Code for Effective Neural Topic Modeling with Embedding Clustering Regularization (ICML2023)

Check our latest topic modeling toolkit TopMost !

ICML 2023

Usage

1. Prepare environment

torch==1.7.1
scipy=1.7.3
scikit-learn==0.23.2
gensim==4.0.1
pyyaml==6.0

Prepare coherence evaluation:

1. Install java.

    sudo apt install openjdk-11-jdk
   

2. Download $C_V$ java jar to ./ECRTM/palmetto. It is developed by palmetto.

3. Download and extract preprocessed Wikipedia articles to ./ECRTM/palmetto/wikipedia as the reference corpus.

2. Train and evaluate the model

We provide a shell script ./ECRTM/scripts/run.sh to train and evaluate our model.

Change to directory ./ECRTM, and run commands as

./scripts/run.sh ECRTM 20NG 50
./scripts/run.sh ECRTM IMDB 50
./scripts/run.sh ECRTM YahooAnswer 50
./scripts/run.sh ECRTM AGNews 50

Preprocess datasets (Optional)

Datasets in ./data have been preprocessed before. Here we provide a shell script to show how we preprocess these datasets:

./scripts/preprocess.sh

This can be used to preprocess other datasets.

Citation

If you want to use our code, please cite as

@inproceedings{wu2023effective,
    title={Effective neural topic modeling with embedding clustering regularization},
    author={Wu, Xiaobao and Dong, Xinshuai and Nguyen, Thong and Luu, Anh Tuan},
    booktitle={International Conference on Machine Learning},
    year={2023},
    organization={PMLR}
}