Leveraging LLM for text augmentation in Topic Modeling
summary
Demonstrates generating synthetic text with GPT‑3.5 Turbo to expand small seeds (300 TweetEval examples and subsets of OpenFoodFacts), filters out low‑quality outputs (about 15% mismatch rate on sentiment checks), then trains a topic pipeline using BERTopic + UMAP + HDBSCAN on LLM‑augmented descriptions. Reports a high interpretability metric (0.943 coherence score) and a production use case: augmented topics feed a backend recommender, improving cluster clarity versus raw data.
AI-written summary · about
chapters (13)
- 0:00LLM-powered text augmentation intro
- 1:05What is data augmentation
- 2:52Topic modeling fundamentals
- 3:56Insufficient training data challenges
- 4:59LLM data augmentation steps
- 8:55Live sentiment augmentation coding
- 12:07LLM augmentation for topic modeling
- 13:32Enhancing Open Food Facts dataset
- 20:43Live topic modeling demo
- 25:45Evaluating coherence scores
- 27:57Production use cases
- 29:43Bias and data quality concerns
- 32:38Q&A
related talks
24mHarnessing the Power of LLMs in Advanced Analytics
37mExpanding the Capabilities of LLMs Through Community Collaboration
52mUnleash the Power of Large Language Models (LLMs)
25mWhen Your Dataset Has Blind Spots LLM-Based Data Augmentation
57mLLMs gone wild
45mHalf hour of labeling power: Can we beat GPT?
27mEmbeddings, Transfer Learning and More: LLM Tailored to You
36mBuilding & Deploying LLM Apps