A Cross-Domain Transferable Neural Coherence Model
Paper published in ACL 2019: arxiv.org/abs/1905.11912
This implementation is based on PyTorch 0.4.1.
To download the dataset:
python prepare_data.py
which includes WikiCoherence dataset we construct, 300-dim GloVe embeddings and pre-trained Infersent model.
For WikiCoherence, it contains:
- 7 categories under Person
- Artist
- Athlete
- Politician
- Writer
- MilitaryPerson
- OfficeHolder
- Scientist
- 3 categories from different irrelevant domains:
- Plant
- EducationalInstitution
- CelestialBody
- parsed_wsj: original split for Wall Street Journal (WSJ)
- parsed_random: randomly split all paragraphs of the seven categories under Person into training part and testing part
Check config.py for the data_name for each setting.
Premute the original documents or paragraphs to obtain the negative samples for evaluation:
python preprocess.py --data_name <data_name>
Train the LM with the following command:
python train_lm.py --data_name <data_name>
python train_lm.py --data_name <data_name> --reverse
The pre-trained models will be saved in ./checkpoint.
To evaluate our proposed models:
python run_bigram_coherence.py --data_name <data_name> --sent_encoder <sent_encoder> [--bidirectional]
where sent_encoder can be average_glove, infersent or lm_hidden.
python eval.py --data_name <data_name> --sent_encoder <sent_encoder> [--bidirectional]
Run the above script will run the experiment multiple times and report the mean and std statistics.
The log will be saved in ./log.
If you found this codebase or our work useful, please cite:
@InProceedings{xu2019cross,
author = {Xu, Peng and Saghir, Hamidreza and Kang, Jin Sung and Long, Teng and Bose, Avishek Joey and Cao, Yanshuai and Cheung, Jackie Chi Kit},
title = {A Cross-Domain Transferable Neural Coherence Model}
booktitle = {The 57th Annual Meeting of the Association for Computational Linguistics (ACL 2019)},
month = {July},
year = {2019},
publisher = {ACL}
}