See lme4 manual (http://lme4.r-forge.r-project.org/book/Ch2.pdf) and Bolker's article.
Nested random effects (1|a/b) are actually equivalent to (1|a) + (1|b), as long as levels in b are clearly specified so they can't be confounded. See also https://m-clark.github.io/mixed-models-with-R/extensions.html#crossed-vs.nested

https://github.com/sfischer13/python-arpa/blob/2284b815866aeb08f65f786da416e78d7937ee1d/arpa/models/base.py#L34-L45

This try...catch mechanism to implement the backoff may not be efficient enough.
According to the python documentation:

A try/except block is extremely efficient if no exceptions are raised. Actually catch

On the GIT main page, there is a reference to:
Memory Based Word Predictor/Language Model http://ilk.uvt.nl/wopr/

This link is broken.
It would be nice to have this page back again, or at least a link to the Wopr pdf.