Inducing Synchronous Grammars for Machine Translation

Explore the latest advancements in machine translation through this comprehensive lecture by Phil Blunsom from the University of Edinburgh. Delve into the application of probabilistic machine learning to model translation, focusing on inducing synchronous context-free grammars to capture latent structures essential for translating between divergent languages like Chinese and English. Learn about non-parametric Bayesian models, variational Bayes and Gibbs sampling inference procedures, and their effectiveness in full-scale translation evaluations. Gain insights into complex structured problems in language processing, including language modeling, parsing, and grammar induction. The lecture covers topics such as generative processes, degenerate solutions, P0 distribution, synthetic and real translation experiments, single category models, collapsed Gibbs samplers, priors, split-merge operators, and reparent operators, concluding with a demonstration and results analysis.

Intro
Outline
Probabilistic Model
generative process
generative model
Degenerate solution
P0 distribution
In practice
Synthetic experiment
Real translation experiment
Single category model
Collapsed Gib sampler
Priors
Split Merge Operator
Reparent Operator
Demo
Initialisation
Results
Distribution