BertTopic, From Free-Text feedbacks to Calls for Action
summary
Shows a concrete pipeline that converts free text into actionable topics using sentence-transformers embeddings, dimensionality reduction with UMAP + HDBSCAN, and iterative hyperparameter tuning. Demonstrates an initial silhouette score of 0.25 with >200 topics and 21% outliers, then tuning yields ~11 topics, near zero outliers and a much higher silhouette score. Builds a working BertTopic model, uses representative docs plus an LLM for readable names, and delivers interactive visualizations and topic‑over‑time plots.
AI-written summary · about
chapters (15)
- 0:00Topic modeling intro
- 2:27Student feedback dataset
- 4:57Manual vs automated
- 6:22What is BERTopic and its advantages
- 7:29Minimal BERTopic example
- 10:19How BERTopic works
- 13:36Default BERTopic models
- 15:06Better topic modeling
- 16:54Optimizing hyperparameters for better results
- 17:57Insights from topics
- 18:41Naming topics with ChatGPT
- 20:50Actionable insights features
- 21:32Visualize and merge topics
- 22:15Track topic trends
- 23:21Key takeaways
related talks
34mTopic modeling with BERTopic
32mBERTopic to Accelerate Ukrainian Aid by the Red Cross
01h 28mTricks, tips and topics in Text Analysis
38mData Mapping for Data Exploration
20mBI Toolkit Powered by Natural Language Processing
19mHow to look at your data
47mAI vs. Karen: Real-Time Customer Sentiment Detection at Scale
39mTemporal Analysis on Topics Utilizing Word2Vec