From LDA to BERTopic — Podcast Discourse Analysis
Transformer-based topic modelling of ASR-transcribed political podcasts
A comparison of classical and transformer-based topic modelling applied to political podcasts. Working from Whisper-transcribed audio in the SPoRC podcast corpus, I built a BERTopic model on transformer embeddings and benchmarked it against Latent Dirichlet Allocation.
The BERTopic pipeline achieved a topic-coherence score (C_v) above 0.731, a substantial improvement in interpretability over the LDA baseline. The resulting topics were used to explore political-communication and audience-engagement patterns across the podcast ecosystem.
Stack: Python, BERTopic, LDA, Whisper, sentence embeddings.
Code: github.com/nimathing2052/SPoRC_TopicModeling.
Coursework: Natural Language Processing