Skip to content

Latest commit

 

History

History
75 lines (64 loc) · 5.4 KB

README.md

File metadata and controls

75 lines (64 loc) · 5.4 KB
Raw

Introduction

  • One implementation of the paper DMRST: A Joint Framework for Document-Level Multilingual RST Discourse Segmentation and Parsing and Multilingual Neural RST Discourse Parsing.
  • Users can apply it to parse the input text from scratch, and get the EDU segmentations and the parsed tree structure.
  • The model supports both sentence-level and document-level RST discourse parsing.
  • We trained and evaluated the model with the multilingual collection of RST discourse treebanks, and it natively supports 6 languages: English, Portuguese, Spanish, German, Dutch, Basque. Interested users can also try other languages.
  • This repo and the pre-trained model are only for research use. Please cite the papers if they are helpful.

Package Requirements

The model training and inference scripts were tested on following libraries and versions:

  1. pytorch==1.7.1
  2. transformers==4.8.2

Training: How to convert treebanks to our format for this framework

  • Follow the treebank pre-processing steps in the two sub-folders under Preprocess_RST_Data.
  • Note that the XLM-Roberta-base tokenizer is used in both treebank pre-processing and model training scripts. If you want to use other tokenizers, you should change them accordingly.
  • After all treebank pre-processing steps, samples will be stored in pickle files (the output path is set by user).
  • Since some treebanks need LDC license, for model re-training, in this repo we only provide one public dataset GUM (Zeldes, A., 2017) as an example.
  • The exampled pre-processed treebank GUM (Zeldes, A., 2017) (English-only) is located at the folder ./depth_mode/pkl_data_for_train/en-gum/.

Training: How to train a model with a pre-processed treebank

  • Run the script MUL_main_Train.py to train a model.
  • Before you start to train, we recommend that you read the parameter settings.
  • The pre-processed data in folder ./depth_mode/pkl_data_for_train/en-gum/ (English-only) will be used for training by default, as an example.
  • Note that the XLM-Roberta-base tokenizer is used in both treebank pre-processing and model training scripts. If you want to use other tokenizers, you should change them accordingly.

Inference: Supported Languages

  • Instead of re-training the model, you can use the well-trained parser for inference (model checkpoint is located at ./depth_mode/Savings/).
  • We trained and evaluated the model with the multilingual collection of RST discourse treebanks, and it natively supports 6 languages: English, Portuguese, Spanish, German, Dutch, Basque. Interested users can also try other languages.

Inference: Data Format

  • [Input] InputSentence: The input document/sentence, and the raw text will be tokenizaed and encoded by the xlm-roberta-base language backbone.

    • Raw Sequence Example:
    • Although the report, which has released before the stock market opened, didn't trigger the 190.58 point drop in the Dow Jones Industrial Average, analysts said it did play a role in the market's decline.*

  • [Output] EDU_Breaks: The indices of the EDU boundary tokens, including the last word of the sentence.

    • Output Example: [5, 10, 17, 33, 37, 49]
    • Segmented Sequence Example ('||' denotes the EDU boundary positions for better readability):
    • Although the report, || which has released || before the stock market opened, || didn't trigger the 190.58 point drop in the Dow Jones Industrial Average, || analysts said || it did play a role in the market's decline. ||*

  • [Output] tree_parsing_output: The model outputs of the discourse parsing tree follow this top-down constituency parsing format.

    • (1:Satellite=Contrast:4,5:Nucleus=span:6) (1:Nucleus=Same-Unit:3,4:Nucleus=Same-Unite:4) (5:Satellite=Attribution:5,6:Nucleus=span:6) (1:Satellite=span:1,2:Nucleus=Elaboration:3) (2:Nucleus=span:2,3:Satellite=Temporal:3)
    • Moreover, (1:Satellite=Contrast:4,5:Nucleus=span:6) means the first parsing step (EDU1 to EDU6), where EDU4 is the splitting prediction, EDU1:4 (predicted as Satellite) and EDU5:6 (predicted as Nucleus) is one pair with discourse relation "Contrast".

Inference: How to use it for parsing

  • Put the text paragraph to the file ./data/text_for_inference.txt.
  • Pre-trained model checkpoint is located at ./depth_mode/Savings/.
  • Run the script MUL_main_Infer.py to obtain the RST parsing result. See the script for detailed model output.
  • We recommend users to run the parser on a GPU-equipped environment.

Citation

If the work is helpful, please cite our papers in your publications, reports, slides, and thesis.

@inproceedings{liu-etal-2021-dmrst,
    title = "{DMRST}: A Joint Framework for Document-Level Multilingual {RST} Discourse Segmentation and Parsing",
    author = "Liu, Zhengyuan and Shi, Ke and Chen, Nancy",
    booktitle = "Proceedings of the 2nd Workshop on Computational Approaches to Discourse",
    month = nov,
    year = "2021",
    address = "Punta Cana, Dominican Republic and Online",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.codi-main.15",
    pages = "154--164",
}

@inproceedings{liu2020multilingual,
  title={Multilingual Neural RST Discourse Parsing},
  author={Liu, Zhengyuan and Shi, Ke and Chen, Nancy},
  booktitle={Proceedings of the 28th International Conference on Computational Linguistics},
  pages={6730--6738},
  year={2020}
}