Natural Language Understanding at Scale
summary
A custom NLU service is built and deployed: Python, PyTorch and C++ back ends, containerised with OpenAPI/Connexion, gunicorn and nginx, and client SDKs generated from the spec. Intent and entity models run at 1–5 ms per request and need about ~1 GB per language model, yielding roughly 200–1000 req/s depending on implementation. Fine‑tuning per vertical raises accuracy from ~90% to >99%, and C++ components improve throughput for heavy loads.
AI-written summary · about
chapters (19)
- 0:00Start
- 0:24About Feersum Engine
- 1:05About Me
- 2:59The Dialog Agent
- 5:42NLU for Automation
- 6:28NLU for Agents
- 7:28Intent Detection
- 8:55Lasso Model Builder
- 9:49Built on Language Models
- 11:05TL;DR - Agents & NLU
- 14:32The NLU Service
- 15:17Swagger YAML Spec.
- 16:47API Language Wrappers
- 17:59Compute Resources
- 20:40Stress Tests Results
- 21:51TL;DR - Scaling Up
- 24:38Deep Learning - Vision
- 26:00Transfer Learning
- 30:46Questions?
related talks
39mFrom data science to scalable NLU and vision cloud service.
30mBuilding Complete State of the art Natural Language Processing
33mState of the art natural language understanding at scale
26mNatural Language Processing (NLP) going mainstream
01h 03mNatural Language Support with LUIS
17mNatural Language Processing in Automation
55mDemystifying Natural Language Processing across several industry verticals
28mThe NLU Orchestra