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:

@inproceedings{zhang-etal-2018-stog,
    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 1.9. And requires python 2.7 for evaluation (aliased as python2) All other dependencies are listed in requirements.txt.

Via conda:

conda create -n stog python=3.6
conda activate stog
# Install pytorch first
conda install pytorch cudatoolkit=11.1 -c pytorch -c conda-forge
pip install -r requirements.txt

Do not install Apex for NVIDIA as it will introduce problems casued by fused_layer_norm_cuda (which is not supported for pytorch < 1.0.0>)

2. Data Preparation

Before downloading artifact, consider choosing which language model to use and data version to use to reduce the download time, after that run the script. If you want to experiments, it is advisable to uncomment all of the lines in scripts/download_artifacts.sh to download all artifacts and models. But if you have models in mind, you can uncomment specific part of the script.

./scripts/download_artifacts.sh

2a. English AMR

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/AMR/LDC2017T10
├── 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/prepare_data.sh -v 2 -p data/AMR/LDC2017T10

2b. Indonesian AMR

If you're working with Indonesian AMR instead, either use the pre-annotated AMR dataset from our repo. This file is annotated following Ilmy's method (using Anago for NER, CRFTagger for POS). If you're using preannotated dataset, skup the feature annotation part and go straight to preprocessing.

Or start from scratch using this AMR ID 2.0 dataset. If you're starting from scratch, refer to amr_parser repo on how to annotate the features

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.

or you can also use docker to start a corenlp server using this script:

docker run -p 9000:9000 banditelol/corenlp

Then, annotate AMR sentences:

./scripts/annotate_features.sh data/AMR/amr_2.0

4. Data Preprocessing

Make sure the variables inside this script matches with what you have in your machine. Seriously, look inside the script to understand it before running the script. At least check if the variables match your environment.

# English AMR
./scripts/preprocess_2.0.sh

# Indonesian AMR
./scripts/preprocess_2.0_id.sh

# Indonesian AMR (Windows)
./scripts/preprocess_2.0_id.bat

5. Training

Make sure that you have at least two GeForce GTX TITAN X GPUs to train the full model using AMR 2.0. But if you're training for AMR ID, GT 950M with 4GB of memory can do decent works. To train with specific configuration, you can look inside params/ folder to see the kind of configurations possible.

# python -u -m stog.commands.train params/[PARAMS_FILE]
python -u -m stog.commands.train params/stog_amr_2.0.yaml
# or training the best configuration for AMR ID
python -u -m stog.commands.train params/stog_amr_id_2.0_best.yaml

If you get error parsing yaml, try installing pyyaml==5.4.1

6. Prediction

# python -u -m stog.commands.predict \
#     --archive-file [CHECKPOINT_FOLDER_] \
#     --weights-file [CHECKPOINT_FOLDER_]/best.th \
#     --input-file [PREPROCESSED_FILE] \
#     --batch-size 16 \
#     --use-dataset-reader \
#     --cuda-device 0 \
#     --output-file [CHECKPOINT_FOLDER]/test_id_trans.pred.txt \
#     --silent \
#     --beam-size 5 \
#     --predictor STOG
# For example:

python -u -m stog.commands.predict \
    --archive-file ckpt-amr-id-2.0-large-best-rerun/model.tar.gz \
    --weights-file ckpt-amr-id-2.0-large-best-rerun/best.th \
    --input-file data/AMR/amr_id_2.0/test.txt.features.preproc \
    --batch-size 16 \
    --use-dataset-reader \
    --cuda-device 0 \
    --output-file ckpt-amr-id-2.0-large-best-rerun/test.pred.txt \
    --silent \
    --beam-size 5 \
    --predictor STOG

7. Data Postprocessing

./scripts/postprocess_2.0.sh test.pred.txt

For AMR ID use this instead:

./scripts/postprocess_2.0_id.sh {test_dir} {test_name}

8. Evaluation

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

# ./scripts/compute_smatch.sh {checkpoint_dir} {pred_filename} {amr_data_dir}
./scripts/compute_smatch.sh ckpt-amr-id-2.0-large-best-rerun test.pred.txt data/AMR/amr_id_2.0

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.

Indonesian Language

For models trained on Indonesian Language, you could download the pretrained model here. With the requirement of having the following models:

IndoBert
IndoGlove 4B

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

Acknowledgements

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

License

MIT

banditelol / stog

readme