This is the implementation of Making Better Use of Bilingual Information for Cross-Lingual AMR Parsing in Findings of ACL 2021
Before training, download the pretrained S2S-AMR-Parser in this repo
The training data in English we used is AMR 2.0(here). To get training data and development data in DE, IT, ES and ZH, use MarianMT(here) or other machine translation systems.
The test data in DE, IT, ES and ZH can be found here.
Use codes in AMR_scripts/ to preprocess and postprocess the AMR datasets.
Preprocessing:
python var_free_amrs.py -f sample_input/sample.txt
This will produce a text file and an AMR sequence file. Follow this repo to apply tokenization and BPE to original text file and AMR sequence file. Translate the text file to get {lang}_train.txt and {lang}_dev.txt (DE, IT, ES and ZH).
Postprocessing:
first remove BPE of outputs
sed -r 's/(@@ )|(@@ ?$)//g' sample.txt.bpe > sample.txt
then run the code
python postprocess_AMRs.py -f sample.txt
python reformat_single_amrs.py -f sample.txt.restore.final -e .form
The AMR file for evaluation is the sample.txt.restore.final.form
Here is a command demo for training (several paths should be changed in train.py:L234-237)
python train.py --model s2s_amr_parser_path --prefix train_data_folder_path --prefix_dev dev_data_folder_path --save_prefix save_folder_path --xlm_r_path xlmr_folder_path
and predicting
python cross_translate.py --decode_extra_length 1000 --minimal_relative_prob 0.01 --gpu 0 --src your_input_file_path --translate_input your_eng_input_file_path --output your_output_path --model_path your_model_path --xlmr_path xlmr_folder_path
Notice that model_path for predicting is the model trained with this code instead of the S2S-AMR-Parser. Our temporary best model is in https://drive.google.com/file/d/1SOJ0fiXpWUCkstBVq-6G9-ed5xjv330w/view?usp=sharing
We adapt the codes from S2S-AMR-Parser and RikVN/AMR. Thanks to their open-source projects.