/XSRL-ACL

Primary LanguagePython

codes for the ACL 2020 paper: Cross-Lingual Semantic Role Labeling with High-Quality Translated Training Corpus

Cite:

@inproceedings{fei-etal-2020-cross,
    title = "Cross-Lingual Semantic Role Labeling with High-Quality Translated Training Corpus",
    author = "Fei, Hao and Zhang, Meishan and Ji, Donghong",
    booktitle = "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
    year = "2020",
    url = "https://www.aclweb.org/anthology/2020.acl-main.627",
    pages = "7014--7026",
}

A. Translation-based projection

  • Step1. Train POS taggers for each languages based on the corresponding labelled POS data (with universal POS tagset).

    • trainPOS.py (config the Config/config.cfg in advance)

  • Step2. Prepare the translated SRL parallel data.

    • The translation process should be fulfilled by you own. You can achieve it via Google translation API.
    • The format of source side data and translated target side data should follow the example data in Data/upb_parallel/en-de fold: en-de-train-src.conllu and en-de-train-tgt.conllu. It's conllu style.
    • We provide the UPB-English data which was not offered by UPBV1.0 at the time (i.e., 2019/10/1) we perform our experiments.

  • Step3. Generating the aligning file for each pair of source and target language dataset.

    • Install the fast_align, and conduct alignment.
    • The format of the alignment output file follows the example file en-de-train-src2tgt-align.prob in Data/upb_parallel/en-de fold.

  • Step4. Start annotation projection.

    • project.py

Note:

  • Please note that one sentence may contain multiple sets of prd-args structure. So you should pre-process the data and split them in advance, making sure that one sentence in the data only at maximum contains one set of prd-args proposition.

B. PGN-LSTM SRL model

  • step 1. Configure the setting file Config/config.cfg

  • step 2. Run with main.py

Note:

  • Environment dependency:

    • python3.7
    • pytorch1.0.1
    • allennlp
    • huggingface-bert
    • tqdm

  • Download the UPBV1.0 data. Please pre-process the data, making sure that one sentence in the data only at maximum contains one set of prd-args proposition.