Code for our ACL2020 paper,
AMR Parsing via Graph-Sequence Iterative Inference [preprint]
Deng Cai and Wai Lam.
The code has two branches:
The code has been tested on Python 3.6. All dependencies are listed in requirements.txt.
We use Stanford CoreNLP (version 3.9.2) for lemmatizing, POS tagging, etc.
sh run_standford_corenlp_server.sh
The input file should constain the raw sentences to parse (one sentence per line).
Data Preprocessing: sh preprocess_raw.sh ${input_file}
(when use graph recategorization, first download artifacts via sh download_artifacts.sh
)
sh work.sh
=> {load_path}{output_suffix}.pred
sh postprocess_2.0.sh
{load_path}{output_suffix}.pred
=> {load_path}{output_suffix}.pred.post
Model | Link |
---|---|
AMR2.0+BERT+GR=Smatch80.2 | amr2.0.bert.gr.tar.gz |
AMR2.0+BERT=Smatch78.7 | amr2.0.bert.tar.gz |
The following instruction assumes that you're training on AMR 2.0 (LDC2017T10). For AMR 1.0, the procedure is similar.
Unzip the corpus to data/AMR/LDC2017T10
.
Prepare training/dev/test splits:
sh prepare_data.sh -v 2 -p data/AMR/LDC2017T10
Download Artifacts:
sh download_artifacts.sh
Feature Annotation:
We use Stanford CoreNLP (version 3.9.2) for lemmatizing, POS tagging, etc.
sh run_standford_corenlp_server.sh
sh annotate_features.sh data/AMR/amr_2.0
Data Preprocessing:
sh preprocess_2.0.sh
Building Vocabs
sh prepare.sh data/AMR/amr_2.0
sh train.sh data/AMR/amr_2.0
The training process will produce many checkpoints and the corresponding output on dev set. To select the best checkpoint, one can evaluate the dev output files (need to do postprocessing first). It is recommended to use fast smatch for model selection.
For evaluation, following Parsing with Pretrained Models step 2-3, then sh compute_smatch {load_path}{output_suffix}.pred.post data/AMR/amr_2.0/test.txt
.
We adopted the code snippets from stog for data preprocessing.
The dbpedia-spotlight occasionally does not work. Therefore, we have disabled it.
For any questions, please drop an email to Deng Cai.