Hi this is the source code of our paper "ATP: AMRize Then Parse! Enhancing AMR Parsing with PseudoAMRs" accepted by findings of NAACL 2022.
- 🎈 Release slide (Chinese). Google Drive 2022.07.08
- 🎈 Release camera ready paper. arXiv 2022.04.20
- 🎈 We have released four trained models and the test scripts. 2022.04.10
- 🎯 We are working on merging our training/preprocessing code with the amrlib repo.
TL;DR: SOTA AMR Parsing single model using only 40k extra data. Rank 1st model on Structrual-Related Scores (SRL and Reentrancy).
As Abstract Meaning Representation (AMR) implicitly involves compound semantic annotations, we hypothesize auxiliary tasks which are semantically or formally related can better enhance AMR parsing. With carefully designed control experiments, we find that 1) Semantic role labeling (SRL) and dependency parsing (DP), would bring much more significant performance gain than unrelated tasks in the text-to-AMR transition. 2) To make a better fit for AMR, data from auxiliary tasks should be properly ``AMRized'' to PseudoAMR before training. 3) Intermediate-task training paradigm outperforms multitask learning when introducing auxiliary tasks to AMR parsing.
From an empirical perspective, we propose a principled method to choose, reform, and train auxiliary tasks to boost AMR parsing. Extensive experiments show that our method achieves new state-of-the-art performance on in-distribution, out-of-distribution, low-resources benchmarks of AMR parsing.
Build envrionment for Spring
cd spring
conda create -n spring python=3.7 && conda activate spring
pip install -r requirements.txt
pip install -e .
# we use torch==1.11.0 and A40 GPU. lower torch version is fine.Build envrionment for BLINK to do entity linking, Note that BLINK has some requirements conflicts with Spring, while the blinking script relies on both repos. So we build it upon Spring.
conda create -n blink37 -y python=3.7 && conda activate blink37
cd spring
pip install -r requirements.txt
pip install -e .
cd ../BLINK
pip install -r requirements.txt
pip install -e .
bash download_blink_models.sh- AMRization for SRL with reentrancy restoration
cd amrization
unzip amrization.zip # OntoNotes 5.0 and amrization code
cd amrization
python get_srl.py # get the linearized srl after amrized
python split_train_dev_test.py # get the train-val split of srl dataset
cd spring/bin- Train ATP-SRL Task
python train.py --direction dp --config /home/cl/AMR_Multitask_Inter/spring/configs/config_srl.yaml
# yes, the direction is also dp, you should change the train/dev/test file path in the config_srl.yaml file- Train AMR Task based on intermediate ATP-SRL/DP Model
python train.py --direction amr --checkpoint PATH_TO_SRL_DP_MODEL --config ../configs/config.yaml
- Train AMR,SRL,DP Task in multitask Manner
python train.py --direction multi --config ../configs/config_multitask.yamlconda activate spring
cd script
bash intermediate_eval.sh MODEL_PATH
# it will generate the gold and the parsed amr files, you should the change the path of AMR2.0/3.0 Dataset in the script.
conda activate blink37
# you should download the blink models according to the ATP/BLINK/download_blink_models.sh in BLINK repo
bash blink.sh PARSED_AMR BLINK_MODEL_DIR
cd ../amr-evaluation
bash evaluation.sh PARSED_AMR.blink GOLD_AMR_PATHYou could refer to the inference section and download the models below to reproduce the result in our paper.
- ATP_SRL_AMR2.0 Google Drive
#scores
Smatch -> P: 0.858, R: 0.844, F: 0.851
Unlabeled -> P: 0.890, R: 0.874, F: 0.882
No WSD -> -> P: 0.863, R: 0.848, F: 0.855
Concepts -> P: 0.914 , R: 0.895 , F: 0.904
Named Ent. -> P: 0.928 , R: 0.901 , F: 0.914
Negations -> P: 0.756 , R: 0.758 , F: 0.757
Wikification -> P: 0.849 , R: 0.824 , F: 0.836
Reentrancies -> P: 0.756 , R: 0.744 , F: 0.750
SRL -> P: 0.840 , R: 0.830 , F: 0.835- ATP_SRL_D_AMR2.0 Google Drive
#scores
Smatch -> P: 0.859, R: 0.844, F: 0.852
Unlabeled -> P: 0.891, R: 0.876, F: 0.883
No WSD -> -> P: 0.863, R: 0.849, F: 0.856
Concepts -> P: 0.917 , R: 0.898 , F: 0.907
Named Ent. -> P: 0.942 , R: 0.921 , F: 0.931
Negations -> P: 0.742 , R: 0.755 , F: 0.749
Wikification -> P: 0.851 , R: 0.833 , F: 0.842
Reentrancies -> P: 0.753 , R: 0.741 , F: 0.747
SRL -> P: 0.837 , R: 0.830 , F: 0.833- ATP_SRL_D_AMR2.0_Ensemble Google Drive
#scores
Smatch -> P: 0.859, R: 0.847, F: 0.853
Unlabeled -> P: 0.891, R: 0.877, F: 0.884
No WSD -> -> P: 0.863, R: 0.851, F: 0.857
Concepts -> P: 0.917 , R: 0.899 , F: 0.908
Named Ent. -> P: 0.938 , R: 0.917 , F: 0.927
Negations -> P: 0.740 , R: 0.755 , F: 0.747
Wikification -> P: 0.849 , R: 0.830 , F: 0.840
Reentrancies -> P: 0.755 , R: 0.748 , F: 0.751
SRL -> P: 0.837 , R: 0.836 , F: 0.836- ATP_SRL_AMR3.0_Ensemble Google Drive
#scores
Smatch -> P: 0.844, R: 0.836, F: 0.840
Unlabeled -> P: 0.875, R: 0.866, F: 0.871
No WSD -> -> P: 0.849, R: 0.840, F: 0.845
Concepts -> P: 0.908 , R: 0.892 , F: 0.900
Named Ent. -> P: 0.900 , R: 0.879 , F: 0.889
Negations -> P: 0.734 , R: 0.729 , F: 0.731
Wikification -> P: 0.816 , R: 0.798 , F: 0.807
Reentrancies -> P: 0.729 , R: 0.749 , F: 0.739
SRL -> P: 0.822 , R: 0.830 , F: 0.826We thank all people/group that share open-source scripts for this project, which include the authors for SPRING, amrlib, smatch, amr-evaluation, BLINK and all other repos.
If you feel our work helpful, please kindly cite
@inproceedings{chen-etal-2022-atp,
title = "{ATP}: {AMR}ize Then Parse! Enhancing {AMR} Parsing with {P}seudo{AMR}s",
author = "Chen, Liang and
Wang, Peiyi and
Xu, Runxin and
Liu, Tianyu and
Sui, Zhifang and
Chang, Baobao",
booktitle = "Findings of the Association for Computational Linguistics: NAACL 2022",
month = jul,
year = "2022",
address = "Seattle, United States",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2022.findings-naacl.190",
pages = "2482--2496",
abstract = "As Abstract Meaning Representation (AMR) implicitly involves compound semantic annotations, we hypothesize auxiliary tasks which are semantically or formally related can better enhance AMR parsing. We find that 1) Semantic role labeling (SRL) and dependency parsing (DP), would bring more performance gain than other tasks e.g. MT and summarization in the text-to-AMR transition even with much less data. 2) To make a better fit for AMR, data from auxiliary tasks should be properly {``}AMRized{''} to PseudoAMR before training. Knowledge from shallow level parsing tasks can be better transferred to AMR Parsing with structure transform. 3) Intermediate-task learning is a better paradigm to introduce auxiliary tasks to AMR parsing, compared to multitask learning. From an empirical perspective, we propose a principled method to involve auxiliary tasks to boost AMR parsing. Extensive experiments show that our method achieves new state-of-the-art performance on different benchmarks especially in topology-related scores. Code and models are released at \url{https://github.com/PKUnlp-icler/ATP}.",
}