AMR2text from Pre-trained Transformer

Code to reproduce

GPT-too: A Language-Model-First Approach for AMR-to-Text-Generation

training GPT-2 models to attain state-of-the art AMR-to-text generation.

Install

Code was tested for Python 3.6 on Linux machines both for x86 and PowerPC architectures.

All scripts source an environment script that can be used to activate virtual environments or setup system variables specific to the project. If you do not need any of this just set it to an empty file

touch set_environment.sh

examples of set_environment.sh for a pip install with virtual environment

[ ! -d venv ] && virtualenv venv
. venv/bin/activate

and with a conda environment

eval "$(/path/to/my/ppc64/miniconda3/bin/conda shell.bash hook)"
[ ! -d cenv_ppc ] && conda create -y -p ./cenv_ppc
conda activate ./cenv_ppc

the installers also source set_environment.sh and download all the necessary modules and tools, available installers

bash scripts/amr2txt/install_x86_with_pip.sh
bash scripts/amr2txt/install_x86_with_conda.sh
bash scripts/amr2txt/install_ppc_with_conda.sh

Full pip install for x86, conda install for x86 and conda install for Power PC respectively. You can also run the commands on those scripts one by one for a manual install.

Run full train and test on LDC2016T10

You will need to store the LDC2016T10 dataset with following names

DATA/train.amr
DATA/dev.amr
DATA/test.amr

The following script trains a GPT-2 medium version of the AMR2txt system. It also selects best model in dev according to BLEU and runs beam decoding on it.

bash scripts/amr2txt/experiment.sh scripts/amr2txt/configs/acl2020.sh

a similar config is also available for GPT-2 large.

Citation

If you use the source code or want to refere to our findings, please consider citing the follow paper:

@inproceedings{mager-etal-2020-gpt-too,
    title = "GPT-too: A language-model-first approach for AMR-to-text generation",
    author = "Manuel, Mager and Ram\'on, Fernandez Astudillo and Tahira, Naseem and Md Arafat, Sultan and Young-Suk, Lee and Radu, Florian and Salim, Roukos",
    booktitle = "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
    month = jul,
    year = "2020",
    address = "Seattle, USA",
    publisher = "Association for Computational Linguistics",
}

keshavramji/GPT-too-AMR2text

AMR2text from Pre-trained Transformer

Install

Run full train and test on LDC2016T10

Citation