AMR Parsing as Sequence-to-Graph Transduction

Code for the AMR Parser in our ACL 2019 paper "AMR Parsing as Sequence-to-Graph Transduction".

If you find our code is useful, please cite:

    title = "{AMR Parsing as Sequence-to-Graph Transduction}",
    author = "Zhang, Sheng and
      Ma, Xutai and
      Duh, Kevin and
      Van Durme, Benjamin",
    booktitle = "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = jul,
    year = "2019",
    address = "Florence, Italy",
    publisher = "Association for Computational Linguistics"

1. Environment Setup

The code has been tested on Python 3.6 and PyTorch 0.4.1. All other dependencies are listed in requirements.txt.

Via conda:

conda create -n stog python=3.6
source activate stog
pip install -r requirements.txt

2. Data Preparation

Download Artifacts:


Assuming that you're working on AMR 2.0 (LDC2017T10), unzip the corpus to data/AMR/LDC2017T10, and make sure it has the following structure:

(stog)$ tree data/AMR/LDC2017T10 -L 2
├── data
│   ├── alignments
│   ├── amrs
│   └── frames
├── docs
│   ├── AMR-alignment-format.txt
│   ├── amr-guidelines-v1.2.pdf
│   ├── file.tbl
│   ├── frameset.dtd
│   ├── PropBank-unification-notes.txt
│   └── README.txt
└── index.html

Prepare training/dev/test data:

./scripts/ -v 2 -p data/AMR/LDC2017T10

3. Feature Annotation

We use Stanford CoreNLP (version 3.9.2) for lemmatizing, POS tagging, etc.

First, start a CoreNLP server following the API documentation.

Then, annotate AMR sentences:

./scripts/ data/AMR/amr_2.0

4. Data Preprocessing


5. Training

Make sure that you have at least two GeForce GTX TITAN X GPUs to train the full model.

python -u -m stog.commands.train params/stog_amr_2.0.yaml

6. Prediction

python -u -m stog.commands.predict \
    --archive-file ckpt-amr-2.0 \
    --weights-file ckpt-amr-2.0/ \
    --input-file data/AMR/amr_2.0/test.txt.features.preproc \
    --batch-size 32 \
    --use-dataset-reader \
    --cuda-device 0 \
    --output-file test.pred.txt \
    --silent \
    --beam-size 5 \
    --predictor STOG

7. Data Postprocessing

./scripts/ test.pred.txt

8. Evaluation

Note that the evaluation tool works on python2, so please make sure python2 is visible in your $PATH.

./scripts/ test.pred.txt data/AMR/amr_2.0/test.txt

Pre-trained Models

Here are pre-trained models: ckpt-amr-2.0.tar.gz and ckpt-amr-1.0.tar.gz. To use them for prediction, simply download & unzip them, and then run Step 6-8.

In case that you only need the pre-trained model prediction (i.e., test.pred.txt), you can find it in the download.


We adopted some modules or code snippets from AllenNLP, OpenNMT-py and NeuroNLP2. Thanks to these open-source projects!