Skip to content


Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time

XL-AMR: Enabling Cross-Lingual AMR Parsing with Transfer Learning Techniques

XL-AMR (Camera ready print) is a cross-lingual AMR parser that exploits the existing training data in English to transfer semantic representations across languages. The achieved results shed light on the applicability of AMR as an interlingua and set the state of the art in Chinese, German, Italian and Spanish cross-lingual AMR parsing. Furthermore, a detailed qualitative analysis shows that the proposed parser can overcome common translation divergences among languages.

If you find either our code or our release datasets useful in your work, please cite us with:

    title = "{XL-AMR}: {E}nabling Cross-Lingual {AMR} Parsing with Transfer Learning Techniques",
    author = "Blloshmi, Rexhina  and
      Tripodi, Rocco  and
      Navigli, Roberto",
    booktitle = "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
    month = nov,
    year = "2020",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "",
    pages = "2487--2500",

1. Install

Create a conda environment with Python 3.6 and PyTorch 1.5.0, install the dependencies requirements.txt and download the artifacts**.

Via conda:

conda create -n xlamr python=3.6
source activate xlamr
pip install -r requirements.txt
bash scripts/    

**Also please unzip all the zipped files you find inside the data folder before continuing with the other steps.

If you need our silver data based on AMR 2.0 translations please contact us by email!

Download XL-AMR best checkpoints* per language here: CHECKPOINTS

*Please take care of the paths in the config file according to your file structure in case of exceptions, before sending an email or opening an issue.

2. Gold Dataset

1 - Download AMR 2.0 (LDC2017T10) and AMR 2.0 - Four Translations (LDC2020T07).

2 - Unzip the AMR 2.0 corpus to data/AMR/LDC2017T10. It should look like this:

(xlamr)$ tree data/AMR/LDC2017T10 -L 2 data/AMR/LDC2017T10
├── data
│   ├── alignments
│   ├── amrs
│   └── frames
├── docs
│   ├── AMR-alignment-format.txt
│   ├── amr-guidelines-v1.2.pdf
│   ├── file.tbl
│   ├── frameset.dtd
│   ├── PropBank-unification-notes.txt
│   └── README.txt
└── index.html

Prepare training/dev/test data:

./scripts/ -v 2 -p data/AMR/LDC2017T10

3 - Unzip the Translations corpus to data/AMR/LDC2020T07 and copy *.txt files into data/amr_2.0/translations/ .

Project English test AMR graphs across languages:


3. Silver dataset for {Chinese, German, Italian, Spanish}

3.1 Annotation Projection through parallel sentences

These data make use of Europarl corpus [1] and the English AMR parser by Zhang et al. 2019 [2] to parse the English side of the parallel corpus.

Data used for training and development are found in the following folder:

cd xl-amr/data/AMR/europarl_en_de_es_it/

*We do not produce silver AMR graphs for Chinese in this approach since Europarl does not cover Chinese language.

3.2 Annotation Projection through automatic translations

We machine translated sentences of AMR 2.0 using OPUS-MT pretrained models and filtered less accurate translations. The translated sentences are found in the following folder:

cd data/AMR/amr_2.0_zh_de_es_it/translations/

To respect the LDC agreement for AMR 2.0, we release the translations without the gold graphs. Therefore to project the AMR graphs from AMR 2.0 run:



4 Preprocessing

4.1 Lemmatization, PoS-tagging, NER:

  • For English we use Stanford CoreNLP (version 3.9.2). Before running the script, start a CoreNLP server following the API documentation.

      cd stanford-corenlp-full-2018-10-05
      java -mx4g -cp "*" edu.stanford.nlp.pipeline StanfordCoreNLPServer -port 9000 -timeout 15000
  • For Italian we use Tint (package tint-runner-0.2-bin.tar.gz). Before running the script, start a Tint server

     ./tools/tint/ -p 9200
  • For German and Spanish we use Stanza. Before running the script, dowload the models.

      import stanza"zh")"de")"es")

Run the script lang -> {en, zh, de, es, it} and dataset_type -> {silver, gold}:

./scripts/ {data_dir} {lang} {dataset_type}

4.2 Re-categorization and Anonymization

To preprocess the training and dev data run:

./scripts/ {data_dir} {dataset_type}

dataset_type -> {silver, gold}

To preprocess the test data run:


5. Training

XL-AMR models are trained in one GeForce GTX TITAN X GPU. We provide the params .yaml files for all the models in the paper.

python -u -m stog.commands.train params/{method}/{xl-amr_configuration}.yaml

method -> {zeroshot, xl-amr_par, xl-amr_trns}

6. Prediction

To evaluate the XL-AMR models run:

./scripts/ {lang} {model_dir}

7. Postprocessing

For postprocessing two steps are needed:

1 - Run wikification using DBPedia Spotlight API on the test sentences for each language. Before running the script, start a docker server (following the instructions here spotlight-docker) for the the specific language model at {port} for example (the italian case):

docker run -itd --restart unless-stopped -p 2230:80 dbpedia/spotlight-italian

then run:

./scripts/ {lang} {port}

lang -> {en, de, es, it}

*For Chinese we use Babelfy as shown in Babelfy HTTP API example.

2 - Run postprocessing script:

./ {lang} {model_dir}

lang -> {en, zh, de, es, it}

8. Evaluation using Smatch and Fine-Grained metrics

./ {lang} {model_dir}


[1] Philipp Koehn. 2005. Europarl: A Parallel Corpus for Statistical Machine Translation. In Conference Proceedings: the tenth Machine Translation Summit, pages 79–86, Phuket, Thailand. AAMT, AAMT.

[2] Sheng Zhang, Xutai Ma, Kevin Duh, and Benjamin Van Durme. 2019. AMR Parsing as Sequence-to-Graph Transduction. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 80–94, Florence, Italy. Association for Computational Linguistics.


This project is released under the CC-BY-NC-SA 4.0 license (see LICENSE). If you use XL-AMR, please put a link to this repo.


The authors gratefully acknowledgethe support of the ERC ConsolidatorGrant MOUSSE No. 726487 and the ELEXIS project No. 731015 under the European Union’s Horizon 2020 research and innovation programme.

This work was supported in part by the MIUR under the grant "Dipartimenti di eccellenza 2018-2022" of the Department of Computer Science of the Sapienza University of Rome.

We adopted modules or code snippets from the open-source projects:

Thank you for making research easier!


XL-AMR is a sequence-to-graph cross-lingual AMR parser that exploits transfer learning (EMNLP2020).








No releases published


No packages published