Leverage Reviews Data for Multi Label Topics Classification
summary
Fine-tuned multilingual BERT-like models classify 240 tourism topics across reviews and QA, outperforming off-the-shelf zero-shot encoders and needing only ~200 positives per topic to plateau. Two architectures are compared: cross-encoder vs bi-encoder — cross-encoder yields higher accuracy but bi-encoder is chosen for production speed and saved embeddings to backfill ~240 million reviews. Labeling uses active learning and AWS Sagemaker Ground Truth, explainability via LIME and a Streamlit demo, and thresholds tune F‑beta per use case.
AI-written summary · about
chapters (5)
related talks
31mWhen humans teach machines
01h 26mBagging to BERT- A tour of applied NLP
45mDecoding Topics: A Comparative Analysis of Python’s Leading Topic Modeling Libraries Using Climate C
21mFrom 0 to a crowd sourced neural network in a jiffy
41mBetter Machine Learning with Less Data
30mProductionizing an unsupervised machine learning model to...
35mLeveraging LLM for text augmentation in Topic Modeling
39mModel Fine-tuning For Search: From Algorithms to Infra