Rushin Shah, Engineering Manager, Facebook
summary
Facebook deploys a production NLP stack that labels tens of thousands of topics, runs sequence labeling with LSTM/CRF and CNN pipelines, and uses a DeepText platform to centralize embeddings and features. Document models favor CNNs for speed and long-post accuracy; fastText 30–40x faster training provides minute-scale iterations. A two-tower CNN/LSTM ranking model gives state-of-the-art semantic similarity; entity linking resolves noisy band and place names. A self-service system called Clue (content and language understanding engine) trains ~200 models per week and powers Marketplace listings and social recommendations.
AI-written summary · about
chapters (25)
- 0:00Introduction
- 0:45Mission Statement
- 1:35Machine Learning Conferences
- 2:06Natural Language Understanding
- 2:25Text Classification
- 3:45Word Classification
- 4:00Content Similarity
- 4:20Named Entities
- 5:27How did you start investing in Deep Learning
- 6:30What are the main problems you want to solve with Deep Learning
- 7:14What is the basic idea of Deep Learning
- 8:14What is the platform
- 8:26What is text classification
- 9:11Khan nets for document classification
- 9:36Fast text
- 9:57Simple model
- 10:52Sequence labeling
- 12:04Two tower model
- 12:43Entity linking
- 13:20Marketplace tab
- 14:17Social recommendations
- 16:09Maximum leverage
- 16:42Reusing models
- 17:07Selfservice model
- 17:46Summary
related talks
03mApplied machine learning at Facebook
21mFrom 0 to a crowd sourced neural network in a jiffy
26mEwa Dominowska, Engineering Manager, Facebook @ MLconf SEA
35mIndustrial Applications of Machine Learning in Search
22mGenerating a Billion Personal News Feeds
24mMachine Intelligence gets real
43mSentiment Analysis in Action: Building Your Real-time Pipeline
55mDemystifying Natural Language Processing across several industry verticals