Online Large-Margin Training of Syntactic and Structural Translation Features - 2008

Explore advanced techniques in statistical machine translation with this lecture on online large-margin training of syntactic and structural translation features. Delve into the limitations of minimum-error-rate training (MERT) and discover how the MIRA algorithm can be applied to overcome these challenges. Learn about two novel feature classes that address deficiencies in the Hiero hierarchical phrase-based translation model. Examine the implementation of soft syntactic constraints and a structural distortion model, and understand how these improvements lead to significant enhancements in translation performance. Gain insights into parallel processing methods and forest-based approaches that make MIRA computationally competitive with MERT. Follow along as the speaker, David Chiang, presents joint work with researchers from the University of Maryland, offering a comprehensive look at cutting-edge developments in machine translation technology.

Introduction
Overview
Example Sentence
Whats Missing
Syntactic Ignorance System
Soft Syntactic Constraint
Training Limitations
Training Data
The Problem
Improvements
Model Score
Selecting Translations
Selecting EJs
Parallelization
When to Stop
Results
Parallelization method
Summary