search

bjascob / amrlib

A python library that makes AMR parsing, generation and visualization simple.
MIT License
219 stars 34 forks source link

more AMR-2-text models? #48

Closed haixpham closed 2 years ago

haixpham commented 2 years ago

Hello,

Thanks for the awesome work! Your library helps simplify the task on AMR a lot!

I wonder, do you have plan to support more AMR2text generative models? I find the T5 model can't handle deeply nested AMR.

bjascob commented 2 years ago

I don't have any plans right now for new generate models. From the scores, the T5generate model should be near SOTA. What does "can't handle" mean? Crash, produces poorly formed sentences or is it missing content from the original AMR graph?

haixpham commented 2 years ago

In most cases it produces poorly formed sentences from deeply nested AMR graphs.

bjascob commented 2 years ago

You'd probably have to figure out some type of "surface planning" model to break the AMR graphs into sub-graphs and determine the correct order for serialization. The current method of serializing based on depth-first recursion is probably not sufficient for the deep graphs.

It's also possible that training a T5-large model would give improved results but you might also need to train with significantly more deep graph AMR data.

Sound like a good research project for someone. Here's a link to a paper where they did planning on raw data (not AMR) which might be an interesting place to start.