This is code that accompanies the paper titled "Topic Modeling in Embedding Spaces" by Adji B. Dieng, Francisco J. R. Ruiz, and David M. Blei. (Arxiv link: https://arxiv.org/abs/1907.04907)
ETM defines words and topics in the same embedding space. The likelihood of a word under ETM is a Categorical whose natural parameter is given by the dot product between the word embedding and its assigned topic's embedding. ETM is a document model that learns interpretable topics and word embeddings and is robust to large vocabularies that include rare words and stop words.
The major project dependency are :
- python 3.6.7
- pytorch 1.1.0
With or without a virtual environment install you can install the other project requirements with:
pip install -r requirement.txt
All the datasets are pre-processed and can be found below:
- https://bitbucket.org/franrruiz/data_nyt_largev_4/src/master/
- https://bitbucket.org/franrruiz/data_nyt_largev_5/src/master/
- https://bitbucket.org/franrruiz/data_nyt_largev_6/src/master/
- https://bitbucket.org/franrruiz/data_nyt_largev_7/src/master/
- https://bitbucket.org/franrruiz/data_stopwords_largev_2/src/master/ (this one contains stop words and was used to showcase robustness of ETM to stop words.)
- https://bitbucket.org/franrruiz/data_20ng_largev/src/master/
All the scripts to pre-process a given dataset for ETM can be found in the folder 'scripts'. The script for 20NewsGroup is self-contained as it uses scikit-learn. If you want to run ETM on your own dataset, follow the script for New York Times (given as example) called data_nyt.py
To learn interpretable embeddings and topics using ETM on the 20NewsGroup dataset, run
python main.py --mode train --dataset 20ng --data_path data/20ng --num_topics 50 --train_embeddings 1 --epochs 1000
To evaluate perplexity on document completion, topic coherence, topic diversity, and visualize the topics/embeddings run
python main.py --mode eval --dataset 20ng --data_path data/20ng --num_topics 50 --train_embeddings 1 --tc 1 --td 1 --load_from CKPT_PATH
To learn interpretable topics using ETM with pre-fitted word embeddings (called Labelled-ETM in the paper) on the 20NewsGroup dataset:
- first fit the word embeddings. For example to use simple skipgram you can run
python skipgram.py --data_file PATH_TO_DATA --emb_file PATH_TO_EMBEDDINGS --dim_rho 300 --iters 50 --window_size 4
- then run the following
python main.py --mode train --dataset 20ng --data_path data/20ng --emb_path PATH_TO_EMBEDDINGS --num_topics 50 --train_embeddings 0 --epochs 1000
@article{dieng2019topic,
title={Topic modeling in embedding spaces},
author={Dieng, Adji B and Ruiz, Francisco J R and Blei, David M},
journal={arXiv preprint arXiv:1907.04907},
year={2019}
}
