Skip to content

OsloInfoMorph

EmilyBender edited this page Aug 8, 2017 · 2 revisions

Discussion:

DI: look at langs with radically dicont morph (athabaskan)

BC: Navajo has been looked at (Koenig?)

GE: Panini: beyond current delphin machinery?

BC: sufficient: if in your domain you say I have constraints A and B that apply only [???] you can make it explicit by applying closure [???] -- it is a finite domain so it should be possible. e.g. constraint A can apply only to NOM.SG. ultimate default: would accumulate all the negative constriants

DI: not y, not z etc

GE: one big type hierarchy

BC: Malouf proposed using topological sort, impose order in which rules apply. But how that would work for generation? problem: if you want to associate XXX with parsing, specific morphosyntactic property except ones that have more specific shape

GE: DELPHIN grammars should be bidirectional

BC: are type hierarchies bit vector-encoded?

SO: yes

GE: [???]

BC: would it be hard to compute the complement in this case

GE: you can create tdl automatically

BC: where you use sets is the more challenging part.

DF: sets are hard in our universe

BC lists will be ok but I need to be able to say: "I know there is X at the end of the list but there is Y that comes before that"

WP: you want sets of atomic objects?

BC: of non-recursive feature structures. equivalent to atoms?

WP: if it is small inventory of atoms then it is similar to a large type hierarchy

EB: size of the set would be bound by number of atoms in the inventory

BC dont want to do closure operation by hand.

WP satement in tdl: i have this list of atoms and [???]

BC: typically you have finite domains. you can have a lot of complexity e.g. 30K of [???] in Koenig's work. With reasonable preprocessing we can get close to [???]. In parsing, can split rules but that has to factor in what you done in the inheritance because ... that's a hunch.

EB: what is the problem with processing this on the textual level?

BC spot which rule specifies something replace by two rules; a rule can introduce two exponents. Problem with doing it textually: constraints can come from two diff dimensions,

EB: constraints on what is being expressed? no harm in having both rules say "case nom". we are doing it in the customization system.

BC: you dont have all the info on the textual level about the leaves

EB: why does that matter? we can spit out a modified grammar

BC: once you have done the [expansion?] e.g. run the korean grammar with YADU dump out the tdl and then [???] ... but this didn't happen and Jong-Bok removed defaults from his grammar.

EB: Fokkens manipulates read-in in tdl in "spring cleaning". at the end what we are manipulating are grammar entities

BC: when to do the check? when is something fully inflected

DI: keep the inflection of the [???] rule

BC: if I have no way of establishing of [???] to inflect this, I need e.g. agreement information.

EB: we have the FLAG feature (Goodman & Bender), it's an unordered set features of the same type just like other feature structures are. and you can check the inflections are satisfied

BC: but we also have other sets

MG: one rule can satisfy any number of requirements

BC: accumulation or any fusion would quality for that. common property of any inflectional systems, even in agglutinating langs.

Clone this wiki locally