Using the gr command with probabilities

[Dan Haub]

Hi Everyone!

Over the last few months, I've been working in Grammatical Framework with the ACE-in-GF grammar (big thanks to Inari for all your support with that). My project involves using machine translation techniques to translate English sentences into Attempto Controlled English (ACE) sentences in a very similar way to existing natural language to code systems. Currently, I'm struggling to figure out how to decode the output of a neural network such that every sentence it produces conforms to the ACE grammar. I'm aware of a few systems that decode a NN directly into an abstract syntax tree, but they all seem to require the grammar of the output language to be significantly simpler than the grammar for ACE.

While exploring the documentation for the GF shell, I noticed that there's a way to provide probabilities for the generate random command, although I couldn't seem to get it to work. Can anyone provide some more information on how this works and how to use it? Also, does anyone know if it's possible to leverage this feature for a neural network decoder system? Using the ACE grammar that has already been developed would be a massive help for my project.

Thank you all very much,

Dan Haub

Chapman University