Files for the repo on Faithful Neural Topic Modeling. The main objective of the thesis is to understand how the inbuilt metric of cTF-IDF of BERTopic matches up against other intrinsic metrics borrowed from XAI into Topic Modeling, namely Topic Changes in the context of Comprehensiveness and Sufficiency and Centroid Movements.

You should clone the repo and cd into it, as a first step. We suggest creating a new virtual environment in python and installing the requirements with pip install -r requirements.txt.

1. Firstly you have to use the main function to save the results in a suitable location which can be accessed later on. The two main files for comprehensiveness and sufficiency studies respectively are src/main_comprehensiveness and src/main_sufficiency.py. Example : python -m src.main_sufficiency --path results/keybert/nyt/ --dataset data/nyt2020.csv --column text --k 100 --model 1 where

• --path : path to save our results in.
• --dataset : path to load our data from.
• --column : column of the dataset csv to load the data from.
• --k : top k topics for which to run the ablations, out of the total generated topics.
• --model : a number denoting which model to use :
  1. TF-IDF
  2. KeyBert
  3. PoS
  4. MMR
  5. Randomization

2. Secondly it is easy to run the Gradio interface to interact with the representative graphs per topic. Just run python -m src.gradio_app.app After that, based on your choice, you can load and see the difference in rankings for different topics.

3. To get overall statistics, one should run the following command to run the process_results.py script. It can be run as : python -m process_results --mode comprehensiveness --paths /home/abpal/WorkFiles/Faithful-Topic-Modeling/result1/comprehensiveness/wiki/model_4,/home/abpal/WorkFiles/Faithful-Topic-Modeling/result2/comprehensiveness/wiki/model_4 --total_topics 100 --total_words 10 --intervals 5,7,10 --dataset wiki where

• --mode : mode of function to process results for, either comprehensiveness or sufficiency
• --paths : paths to our saved raw results to calculate the correlation and find it's mean and standard deviation.
• --total_topics : total topics for which to calculate the results for. For example, setting it to 100 means we calculate the results for the top 100 topics.
• --total_words : total words per topics for which to calculate the results for, ranges from 1 to 10.
• --intervals : The interval represents the top-k representative words for which we will calculate the percent of documents changed/same depending on the mode.
• --dataset : Mainly used to store the results. For example for wiki dataset we will store the results in /final_result/{mode}/wiki/model_x.

LDA results were generated using the LDA.ipynb notebook. They were consumed for further processing inside randomized_rankings_and_visualizations.ipynb. The latter also contains a few analyses regarding the randomization model, and scripts to create the boxplots we used to explain the results.

Centroid Movement Results were obtained using the centroid_movement.ipynb notebook.

The 20NewsGroups dataset is accessed from scikit-learn. Run src/save_data.py as python -m src.save_data --path /data to save the dataset and access it later.

• Thanks to the Social Computing Group at TUM for Compute Resources.
• Logo Courtesy : DALL-E