Edit model card

AMRBART (large-sized model)

AMRBART model is continually pre-trained on the English text and AMR Graphs based on the BART model. It was introduced in the paper: Graph Pre-training for AMR Parsing and Generation by bai et al. in ACL 2022 and first released in this repository.

Model description

AMRBART follows the BART model which uses a transformer encoder-encoder architecture. AMRBART is pre-trained with 6 tasks:

  • learning to reconstruct the text based on the corrupted text.
  • learning to reconstruct AMR graphs based on the corrupted AMR graph.
  • learning to reconstruct the text based on the corrupted text and its corresponding AMR graph.
  • learning to reconstruct an AMR graph based on the corrupted AMR graph and its corresponding text.
  • learning to reconstruct the text based on the corrupted text and its corresponding corrupted AMR graph.
  • learning to reconstruct an AMR graph based on the corrupted AMR graph and its corresponding corrupted text.

AMRBART is particularly effective when fine-tuned for AMR parsing and AMR-to-text generation tasks.

Training data

The AMRBART model is pre-trained on AMR3.0, a dataset consisting of 55,635 training instances and English Gigaword (we randomly sampled 200,000 sentences).

Intended uses & limitations

You can use the raw model for either AMR encoding or AMR parsing, but it's mostly intended to be fine-tuned on a downstream task.

How to use

Here is how to initialize this model in PyTorch:

from transformers import BartForConditionalGeneration
model = BartForConditionalGeneration.from_pretrained("xfbai/AMRBART-large")

Please refer to this repository for tokenizer initialization and data preprocessing.

BibTeX entry and citation info

Please cite this paper if you find this model helpful

@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"
}
Downloads last month
7
Hosted inference API
This model can be loaded on the Inference API on-demand.