AMR parsing expression grammar

[mpertierra]

Is there an official parsing expression grammar (PEG) for AMR? I've tried writing one myself and this is the closest I've come up with, although I'm sure it's not complete.

constant <- r'([^ \t\n()":]+)|("[^"]+")|(-?[0-9]+)|(-?[0-9]+\.[0-9]+)';
variable <- r'[^ \t\n()":]+';
concept <- r'([^ \t\n()":]+)|("[^"]+")';
relation <- r':[^ \t\n()":]+';
amr <- constant / ( '(' variable '/' concept (relation amr)* ')' );
root <- amr EOF;

[nschneid]

An unofficial one which may not be perfect: https://github.com/nschneid/amr-hackathon/blob/master/src/amr.peg

[mpertierra]

@nschneid That looks great, thank you! Have there been significant changes to the AMR specs since you wrote that?

[nschneid]

Not significant changes, but there may be edge cases that show up in newer data. E.g., negative numbers (nschneid/amr-hackathon#2).

[mpertierra]

@nschneid it seems that smatch is now being hosted on GitHub (https://github.com/snowblink14/smatch). The link to the repository is found here (http://amr.isi.edu/evaluation.html). It does not use a PEG to read in AMR annotations and instead it processes them by character. Is this the closest there is to an "official" AMR annotation parser?

[nschneid]

I believe so. The smatch implementation has been tested more thoroughly then mine.

[danielhers]

Unfortunately the smatch version does not handle alignments, which @nschneid's version does. But the conventions for these seem to be completely nonexistent.

[danielhers]

What are the differences with respect to the grammar used for general PENMAN notation? https://github.com/goodmami/penman#penman-notation

[goodmami]

@danielhers, note that I'm not affiliated with ISI or the AMR group, so my module is not official in any way. I haven't done an actual comparison, but I intentionally made the grammar a bit more forgiving. E.g., I allow node identifiers (variables) to contain anything that isn't a space, a / or a ), and for strings I allow anything inside the quotes, even escaped quotes. I don't (yet) do character alignments, because, as you say, there is no established format I've seen outside of @nschneid's implementation. In the default configuration, I also allow empty nodes (e.g. (a)) and anonymous relations ((a : b)). I don't believe these are valid in AMR, or possibly even in the original PENMAN SPL. I hope to offer an AMR-specific subclass of my PENMANCodec class that restricts the grammar to what is allowed in AMR (and also to manage inversions like :domain <-> :mod), but I haven't had a chance to do this yet (contributions are welcome; and thanks for your previous PR).

Also, on the README I defined the PEG grammar that the module abides by, but I don't use a PEG parsing algorithm (e.g. packrat or even recursive descent). But decoding is pretty well tested: https://github.com/goodmami/penman/blob/master/tests/test_penman.py#L44-L139

[danielhers]

Thanks @goodmami, your module is very useful to me as it allows dynamically creating AMRs easily and printing them (which @nschneid's doesn't really yet). It might not be official but it does the job :)