Scale-Invariant Infinite Hierarchical Topic Model (ihLDA)

Paper

Shusei Eshima and Daichi Mochihashi. 2023. Scale-Invariant Infinite Hierarchical Topic Model. In Findings of the Association for Computational Linguistics: ACL 2023. Link.

Requirements

Python 3.9.6

Cython==0.29.28
gensim==4.2.0
matplotlib==3.5.1
nltk==3.7
numpy==1.23.4
pandas==1.4.2
scikit-learn==1.0.2

Usage

Data Preparation

$ python preprocessing.py

The input/sample_raw folder contains ten sample documents for testing purposes (note that this is not enough data to obtain any meaningful results).

Fitting the Model

$ python setup.py build_ext --inplace
$ python main.py --output_path ./output/

# or if the default settings are fine, just run
$ python run.py

The output folder contains the following files:

• fig_tssb/: the structure of the root tree.
• model/: the output is saved every 1000 iterations.
• filenames.csv: the list of file names and doc_ids.
• info.csv: the number of topics.
• parameters.csv: the hyperparameters for each iteration.
• perplexity.csv: the perplexity.
• TopWords_prob.csv: the topic-word distribution of top words.
• model_temp.pkl: the temporary model object. This allows us to resume the iteration, but the random seed will be reset if you resume the iteration.
• txtdata.pkl: the data object.
• settings.txt: the settings of the model.

Evaluation

evaluate.ipynb calls the evaluation function in evaluate_helper.py.