Apertium-neural

Apertium was originally developed to offer a free/open-source framework for creating RBMT systems. It was modelled on existing systems, but targetted at related languages, trying to do one thing well.

What might an Apertium NMT system for lesser-resourced and marginalised languages look like?

Thoughts:

• Trains without GPU or large compute
• Optimised for small corpora (under 100k parallel sentences)
• Includes linguistic tricks
• C++, autotools
• Works with existing tools (formatters, APY etc.)

Pipeline(?):

apertium-destxt | apertium-preprocess | apertium-encode | apertium-decode | apertium-retxt

Backend:

• Own built(?)
• DyNet --- forked?