This is the official implementation of AMR Parsing with Causal Hierarchical Attention and Pointers, presented at EMNLP 2023.
chmod +x install.sh
. ./install.sh
The installation script will create a new conda environment, named as chap_amr, and compile some c++ extension, developed based on Transformer Grammars.
First, prepare AMR graph banks using the script in amrlib: https://github.com/bjascob/amrlib/blob/master/scripts/33_Model_Parse_XFM/10_Collect_AMR_Data.py. You will obtain a directory tdata_xfm in amrlib's working directory.
Move it to <project root>/data/AMR3.0/tdata_xfm (change to AMR2.0 if processing the AMR 2.0 graph bank; and you don't need this step when processing OOD data).
Then, run scripts/10_prepare_data.sh at the project root. It contains commands of processing all datasets. So you may just need some of them according to your usage.
python src/train.py experiment=<model config>
# our best model
python src/train.py experiment=t2g_point2tgt_paware_strict_adapter
We use hydra to configure training and models. This project is based on the lightning-hydra-template. experiment will use the config files in configs/experiment, in which import the corresponding config files in configs/model. The config file name is organized as follows:
| Model Config (configs/model/) | Coref Layer | Encode Pointer | Tree Modeling |
|---|---|---|---|
| t2g_baseline | - | - | - |
| t2g_inline_adapter | - | - | CHA (adapter) |
| t2g_inline_inplace | - | - | CHA (inplace) |
| t2g_point2tgt | Yes | - | - |
| t2g_point2tgt_paware | Yes | Yes | - |
| t2g_point2tgt_paware_strict* | Yes | Yes | - |
| t2g_point2tgt_paware_strict_adapter* | Yes | Yes | CHA (adapter) |
- strict means that the pointers of coref layer are constrained to valid values.
In the training script, a prediction on the test set will be stored under the working directory. To evaluate a trained model on other datasets, use the following command:
python src/eval.py experiment=<model config> data=<data config> ckpt_path=<ckpt path>
Then, you can use the scripts in amrlib to call BLINK to generate AMR graphs with wiki tags and compute the fine-grained metrics. The scripts could be found at https://github.com/bjascob/amrlib/blob/master/scripts/33_Model_Parse_XFM/.
This code is released under the MIT license.
If you use this work for research, please cite our paper:
@inproceedings{
lou2023amr,
title={{AMR} Parsing with Causal Hierarchical Attention and Pointers},
author={Chao Lou and Kewei Tu},
booktitle={The 2023 Conference on Empirical Methods in Natural Language Processing},
year={2023},
url={https://openreview.net/forum?id=SI2CXa5eok}
}