End-to-End Argument Mining over Varying Rhetorical Structures

This repository contains the official code and data for the ACL 2023 Findings paper End-to-End Argument Mining over Varying Rhetorical Structures.

Environment setup

bash src/dependencies_and_data.bash
source py39/bin/activate

The bash script sets up the environment and collects the corpus (283 documents total) in a single data/data.pkl file. Run the script first, then install the appropriate pytorch version for your cuda if necessary.

1. Data preparation

1.1 Back translation

⚠️ Processing documents in our setting takes around 15 s/document on average.

The results from the paper are given in data/data.pkl.

Click to view the steps to generate the data/data.pkl file (Optional)

Run the tchewik/isanlp_hf_translator Docker image on the same or remote machine with the following command:

docker run --rm --name translator -p 3332:3333 -d tchewik/isanlp_hf_translator

The script requires the IP or domain name of the machine (servername) and the source language (english or russian) as arguments:

python src/data/run_back_translation.py --servername servername

It updates the data/data.pkl file with back translations: Ru -> En and En -> Ru.

For BLUE evaluation, run:

python src/data/evaluate_back_translation.py --target_lang english|russian

1.2 RST parsing

⚠️ In our setting, processing one document takes roughly 3 seconds (English) or 5 seconds (Russian) on average.

The NLP annotations obtained in our experiments are given in data/nlp_annot.zip. Unpack it into data/ to skip this step: cd data && unzip nlp_annot.zip.

Click to view the steps to reproduce the data/nlp_annot/ directory (Optional)

To analyze texts on either local or remote machines, use these commands:

For English:

# spaCy: syntax port 3333
docker run --rm -d -p 3333:3333 --name spacy_en tchewik/isanlp_spacy:en
# RST parser: parser port 3222
docker run --rm -d -p 3222:3333 --name rst_en tchewik/isanlp_zhang21

For Russian:

# spaCy: syntax port 3334
docker run --rm -d -p 3334:3333 --name spacy_ru tchewik/isanlp_spacy:ru
# RST parser: parser port 3222
docker run --rm -d -p 3222:3333 --name rst_ru tchewik/isanlp_rst:2.1-rstreebank

Wait for the parsers to load completely, then run the script:

python src/data/run_rst_parsing.py \
       --servername servername \
       --syntax_port <number> \
       --rst_port <number> \
       --lang english|russian

It will save the linguistic annotations in the data/nlp_annot/ directory.

1.3 RST to Microtexts mapping

Run the next command to collect the RST-like structures with ADU as terminal nodes into the data/rst_shrinked/ directory:

python src/data/map_rst_to_microtexts.py --lang english|russian

1.4 Convert RST do dependencies

Run the next command to convert the RST-like structures into dependencies for DBAP models under data/conll_cv_rst/ directory:

python src/data/convert_rst_to_dep.py 

The next command converts the RST-like structures into EDU-based dependencies for e2e models under data/conll_cv_rst_end2end/ directory:

python src/data/convert_rst_to_e2e_dep.py

2. Experiments

2.1 Training

All the required training configuration files are in the initial_configs/ directory.

# Parsing over handcrafted segmentation
bash src/{bap|dbap}_train_and_predict.bash --lang en|ru --train_data en|ru|en_aug|ru_aug

# End-to-end
bash src/e2e_{bap|dbap}_train_and_predict.bash --lang en|ru --train_data en|ru|en_aug|ru_aug

2.2 Evaluation

The script requires the dependencies the evidencegraph virtual environment: source baseline/evidencegraph/py38-evgraph/bin/activate.

# For MODELNAME in {bap|dbap|e2e_bap|e2e_dbap}:
(py38-evgraph) python utils/crossval/eval_parsing.py --dir models/MODELNAME/MODELNAME_{en|ru|en_aug|ru_aug} --k 10

# e2e evaluation without 'same-arg' relation:
(py38-evgraph) python utils/latex_report_line.py --model_name models/MODELNAME/MODELNAME_{en|ru|en_aug|ru_aug}

Citation

@inproceedings{chistova-2023-end,
    title = "End-to-End Argument Mining over Varying Rhetorical Structures",
    author = "Chistova, Elena",
    booktitle = "Findings of the Association for Computational Linguistics: ACL 2023",
    month = jul,
    year = "2023",
    address = "Toronto, Canada",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.findings-acl.209",
    pages = "3376--3391",
}