Building new NLP solutions with spaCy and Prodigy
summary
Claim: NLP projects fail when teams design pipelines and annotation schemes upfront instead of iterating. Built and used spaCy and Prodigy with Thinc-backed models to compose text classification + NER composition instead of bespoke labels, saving annotation effort. Annotation pace hits one decision every 5–10 seconds so an hour yields a usable prototype, a 200 annotations pilot gives early signal, and A/B evaluation in 10–20 minutes lets small tweaks be judged quickly. Dependency parses need domain tuning; in-house small teams and semi-automatic workflows work better than mass outsourcing.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 2:21NLP projects are like start-ups
- 6:12Machine Learning Hierarchy of Needs
- 11:45Making modeling decisions that are simple, obvious, and wrong (Problem #1)
- 17:05Compose generic models into novel solutions (Solution #1)
- 19:05Workflow #1
- 22:35Big annotation projects make evidence collection expensive (Problem #2)
- 24:00Run your own micro-experiments (Solution #2)
- 27:38It is hard to get good data by boring out underpaid people (Problem #3)
- 28:43Smaller teams, better workflows (Solution #3)
related talks
18mReality is not an end-to-end prediction problem
29mAnnotating data the right way
33mAdvanced Natural Language Processing with Free Low-Code and...
27mThe necessity of pipelines in Natural Language Processing
17mNatural Language Processing in Automation
30mDeveloper Productivity for GenAI Apps
30mBuilding Reliable AI Agents for Publishing: A DSPy-Based Quality Assurance Framework
15mNatural Language-Guided Programming