A code for "Tree-Structured Neural Topic Model" in ACL2020

Corresponding paper: https://www.aclweb.org/anthology/2020.acl-main.73/

Masaru Isonuma, Juncihiro Mori, Danushka Bollegala, and Ichiro Sakata (The University of Tokyo, University of Liverpool)

Python==3.6

Run the following script to install required packages.

pip install -r requirements.txt

Download the raw data and put bags_and_cases.trn to data/bags/ from
https://drive.google.com/uc?id=1Vt_Pnby63OgB1NK-2qwT_K4mryEXMQ-J&export=download
(The data is distributed in https://github.com/stangelid/oposum)

Run the following script:

python preprocess_oposum.py -path_data </path/to/raw/data> -path_output </path/to/preprocessed/data>

Download the raw data and put them to data/20news/ from
https://github.com/akashgit/autoencoding_vi_for_topic_models/tree/master/data/20news_clean
(The data is distributed in https://github.com/akashgit/autoencoding_vi_for_topic_models)

Run the following script:

python preprocess_20news.py -dir_data </dir/of/raw/data> -path_output </path/to/preprocessed/data>

Run the following script:

python train.py -gpu <index/of/gpu> -path_data </path/to/preprocessed/data> -dir_model <path/to/model/directory>

The trained parameters are saved in dir_model.
The corpus in dir_corpus are used for calculating coherence score (NPMI).

Run the following script:

python evaluate.py -gpu <index/of/gpu> -path_model <path/to/model/checkpoint> -dir_corpus <path/to/corpus>

The scores and topic frequent words are displayed in the console.
You can also use our checkpoint in model/bags/checkpoint_stable.
(Although the scores on this checkpoint slightly differ from the scores in the paper, the difference does not influence the claim of the paper.)

The module to calculate NPMI (coherence.py) is based on the code:
https://github.com/jhlau/topic_interpretability