Data and code for our paper
@inproceedings{mesgar2021neural,
title={A Neural Graph-based Local Coherence Model},
author={Mesgar, Mohsen and Ribeiro, Leonardo FR and Gurevych, Iryna},
booktitle={Findings of the Association for Computational Linguistics: EMNLP 2021},
pages={2316--2321},
year={2021}
}
We recommend that you organize your files and name them like we did in the file tree
.
├── Experiments # model saved dir
│ └── WSJ
│ └── ours
│ └── 2022_3_13_11_42
│ │── Epoch_3
│ └── log.out
│── data
│ │── GoogleNews-vectors-negative300.bin
│ │── allen_cache
│ │ │── elmo*.json
│ │ └── elmo*.hdf5
│ └── Dataset_Global
│ │── *.py
│ │── train_pos # 1240 postive samples
│ │ │── wsj_0001.pos.text
│ │ │── wsj_0002.pos.text
│ │ └── ...
│ │── train_neg # 1240*40 negative samples, each positive sample has 40 negative permuatations
│ │ │── wsj_0001.pos.text_1
│ │ │── wsj_0001.pos.text_2
│ │ │── ...
│ │ │── wsj_0002.pos.text_1
│ │ │── wsj_0002.pos.text_2
│ │ └── ...
│ │── dev_pos # 138 postive samples
│ │── dev_neg # 138*40 negative samples
│ │── test_pos # 1053 postive samples
│ │── test_neg # 1053*40 negative samples
│ │── EGrid.train_dev # Grid files for postive/negative samples in train and dev sets
│ │ │── wsj_0001.pos.text.parsed.ner.EGrid
│ │ │── wsj_0001.pos.text.parsed.ner.EGrid-1
│ │ └── ...
│ │── EGrid.test # Grid files for postive/negative samples in test set
│ │── Dataset
│ │ │── train # 1240*40 postive-negative pairs
│ │ │ │── wsj_0001.pos.text_1
│ │ │ │── wsj_0001.pos.text_2
│ │ │ └── ...
│ │ │── dev
│ │ │── test
│ │ └── vocab # vocab list for train set
│ │ │── Vocab
│ │ │── word2idx
│ │ └── idx2word
│ │── wsj.train # paths to grid files for positive samples in train set
│ │── wsj.dev
│ └── wsj.test
│── fairseq
│── src
└── experiments_unified.py
- Install necessary dependencies listed in requirements.txt. Please note that if you fail to install torch-* (e.g. torch-scatter), run the following command
>>> pip install torch-scatter==2.0.5 torch-sparse==0.6.7 torch-cluster==1.5.7 torch-spline-conv==1.2.0 torch-geometric==1.6.1 -f https://pytorch-geometric.com/whl/torch-1.6.0+cu101.html- Download GoogleNews-vectors-negative300.bin from https://s3.amazonaws.com/dl4j-distribution/GoogleNews-vectors-negative300.bin.gz and put it under the folder data/
- Regard to WSJ license, we can't upload the raw data. You can contact us to seek dataset help.
- You can run the following command to train our model, other parameters such as paths to datasts can be viewed in src/utils.py
>>> python experiments_unified.py \
--experiment_path ./Experiments/WSJ/ours/ \
--bilinear_dim 32 \
--batch_size_train 1\
--batch_size_test 1\
--ELMo True- You can run run the following command to test our model. We have prepared a trained model under the folder Experiments/
>>> python experiments_unified.py \
--experiment_path ./Experiments/WSJ/ours/ \
--bilinear_dim 32 \
--batch_size_train 1\
--batch_size_test 16\
--ELMo True \
--only_eval True \
--model_name ./Experiments/WSJ/ours/2022_3_13_11_42/Epoch_3