Treebanking mystery: Have parses from ACE but FFTB doesn't see/has them?

Someone please help me figure out why the following happens:

I am running the following pydelphin command on a corpus:

The .sh script:

delphin process --options="-1 -p -y --yy-rules --max-chart-megabytes=24000 --max-unpack-megabytes=36000" -g ~/delphin/SRG/grammar/srg/ace/srg.dat --full-forest --select i-tokens $profile

So, ACE is run through pydelphin, in YY-mode, in full-forest mode (the -1 flag I think doesn’t matter here; I keep it because it seems like only with it ACE prints a meaningful message about how many sentences it parsed?).

I then compare the result with a previous version of the same treebank, and I see that I lose a sentence:

Screenshot from 2024-05-29 10-59-51

I haven’t made any changes related to this, so I am surprised.

So I try ACE separately, with the same version of the grammar, and I have parses, including the one I want (ACE shows lemmas instead of surface forms; please disregard that):

What am I missing? I triple-checked that the grammar is the same grammar. What else could lead to this?.. Something in the pydelphin? something related to full-forest?..

More details: if I try this sentence in the LKB, it does not parse it, the reason seems to be that it fails at some point to use the desired lexical analysis for hace. (I have so far failed to understand why, with interactive unification; everything seems to work interactively; not sure yet where I do things wrong.) But why am I getting the parse with ACE separately then? Again, it does seem to be the same grammar file.

I remember that particular sentence being problematic in the past (it exceeded the default ACE and LKB resource limits) until you changed the lexical entry for hace in early March. For me in the LKB it now parses easily, in less than 1 second for either the top-ranked parse or for the full set of 257 possible parses. (There’s still a lot of local ambiguity, though, since it creates over 10000 edges). Anyway, there must be something different between our setups.

1 Like

Yes, there must be something weird related to the lexical entry… When I say I didn’t change anything, I mean I don’t think I touched it recently, but indeed, this is the analysis I did change not so long ago. Can’t be an accident.