Discourse-Based Sentence Splitting

This repository contains the code and resources from the following paper.

Install

pip install -r requirements.txt
pip install -e .

Download Dataset

Datasets described in the paper are available to download from data/dsplit_data.zip. This contains all of D_CCNews_C, D_CCNews_S, D_MUSS, and D_WikiSplit.

Building a Synthetic Dataset

Alternatively, you can build your own version of the data. You should use this repo to mine sentences with discourse connectives or sentence pairs with an adverbial from the Common Crawl News Corpus. Once you have some mined data, you can use tools in this package to compile samples for training/testing models.

To build a dataset:

python disco_split/processing/build_dataset.py mine <mined_articles_dir> <output_file> <sample_limit>

The --item_type arg can be used to specify whether you are extracting from a mine of complex or simple texts.

To transform a simple sentence dataset into (T, S) sequence pairs:

# NOTE: see the actual script for optional args
python disco_split/processing/build_dataset.py s2t <dataset_file> <output_file>

You can also use this library as an interface to the PDTB parser (original implementation here), which can be used to generate trees for complex sentences.

# an example with the parsing configuration used in the paper
python disco_split/evaluation/pipeline.py prepare_trees <dataset_file> --out_trees=<trees_output> --out_parsed=<parser_output> --text_col=sentence --fallback=naive --batched=True --num_procs=8 --batch_size=128

In order to generate synthetic complex sentences from simple sentence pairs using our rule-based approach, run the following script:

python disco_split/processing/build_dataset.py s2c <input_file> <output_file>

If you would like to specifically reformat intermediate parser ouput into T sequences, you may use the following:

# NOTE: see the actual script for optional args
python disco_split/processing/build_dataset.py a2t <new_oracles_csv> <output_file>

Training a Model

Code for training models is located in disco_split/models/.

To finetune a baseline BART model:

python disco_split/models/train_bart.py --data_file=<dataset_file>
        --train_split=0.95
        --max_epochs=5
        --gpus=1
        --learning_rate=3e-5
        --batch_size=16
        --max_source_length=64
        --max_target_length=64
        --eval_beams=4
        --eval_max_gen_length=64
        --max_samples=1e6
        --val_check_interval=0.2

Additional arguments can be found in disco_split/models/bart.py. These include options for multiple input files, loading checkpoints, specifying save directory, columns names, etc.

Evaluation

You can evaluate end-to-end models from the terminal as follows:

disco_split/evaluation/evaluate.py evaluate <model_loc> <test_data> <output_file> --samsa=True

Alternatively, to evaluate pipeline models:

# using parser C2T component
python disco_split/evaluation/pipeline.py evaluate <t2s_model> <test_data> <output_dir> --parser=pdtb

# using trained neural C2T component
python disco_split/evaluation/pipeline.py evaluate <t2s_model> <test_data> <output_dir> --parser=bart --bart_model=<c2t_model>

Citation

If you find this repository useful, please cite our publication Discourse-Based Sentence Splitting.

@inproceedings{cripwell-etal-2021-discourse-based,
    title = "Discourse-Based Sentence Splitting",
    author = {Cripwell, Liam  and
      Legrand, Jo{\"e}l  and
      Gardent, Claire},
    booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2021",
    month = nov,
    year = "2021",
    address = "Punta Cana, Dominican Republic",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.findings-emnlp.25",
    doi = "10.18653/v1/2021.findings-emnlp.25",
    pages = "261--273",
}