Travatar is a new decoder, so there are still a number of things that are yet to be implemented. Here is a short list, in approximate order of priority:
- Finish the implementation of online learning algorithms such as MIRA (and integration with mert-travatar.pl.
- On-disk storage of the rule table and language model.
- More efficient storage of the rule table in memory.
- Create better phrase table filtering scripts, based on entropy filtering, etc.
- Implement phrase-sense disambiguation.
- Implement a clone of the Berkeley parser within Travatar to remove dependency on external software and remove hacky replacement of failed parses.
- Build KenLM's language model estimator together with travatar, and use it as the default in the tutorial.
- Expand the ability to use target-side syntax.
- Better rule table filtering code filter-rt.pl (currently this does not filter all the rules that could be filtered).
If you are interested in helping to implement any of these or any other features that you would like to see in Travatar, we'd love to have your help! Please send mail to the Travatar-users mailing list, which you can join by sending mail to travatar-users+subscribe with the domain googlegroups.com. (You can also send questions directly to neubig at gmail.com.)