Skip to content

DocumentationParty

Alexandre Rademaker edited this page Jun 2, 2021 · 11 revisions

Morphological analyzer notes

Existing documentation:

What's SPPP? In the same space as chart mapping, may become redundant. Predates chart mapping. Might also overlap with REPP, supported in LKB and soon in PET.

Send question to Rebecca: Will SPPP be part of extended support for PET?

[oe, 15-jul-11] SPPP is unrelated to chart mapping. It is rather yet another way of feeding partially annotated token lattices into the parser; by and large, SPPP is an XML rendering of the good old YY format. SPPP currently is only supported in the LKB; currently there are no plans for supporting it in PET. Rather, FSC should probably be augmented (to support annotation of token edges with mandatory applications of orhtographemic rules; see my email to 'developers' of today) and also be implemented in the LKB (to replace SMAF; this should not be too hard, but Ann or oe would seem the two most viable candidates to actually do the implementation).

Current REPP processing does damage for English to the input text because TNT requires punctuation stripped. That's lossy --- throwing spaces around a hyphen etc. Doesn't arise in terms of generation.

ERG generation -- some lisp code at the end provides some patching up like sentence-initial capitalization.

Ask Montse about generation and also the question on ChartMappingAppliedNotes from Paris meeting.

Montse's reply: SRG not currently used for generation but FreeLing is bidirectional. Regarding the question on ChartMappingAppliedNotes:

"Actually I had several reasons:

  • to follow the same approach as the other systems, I think I was the only one using SPPP.
  • to keep FL and SRG as two independent tools, though, we've already managed to do that with SPPP.
  • to integrate chunks into the SRG, that is, just as we translate PoS chunks into Feature Structures, I want to translate (some) chunk mark-ups into feature structures. Though this may be also done with SPPP, not sure.

Yes, I still want to integrate chunks. As for the way to do it... I'd rather follow the most efficient strategy, so I'd be happy to change the strategy if it is worth."

What about Jacy? Chasen on the way out? BURGER very large fully inflected form dictionary

Chart mapping is fully supported in PET and planned for the LKB. But mostly for robustness over real phenomena real corpora. Other use is passing along morphological ambiguity to the parser. Montse could want this too in the LKB for Spanish. (por tanto] tiempo] v. [por [tanto tiempo)

Time involved in chart mapping is mostly in the rules to create more or fewer tokens.

No chart mapping in the LKB yet, so that takes us back to SPPP as the recommended connection in the parsing direction. Nothing in the generation direction yet.

No one in DELPH-IN currently going through a morphological analyzer/generator on the way out, but this is of interest to the Matrix group so we can continue to claim that morphophonology does not have to be handled within the Matrix-based grammar.

Posssible MA/MS project:

  • Hook up existing reversible morphophonological analyzer to existing grammar in the generation direction
  • Do this in a way that it can be added to the customization system as an option.

Montse/Luis FreeLing might be good beta testers

OpenFST --- should be reversible SPRouT --- reversible?