Code for A Hierarchical Model for Data-to-Text Generation (Rebuffel, Soulier, Scoutheeten, Gallinari; ECIR 2020); most of this code is based on OpenNMT.
UPDATE 11/03/2021: The original checkpoints used to produce results from the
paper are officialy lost. However, I still have the actual model outputs, which
are now included in this repo. Simply unzip outputs.zip.
Furthermore, Radmil Raychev and Craig Thomson from the University of Aberdeen
are currently working with this repo, and have agreed to share their checkpoints,
namely htransformer.tar.gz2.
(Note that this file is not downloadable by command line, still looking for a better
alternative)
Once it's downloaded, simply tar -xvf htransformer.tar.gz2.
You'll find the data used to train the model, as well as *.cfg files and
*.pt checkpoints. Note that the data is from SportSett, which contains some
additional info (such as day of the week for instance).
(Also see Thomson et al. for more info regarding the impact of additional data
on system performances.)
You will need a recent python to use it as is, especially OpenNMT. However I guess files could be tweaked to work with older pythons. Please note that at the time of writting, torch==1.1.0 can be problematic with very recent version of python. I suggest running the code with python=3.6
Full requirements can be found in requirements.txt. Note that they are not really all required, it's the full pip freeze of a clean conda virtual env, you can probably make it work with less.
Beyond standard packages included in miniconda, usefull packages are torch==1.1.0 torchtext==0.4 and some others required to make onmt work (PyYAML and configargparse for example). I also use more_itertools to format the dataset.
The dataset used in the paper can be downloaded here. More specifically, you just need to download the RotoWire dataset:
cd data
wget https://github.com/harvardnlp/boxscore-data/raw/master/rotowire.tar.bz2
tar -xvjf rotowire.tar.bz2
cd ..You'll need to format the dataset so that it can be preprocessed by OpenNMT.
python data/make-dataset.py --folder data/
At this stage, your repository should look like this:
.
├── onmt/ # Most of the heavy-lifting is done by onmt
├── data/ # Dataset is here
│ ├── rotowire/ # Raw data stored here
├ ├── make-dataset.py # formating script
├ ├── train_input.txt
├ ├── train_output.txt
│ └── ...
└── ...
Before any code run, we build an experiment folder to keep things contained
python create-experiment.py --name exp-1
At this stage, your repository should look like this:
.
├── onmt # Most of the heavy-lifting is done by onmt
├── experiments # Experiments are stored here
│ └── exp-1
│ │ ├── data
├ │ ├── gens
│ │ └── models
├── data # Dataset is here
└── ...
Before training models via OpenNMT, you must preprocess the data. I've handled all useful parameters with a config file. Please check it out if you want to tweak things, I have tried to include comments on each command. For futher info you can always check out the OpenNMT preprocessing doc
python preprocess.py --config preprocess.cfg
At this stage, your repository should look like this:
├── onmt # Most of the heavy-lifting is done by onmt
├── experiments # Experiments are stored here
│ └── exp-1
│ │ ├── data
│ │ │ ├── data.train.0.pt
│ │ │ ├── data.valid.0.pt
│ │ │ ├── data.vocab.pt
│ │ │ ├── preprocess-log.txt
├ │ ├── gens
│ │ └── models
├── data # Dataset is here
└── ...
To train a hierarchical model on Rotowire you can run:
python train.py --config train.cfg
To train with different parameters than used in the paper, please refer to my comments in the config file, or check OpenNMT train doc.
This config file runs the training for 100 000 steps, however we manually stopped the training at 30 000.
Before translating, we average the last checkpoints. If you did anything different from previous commands, please change the first few line of average-checkpoints.py.
You can average checkpoints by running:
python average-checkpoints.py --folder exp-1 --steps 31000 32000 33000 34000 35000
Now you can simply translate the test input by running:
python translate.py --config translate.cfg
RG, CS and CO metrics were originaly (see here) coded in Lua and python2. Because of compatibility issues with modern hardware, I have re-implemented RG in PyTorch and python3:
Follow instructions at: KaijuML/rotowire-rg-metric.
You can evaluate the BLEU score using SacreBLEU from Post, 2018. See the repo for installation, it should be a breeze with pip.
You can get the BLEU score by running:
cat experiments/exp-1/gens/test/predictions.txt | sacrebleu --force data/test_output.txt
(Note that --force is not required as it doesn't change the score computation, it just suppresses a warning because this dataset is already tokenized)
Alternatively you can use any prefered method for BLEU computation. I have also checked scoring models with NLTK and scores were virtually the same.