Yvon, Fran{\c{c}}ois  and
      Sch{\"u}tze, Hinrich},
    booktitle = "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: Findings",
    month = nov,
    year = "2020",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2020.findings-emnlp.147",
    pages = "1627--1643",
}

Feedback

Feedback and Contributions more than welcome! Just reach out to @masoudjs or @pdufter.

FAQ

Do I need parallel data to train the system?

No, no parallel training data is required.

Which languages can be aligned?

This depends on the underlying pretrained multilingual language model used. For example, if mBERT is used, it covers 104 languages as listed here.

Do I need GPUs for running this?

Each alignment simply requires a single forward pass in the pretrained language model. While this is certainly faster on GPU, it runs fine on CPU. On one GPU (GeForce GTX 1080 Ti) it takes around 15-20 seconds to align 500 parallel sentences.

License

Copyright (C) 2020, Masoud Jalili Sabet, Philipp Dufter

A full copy of the license can be found in LICENSE.