This directory contains finite-state transducers implementing (reversible)
variant conversion between languages.  Given a transducer X->Y and a
putative inverse transducer Y->X, we also generate bracketing transducers
which can split an input string into "safe" reversible segments (where
X->Y->X is the identity) and "unsafe" segments where the conversion
loses information.  We can then preserve extra metadata to allow
lossless conversion even for "unsafe" segments.

The primary files in this directory are named `<language code>.foma`
and are the input to the FOMA finite-state transducer toolset; see
https://fomafst.github.io/ .  We require a slightly-forked version of
foma: https://github.com/cscott/foma on the wmf-fork branch.  There
are test cases and examples in `<language code>-examples.foma`.  Other
files ending in `.foma` bundle up helper tranducers for reuse across
languages.

Running foma on the primary `<language code>.foma` files will generate
`.att`-format transducer graphs.  The tool in
`$PARSOID/tools/build-langconv-fst.js` can then be used to
compile/compress these into binary files (by convention, `.pfst`
files, for "Parsoid FST"), which can then be executed by
`lib/language/FST.js`.  The intention is to allow execution from a
direct mmap of the `pFST` files to minimize the memory and
startup-time overhead of loading a potentially large number of
conversion machines.