A Crash Course in Topic Modelling with BERTopic
This is a recording from a November 2024 talk JGI Data Scientist Dr Huw Day gave to the climate dynamics group at the University of Bristol. Topic modelling is the task of sorting text into clusters based on themes and labelling those clusters appropriately. This talk gives an intuition of how the process of BERTopic topic modelling (https://maartengr.github.io/BERTopic/... , https://arxiv.org/abs/2203.05794) works under the hood and some example code and visualisation techniques. A copy of the slides and Huw's code from this project can be found on his GitHub: https://github.com/HuwWDay/ClimateTop...
Huw worked with Dr Eunice Lo and Dr Jo Godwin on a project looking at the lived experiences of people during extreme climate events in Bristol and used topic modelling to augment their analysis. This project has been supported by the Jean Golding Institute for data science and data-intensive research at the University of Bristol.
00:00 - 02:00: Introduction to the Jean Golding Institute
02:01 - 02:57 Introduction to the project
02:58 - 07:34 What is topic modelling?
07:35 - 08:51 Overview of BERTopic modelling
08:52 - 10:57 Step 0: Pre-processing
10:58 - 16:49 Step 1: Embed Documents
16:50 - 22:19 Step 2: Dimensionality Reduction
22:20 - 26:45 Step 3: Cluster documents
26:46 - 34:30 Step 4: Bag-of-words (Tokenizing Topics)
34:31 - 36:13 Step 5: Topic representation
36:14 - 38:05 Step 6: Fine-tune Topic representation
38:06 - 44:15 Putting it all together, diagnostics, and fine tuning
44:16 - 44:20 Questions
44:21 - 48:08 Other cool things you can do with BERTopic
48:09 - 49:00 Outro
Check out our website for more events: https://www.bristol.ac.uk/golding/
Subscribe to our newsletter: https://www.bristol.ac.uk/golding/mai...
Read our latest blog posts: https://jeangoldinginstitute.blogs.br...
Follow us on social media:
Bluesky: @jgibristol.bsky.social
LinkedIn: / jeangoldinginstitute