When humans teach machines
summary
A self‑service pipeline lets domain experts build sparse-topic text classifiers by: creating a visual query to find dense positives, extracting triage terms via a significance score to auto-label negatives from a production corpus, and training a fast SVM linear model with iterative active-labeling (stop after seven consecutive negatives). UX fixes include highlights and save‑for‑later. Deployment on Refinitiv runs in production with 300 topics in production and time per topic falling from 26 to 13 hours; alternatives explored: ELMo/BERT with hierarchical sentence‑level attention.
AI-written summary · about
chapters (10)
- 0:00Start
- 2:20Why Do We Need Text Classification
- 10:31User Experience Ux
- 11:25Morphological Expansions
- 18:57Create a Test Set
- 20:26Intelligent Tagging
- 23:28Alternatives to Svm
- 23:38Deep Learning Methods
- 27:31Gamification
- 28:50How Do You Handle Multiple Languages or Do You Just Look at the English Newspapers
related talks
20mEmploying AI to extract actionable insights from qualitative user...
30mProductionizing an unsupervised machine learning model to...
35mIndustrial Applications of Machine Learning in Search
22mImplementing a Text Classifier based on Lucene & LIBSVM
33mImpact of NLP in AI-powered Bank
47mAI vs. Karen: Real-Time Customer Sentiment Detection at Scale
01h 29mTricks, tips and topics in Text Analysis
31mCreate text classifiers in a few hours using the open-source Label Sleuth system