Unified Neural Topic Model via Contrastive Learning and Term Weighting (EACL2023)

Requirements

• Python 3.7
• PyTorch 1.8.0
• transformers 4.19.0
• numpy 1.21.4
• scipy 1.7.1
• pandas 1.3.4
• umap-learn 0.5.2
• scikit-learn 0.24.2
• sentence-transformers 2.2.0

Code is not tested with other versions.

Data

You can add your own data by refering to the data class in data.py. The data should be in the format of a list of documents, where each document is a string.

Usage

ContrastiveTM-main.ipynb is the main notebook for training and evaluating the model.

TODO

• Modularize the code
• Make it into an installable package
• Test on different Python/PyTorch versions

Acknowledgements

Some parts of the implementation is based on the code from contextualized-topic-models.