AMR-DA: Data Augmentation by Abstract Meaning Representation

This repository contains the code for our ACL-2022 paper: AMR-DA: Data Augmentation by Abstract Meaning Representation.

This figure shows an overview of AMR-DA: AMR parser first transduces the sentence into an AMR graph, followed by an AMR graph extender to diversify graphs with different augmentation strategies; finally, the AMR generator synthesizes augmentations from AMR graphs.

The work adopts SPRING as AMR parser and plms-graph2text as AMR generator.

Generated Data Examples

Augmentations examples for wiki: original data and generated augmentations.

Requirements

pip install -r requirements.txt

Text to AMR

Parse the plain text to amr graph

cd amr-parser-spring
bash predict_amr.sh <plain_text_file_path>(../data/wiki_data/wiki.txt)

Preprocess amr graph, convert to source and target string

cd data-utils/preprocess
bash prepare_data.sh <amr_file_path>(../../data/wiki_data/wiki.amr)

Graph Modification

cd data_utils
python augment.py (modify parameters according to specific requirements)

AMR to Text

Generate text from amr graph

cd plms-graph2text
bash decode_AMR.sh <model-path> <checkpoint> <gpu_id> <source file> <output-name>
(bash decode_AMR.sh /path/to/t5-base amr-t5-base.ckpt 0 ../data/wiki-data/wiki.source wiki-perd-t5-base.txt)

Experiments in this paper

For STS tasks, we directly used the code from SimCSE. For text classification tasks, please refer the code of EDA.

Citation

Please cite this repository using the following reference:

@inproceedings{shou-etal-2022-amr,
    title = "{AMR-DA}: {D}ata Augmentation by {A}bstract {M}eaning {R}epresentation",
    author = "Shou, Ziyi  and
      Jiang, Yuxin  and
      Lin, Fangzhen",
    booktitle = "Findings of the Association for Computational Linguistics: ACL 2022",
    month = may,
    year = "2022",
    address = "Dublin, Ireland",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.findings-acl.244",
    pages = "3082--3098"
}