Data and code for replicating WMT17 Multimodal Translation results
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
config
data.tok.bpe
data
models
scripts
.gitignore
README.md
mmt.png

README.md

LIUM-CVC WMT17 Multimodal Translation Systems

mmt

This repository contains necessary scripts and data files to replicate the results of LIUM-CVC submissions to Multimodal Translation (MMT) task of WMT17 for both en->de and en->fr.

Note : You need to install nmtpy in order to follow this tutorial.

PLEASE READ THIS

This codebase uses the old Theano version of nmtpy which is no longer maintained. The new codebase is based on torch and the config files here are not compatible with that. So this repository is not really useful these days.

Data

The data/ folder contains all English, German and French corpora files from Multi30k dataset necessary to train and evaluate the systems. A secondary ambiguous MSCOCO test set is also provided along with the main Flickr30k sets. All the files under data/ are verbatim copies of files downloaded from the official campaign website.

  • train.* : 29K sentences
  • val.* : 1014 sentences
  • test2016.* : 1000 sentences
  • test2017.* : 1000 sentences
  • testcoco.* : 461 sentences

The *_images.txt files are text files containing the list of image names for each split as they are ordered in the image features files.

Preparation

The script scripts/preprocess-bpe-pkl.sh will first use the following scripts (which should be available in your $PATH) from Moses repository in order to preprocess the corpora:

  • Normalize punctuations (normalize-punctuation.perl)
  • Tokenize (tokenizer.perl)
  • Lowercase (lowercase.perl)

It will then learn a joint BPE model with 10K merge operations (for en->de and en->fr separately) using the tools provided by the subword-nmt repository. You need to adapt the script to point to the correct subword-nmt folder by modifying the variables BPE_APPLY and BPE_LEARN.

Once the BPE-ized files are saved under data.tok.bpe/bpe.en-de and data.tok.bpe/bpe.en-fr, nmt-build-dict from nmtpy project will be used to create the vocabulary .pkl files in the respective folders.

Since the multimodal architectures have their own data iterators, they need a special .pkl corpora file for each Flickr30k and MSCOCO split. These files are created by scripts/raw2pkl which is automatically called from scripts/preprocess-bpe-pkl.sh.

In the end, the files in data.tok.bpe/bpe.en-{de,fr} will be the files that are used by nmtpy. The non-BPE versions of validation and test sets from data.tok.bpe will also be used when scoring the hypotheses with automatic metrics.

Note: The script should be launched directly from the wmt17-mmt checkout folder.

Image Features

Once the above preprocessing step is completed, you will need to download and extract the image features under data/images as described in the relevant README file.

Training

You should now be ready to train monomodal and multimodal architectures using the prepared data. If everything went well and you have a recent enough installation of nmtpy, you can use the following commands to start training your baselines:

# Monomodal En->De system
$ nmt-train -c config/monomodal-en-de.conf

# Monomodal En->Fr system
$ nmt-train -c config/monomodal-en-fr.conf

# MNMT (trgmul variant) En->De system
$ nmt-train -c config/mnmt-en-de.conf

# MNMT (trgmul variant) En->Fr system
$ nmt-train -c config/mnmt-en-fr.conf

# MNMT (fusion with conv features) En->De system
$ nmt-train -c config/fusion-en-de.conf

# MNMT (fusion with conv features) En->Fr system
$ nmt-train -c config/fusion-en-fr.conf

These configurations will save the best .npz checkpoints under models/ inside your wmt17-mmt checkout.

Decoding & Scoring

# Decode test2017 for monomodal en->de
$ nmt-translate -m models/monomodal-en-de/attention-e128-r256-...-s1234.1.BEST.npz \
                -S data.tok.bpe/bpe.en-de/test2017.norm.tok.lc.bpe10000.en \
                -o nmt.test2017.tok.de

# Decode test2017 for mnmt en->de
$ nmt-translate -m models/mnmt-en-de/mnmt_trgmul-e128-i2048-r256-...-s1234.1.BEST.npz \
                -S data.tok.bpe/bpe.en-de/test2017.bpe10000.pkl \
                   data/images/resnet50-imagenet-pool5/flickr30k_ResNet50_pool5_test2017.npy \
                -o mnmt.test2017.tok.de

# Score both systems (Output stripped to fit here)
$ nmt-coco-metrics -l de -r data.tok.bpe/test2017.norm.tok.lc.de -s *test2017.tok.de

|    Bleu_1     ||    Bleu_2     ||    Bleu_3     ||    Bleu_4     ||    METEOR   |
nmt.test2017.tok.de
|    63.321     ||    48.730     ||    38.804     ||    31.255     ||    51.274   |
mnmt.test2017.tok.de
|    64.342     ||    49.977     ||    39.879     ||    32.112     ||    51.529   |

Results

(Note that the results below belong to single runs while the ones reported in the paper are averages and ensembles of 5 runs.)

System Val METEOR/BLEU Test2016 METEOR/BLEU Test2017 METEOR/BLEU
monomodal-en-de 56.83/39.17 57.40/39.00 51.27/31.25
mnmt-en-de 56.99/39.15 57.05/38.97 51.52/32.11
fusion-en-de
monomodal-en-fr 72.87/57.79 74.19/59.02 68.87/51.93
mnmt-en-fr 73.88/58.93 74.75/59.82 69.48/52.61
fusion-en-fr