Research project · IIT Guwahati
BERTopic: Autoencoder-based Topic Modelling
Exploring autoencoder-based parameter estimation for BERTopic's dimensionality-reduction stage.
During my research internship at IIT Guwahati, I explored the use of variational autoencoders for parameter estimation within BERTopic, focusing on the UMAP dimensionality-reduction stage and its effect on topic quality.
Approach
- Explored the use of autoencoder-based parameter estimation in BERTopic's UMAP pipeline.
- Evaluated topic quality using coherence scores across a corpus of 19,000 documents.
- Observed a 5-10% improvement in topic coherence over the baseline.
Built with
Research under Prof. Ashok Singh Sairam · May-July 2024