AMRBART

An implementation for ACL2022 paper “Graph Pre-training for AMR Parsing and Generation”. You may find our paper here (Arxiv).

PWC

PWC

PWC

PWC

Requirements

  • python 3.8
  • pytorch 1.8
  • transformers 4.8.2
  • pytorch-lightning 1.5.0
  • Tesla V100 or A100

We recommend to use conda to manage virtual environments:

conda env update --name <env> --file requirements.yml

We also provide a docker image here.

Data Processing

You may download the AMR corpora at LDC.

We follow Spring to preprocess AMR graphs:

# 1. install spring 
cd spring && pip install -e 
# 2. processing data
bash run-preprocess.sh

Pre-training

bash run-posttrain-bart-textinf-joint-denoising-6task-large-unified-V100.sh /path/to/BART/

Fine-tuning

AMR Parsing

bash finetune_AMRbart_amrparsing.sh /path/to/pre-trained/AMRBART/ gpu_id

AMR-to-text Generation

bash finetune_AMRbart_amr2text.sh /path/to/pre-trained/AMRBART/ gpu_id

Evaluation

AMR Parsing

bash eval_AMRbart_amrparsing.sh /path/to/fine-tuned/AMRBART/ gpu_id

AMR-to-text Generation

bash eval_AMRbart_amr2text.sh /path/to/fine-tuned/AMRBART/ gpu_id

Pre-trained Models

Pre-trained AMRBART

Setting Params checkpoint
AMRBART-base 142M model
AMRBART-large 409M model

Fine-tuned models on AMR-to-Text Generation

Setting BLEU(tok) BLEU(detok) checkpoint output
AMRBART-large (AMR2.0) 49.8 45.7 model output
AMRBART-large (AMR3.0) 49.2 45.0 model output

Fine-tuned models on AMR Parsing

Setting Smatch checkpoint output
AMRBART-large (AMR2.0) 85.4 model output
AMRBART-large (AMR3.0) 84.2 model output

Todo

  • clean code

References

@inproceedings{bai-etal-2022-graph,
    title = "Graph Pre-training for {AMR} Parsing and Generation",
    author = "Bai, Xuefeng  and
      Chen, Yulong and
      Zhang, Yue",
    booktitle = "Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = may,
    year = "2022",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "todo",
    doi = "todo",
    pages = "todo"
}

GitHub

View Github