NMT Attention Alignment Visualizations

An attention alignment visualization tool for command line and web.

A part of the web version was borrowed from Nematus utils

The Machine Translation Marathon 2019 tutorial can be found here

Usage

• Train a neural MT system (using Marian, Sockeye or Neural Monkey)
• Translate text and get word or subword level alignments
• Visualize the alignments
  • in the command line standard output
  • in a web browser (PHP required)

Requirements

• Python 3.6 or newer

  • NLTK for BLEU calculation(requires Python versions 3.5, 3.6, 3.7, or 3.8)

  • Numpy

   pip install numpy nltk
  

• PHP 5.4 or newer (for web visualization)

How to get alignment files from NMT systems

• Nematus (This worked for the Theano version. Not sure about the current Tensorflow version...)

  • Run nematus/translate.py with the --output_alignment or -a parameter

• Neural Monkey

  • In the training.ini file add

   [alignment_saver]
   class=runners.word_alignment_runner.WordAlignmentRunner
   output_series="ali"
   encoder=<encoder>
   decoder=<decoder>
  

  and add alignment_saver to the runners in main

   runners=[<runner_greedy>, <alignment_saver>]
  

  • In the translation.ini file in eval_data add

   s_ali_out="out.alignment"
  

• Marian

  • Run marian-decoder with the parameter --alignment soft
  • If you use the transformer model architecture, you will need to train it with guided alignments. See the MT Marathon 2019 Tutorial for details about this.

• OpenNMT Run translate.lua to translate with the -save_attention parameter to save attentions to a file

• Sockeye

  • Run sockeye.translate to translate with the --output-type parameter set to translation_with_alignment_matrix to save attentions to a file.
  • If you use the transformer model architecture, you will need to implement (copy in the correct places) code from this pull requst. Otherwise, Sockeye will return attention values of 0 for all tokens.

• Other The easiest format to use (and convert to) is the one used by Nematus.

  For each sentence the first row should be <sentence id> ||| <target words> ||| <score> ||| <source words> ||| <number of source words> <number of target words>

  After that follow <number of target words> + 1 (for <EOS> token) rows with <number of source words> + 1 (for <EOS> token) columns with attention weights separated by spaces.

  After each sentence there should be an empty line.

  Note that the values of <sentence id>, <score>, <number of source words> and <number of target words> are actually ignored when creating visualizations, so they may as well all be 0.

  An example:

   0 ||| Obama welcomes Netanyahu ||| 0 ||| Obama empfängt Net@@ any@@ ah@@ u ||| 7 4
   0.723834 0.0471278 0.126415 0.000413103 0.000774298 0.000715227 0.10072 
   0.00572539 0.743366 0.0342341 0.000315413 0.00550132 0.00150629 0.209351 
   0.0122618 0.0073559 0.909192 0.000606444 0.00614908 0.00256837 0.0618667 
   0.00110054 0.0214063 0.0759918 0.000446028 0.104856 0.0435644 0.752634 
  

Publications

If you use this tool, please cite the following paper:

Matīss Rikters, Mark Fishel, Ondřej Bojar (2017). "Visualizing Neural Machine Translation Attention and Confidence." In The Prague Bulletin of Mathematical Linguistics volume 109 (2017).

@inproceedings{Rikters-EtAl2017PBML,
	author = {Rikters, Matīss and Fishel, Mark and Bojar, Ond\v{r}ej},
	journal={The Prague Bulletin of Mathematical Linguistics},
	volume={109},
	pages = {1--12},
	title = {{Visualizing Neural Machine Translation Attention and Confidence}},
	address={Lisbon, Portugal},
	year = {2017}
}

Examples

• in the command line as shaded blocks. Example with Neural Monkey alignments (separate source and target subword unit files are required)

  python process_alignments.py \
  -i test_data/neuralmonkey/alignments.npy  \
  -o color \
  -s test_data/neuralmonkey/src.en.bpe \
  -t test_data/neuralmonkey/out.lv.bpe \
  -f NeuralMonkey
  

• the same with Nematus alignments (source and target subword units are in the same file)

  python process_alignments.py \
  -i test_data/nematus/alignments.txt \
  -o color \
  -f Nematus
  

• in a text file as Unicode block elements

  python process_alignments.py \
  -i test_data/neuralmonkey/alignments.npy  \
  -o block \
  -s test_data/neuralmonkey/src.en.bpe \
  -t test_data/neuralmonkey/out.lv.bpe \
  -f NeuralMonkey
  

  or

  python process_alignments.py \
  -i test_data/neuralmonkey/alignments.npy  \
  -o block2 \
  -s test_data/neuralmonkey/src.en.bpe \
  -t test_data/neuralmonkey/out.lv.bpe \
  -f NeuralMonkey
  

• in the browser as links between words (demo here)

  python process_alignments.py \
  -i test_data/marian/marian.out.lv \
  -s test_data/marian/marian.src.en \
  -o web \
  -f Marian
  

Parameters for process_alignments.py

Configuration file

The parameters can be provided via configuration .ini file to have a smaller mess in the command line when calling the script.

For example, create a config.ini file:

[AlignmentsOne]
InputFile: ./test_data/neuralmonkey/alignments.npy
SourceFile: ./test_data/neuralmonkey/src.en.bpe
TargetFile: ./test_data/neuralmonkey/out.lv.bpe
From: NeuralMonkey

[Options]
OutputType: color

And run:

python process_alignments.py -c config.ini

Screenshots

Color, Block, Block2

Web

Compare