The code for moving from AMR to BMR has sensitive data from AMR and BabelNet, therefore to obtain the BMR 1.0 please send us an email with a proof of the AMR license.

Installation

conda install pytorch cudatoolkit=11.3 -c pytorch
pip install -r requirements.txt
pip install -e .

The code only works with transformers 4.11. The code works fine with torch 1.5. We recommend the usage of a new conda env.

Train

Modify config.yaml in configs. Instructions in comments within the file. Since the model is trained in a multilingual fashion, the AMR graph metadata has to include the field "# ::lng" with the language of the sentence (e.g., "en_XX", "es_XX", "de_DE", "it_IT", etc).

Text-to-AMR

python bin/train.py --config configs/config.yaml --direction graph --mode amr --model facebook/mbart-large-cc25

Results in runs/

AMR-to-Text

python bin/train.py --config configs/config.yaml --direction text --mode amr --model facebook/mbart-large-cc25

Results in runs/

Evaluate

Text-to-AMR

python bin/predict_amrs.py \
    --datasets <AMR-ROOT>/data/amrs/split/test/*.txt \
    --gold-path data/tmp/amr3.0/gold.amr.txt \
    --pred-path data/tmp/amr3.0/pred.amr.txt \
    --checkpoint runs/<checkpoint>.pt \
    --beam-size 5 \
    --batch-size 500 \
    --device cuda \
    --penman-linearization --use-pointer-tokens
    --mode amr
    --model facebook/mbart-large-cc25
    --language en_XX

gold.amr.txt and pred.amr.txt will contain, respectively, the concatenated gold and the predictions.

To reproduce our paper's results, you will also need need to run the BLINK entity linking system on the prediction file (data/tmp/amr3.0/pred.amr.txt in the previous code snippet). To do so, you will need to install BLINK, and download their models:

git clone https://github.com/facebookresearch/BLINK.git
cd BLINK
pip install -r requirements.txt
sh download_blink_models.sh
cd models
wget http://dl.fbaipublicfiles.com/BLINK//faiss_flat_index.pkl
cd ../..

Then, you will be able to launch the blinkify.py script:

python bin/blinkify.py \
    --datasets data/tmp/amr3.0/pred.amr.txt \
    --out data/tmp/amr3.0/pred.amr.blinkified.txt \
    --device cuda \
    --blink-models-dir BLINK/models

To have comparable Smatch scores you will also need to use the scripts available at https://github.com/mdtux89/amr-evaluation, which provide results that are around ~0.3 Smatch points lower than those returned by bin/predict_amrs.py.

AMR-to-Text

python bin/predict_sentences.py \
    --datasets <AMR-ROOT>/data/amrs/split/test/*.txt \
    --gold-path data/tmp/amr3.0/gold.text.txt \
    --pred-path data/tmp/amr3.0/pred.text.txt \
    --checkpoint runs/<checkpoint>.pt \
    --beam-size 5 \
    --batch-size 500 \
    --device cuda \
    --penman-linearization --use-pointer-tokens
    --mode amr
    --model facebook/mbart-large-cc25
    --language en_XX

gold.text.txt and pred.text.txt will contain, respectively, the concatenated gold and the predictions. For BLEU, chrF++, and Meteor in order to be comparable you will need to tokenize both gold and predictions using JAMR tokenizer. To compute BLEU and chrF++, please use bin/eval_bleu.py. For METEOR, use https://www.cs.cmu.edu/~alavie/METEOR/ . For BLEURT don't use tokenization and run the eval with https://github.com/google-research/bleurt. Also see the appendix.

Linearizations

The previously shown commands assume the use of the DFS-based linearization. To use BFS or PENMAN decomment the relevant lines in configs/config.yaml (for training). As for the evaluation scripts, substitute the --penman-linearization --use-pointer-tokens line with --use-pointer-tokens for BFS or with --penman-linearization for PENMAN.