Efficient Decoding with Labelled Grammars in Moses

Participant: Philip Williams

Initial ideas:

1. An alternative, compact and fast-loading file format for in-memory rule tables

  • List distinct source / target phrases and word alignments once only
  • Heading information about numbers of source / target non-terminals, distinct phrases...
  • Moses should continue to support old format, so need tool to convert between formats

2. Sharing of common target phrase objects to save memory

  • Though could involve substantial refactoring...

3. A variant of the current SCFG rule lookup algorithm that uses unlabelled dotted rules

  • Change prefix tree representation of rule table: edges are terminals and 'gaps' (not label pairs), vertices map lists of non-terminals to translation options
  • Dotted rules use gaps instead of specific non-terminals
  • Labels checked on successful rule completion
  • Already have prototype implementation (target syntax only) but needs integration and extension to support source syntax
Page last modified on September 05, 2011, at 12:07 PM