Bilingual Rhetorical Structure Parsing

This repository contains the official code and data for the ACL 2024 Findings paper Bilingual Rhetorical Structure Parsing with Large Parallel Annotations.

Trained Models

This repository focuses on data and experiments. For applying the trained parsers, visit the IsaNLP RST repository for models and usage instructions.

Data

The data directory structure should be as follows:

data/
├── gum_rs3/
│   ├── en/
│   │   └── *.rs3
│   └── ru/
│       └── *_RU.rs3
├── rstdt_rs3/
│   ├── TEST/
│   │   └── wsj_*.rs3
│   └── TRAINING/
│       └── wsj_*.rs3
└── rurstb_rs3/
    ├── train.*_part_*.rs3
    ├── dev.*_part_*.rs3
    └── test.*_part_*.rs3

gum_rs3/ru/ Contains the RRG corpus in Russian. data/RRG.zip
gum_rs3/en/ Place the GUM RST *.rs3 files here. GUM dataset link.
rstdt_rs3/ Place the RST-DT *.rs3 files here. RST-DT dataset link.
rurstb_rs3/ Contains the RRT corpus; one document = one tree. data/rurstb_rs3.zip

The train/dev/test splits for GUM/RRG are listed under data/gum_file_lists for GUM v9.1. If you are using a later extended version, you should update these file lists accordingly.

Experiments

Set WANDB_KEY in dmrst_parser/keys.py for online wandb support.

Monolingual Experiments

Train:

python dmrst_parser/multiple_runs.py --corpus "$CORPUS" --lang "$LANG" --model_type "$TYPE" --cuda_device 0 train

Evaluate:

python dmrst_parser/multiple_runs.py --corpus "$CORPUS" --lang "$LANG" --model_type "$TYPE" --cuda_device 0 evaluate

Bilingual Experiments