Parsing natural languages in Swift
summary
Builds a greedy transition based parser trained entirely in Swift for TensorFlow and converts it for Core ML conversion into an iOS app that visualizes dependency trees. Training uses the Universal Dependencies corpus plus GloVe embeddings and yields about 75% labeled attachment score and 80% unlabeled accuracy with only ~0.01% loss after conversion. Implements a SwiftProtobuf converter, shows the demo app parsing sentences and notes failures on ambiguous examples like "I bought a shirt with buttons."
AI-written summary · about
chapters (9)
- 0:00Start
- 5:31LEFT Transition
- 6:01RIGHT Transition
- 6:39Example
- 9:02Finding a Dataset
- 12:48Swift for Tensorflow
- 13:58Training
- 14:22Testing
- 15:08Available Tools
related talks
17mExploring Natural Language Processing
43mGetting Sentimental: exploration of iOS sentiment analysis
22mImproving Swift Tools with libSyntax
25mGetting Started with Machine Learning in your iOS apps
47mSwifty TensorFlow
27mAST Meta-programming
31mPower-Up Your App With Apple’s Machine Learning Frameworks
53mSwift: New Paradigms for iOS Development