- Ubuntu 16.04
- Python 3.6.5
- chainer 5.4.0
- cupy 5.4.0
- gensim 3.8.0
- scipy 1.3.0
- nltk 3.4.5
- pandas 0.25.0
- progressbar2 3.42.0
$ pip install pipenv --user
$ pipenv install
$ ./download.sh # the resultant data size will be 3.5GB
The configuration files used in the experiments are in ./config.
To start training, run ./src/train.py with a configuration file.
The results will be written into ./result
$ pipenv run python src/train.py "config/descript/descript-Seq2seq-batchsize.64-epoch.100-lr.0.001-n_layers.2-n_units.300"
To evaluate trained models, run ./src/test.py with a glob pattern to grep result directories.
$ pipenv run python src/test.py "./result/descript/*" # specify a pattern to glob result directories
To evaluate trained models leaned with different random seeds, run ./src/test.py with multiple glob patterns.
$ pipenv run python src/test.py "./result/descript-seed-0/*" "./result/descript-seed-1/*" "config/descript-seed-2/*"
To generate next events using a trained model, run ./src/generate_interactively.py with a glob pattern to grep result directories.
$ pipenv run python src/generate_interactively.py "./result/descript/*"
- code: MIT License
- data: GNU General Public License, version 2
@inproceedings{kiyomaru-etal-2019-diversity,
title = "Diversity-aware Event Prediction based on a Conditional Variational Autoencoder with Reconstruction",
author = "Kiyomaru, Hirokazu and
Omura, Kazumasa and
Murawaki, Yugo and
Kawahara, Daisuke and
Kurohashi, Sadao",
booktitle = "Proceedings of the First Workshop on Commonsense Inference in Natural Language Processing",
month = nov,
year = "2019",
address = "Hong Kong, China",
publisher = "Association for Computational Linguistics",
url = "https://www.aclweb.org/anthology/D19-6014",
doi = "10.18653/v1/D19-6014",
pages = "113--122",
abstract = "Typical event sequences are an important class of commonsense knowledge. Formalizing the task as the generation of a next event conditioned on a current event, previous work in event prediction employs sequence-to-sequence (seq2seq) models. However, what can happen after a given event is usually diverse, a fact that can hardly be captured by deterministic models. In this paper, we propose to incorporate a conditional variational autoencoder (CVAE) into seq2seq for its ability to represent diverse next events as a probabilistic distribution. We further extend the CVAE-based seq2seq with a reconstruction mechanism to prevent the model from concentrating on highly typical events. To facilitate fair and systematic evaluation of the diversity-aware models, we also extend existing evaluation datasets by tying each current event to multiple next events. Experiments show that the CVAE-based models drastically outperform deterministic models in terms of precision and that the reconstruction mechanism improves the recall of CVAE-based models without sacrificing precision.",
}