A toolkit for segmenting Elementary Discourse Units (clauses). We implement it as is described in our EMNLP paper: Toward Fast and Accurate Neural Discourse Segmentation
- Python 3.5
- Tensorflow>=1.5.0
- allennlp>=0.4.2
- See
requirements.txt
for the full list of packages
We cannot provide the complete RST-DT corpus due to the LDC copyright.
So we only put several samples in ./data/rst/
to test the our code and show the data structure.
If you want to train or evaluate our model on RST-DT, you need to download the data manually and put it in the same folder. Then run the following command to preprocess the data and create the vocabulary:
python run.py --prepare
We provide the vocabulary and a well-trained model in the ./data/
folder. You can evaluate the performance of this model after preparing the RST-DT data as mentioned above:
python run.py --evaluate --test_files ../data/rst/preprocessed/test/*.preprocessed
The performance of current model should be as follows:
'precision': 0.9176470588235294, 'recall': 0.975, 'f1': 0.9454545454545454}
Note that this is slightly better than the results we reported in the paper, since we re-trained the model and there is some randomness here.
You can use the following command to train the model from scratch:
python run.py --train
Hyper-parameters and other training settings can be modified in config.py
.
You can segment files with raw text into EDUs:
python run.py --segment --input_files ../data/rst/TRAINING/wsj_110*.out --result_dir ../data/results/
The segmented result for each file will be saved to the --result_dir
folder with the same name. Each EDU is written as a line.
Please cite the following paper if you use this toolkit in your work:
@inproceedings{wang2018edu,
title={Toward Fast and Accurate Neural Discourse Segmentation},
author={Wang, Yizhong and Li, Sujian and Yang, Jingfeng},
booktitle={Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing},
pages={962--967},
year={2018}
}