''Not all those who wrote opposed the changes.''
Most but not quite all... <- can't lexicalize these as complex predicates
all those -> partitive
those -> _those_q_dem_rel, generic_entity_rel
Why _those_q_dem_rel and not _this_q_dem_rel? We don't do _you_n_rel
Are there any differences between those and that as quantifiers?
??that one who
Claim is that those/that are different "inflected" forms of the same lemma
More general question:
- How do we decide what predicate names to use for quantifiers?
''A similar technique is almost impossible to apply to other crops, such as cotton, soybeans and rice.''
_similar_a_to_rel(e9, x2) comp_rel(e10,e9,u11)
A chair similar to Abrams arrived.
Comparatives are a relation between a property and a reference set; the way the reference set can be described is highly variable; dimension is the inherent argument of the adj (or NP) being compared (ARG0 of the adjective = ARG1 of comp_rel) ARG2 is the reference set
... taller than : than should be semantically empty
... similar to : to should also be, but isn't currently (the pred value for similar already has 'to' as a marker for the generator)
Other lexical entries contributing comp_rel: different
That type (comp_rel) exists, and is clustered with some non-trivial syntactic properties:
- That similar one to this one.
The machinery to do the long-distance thing here requires access to the comp_rel. Could split that to allow the lexicalized (similar, different) ones to be single predicates while the derived ones (shorter, more beautiful) have two predicates.
This choice is thus driven by composition, rather than by the resulting representations, but does lead to a parallelism in the representations.
Design principles: decomposition is desirable, but not a goal in itself. If the predicates required for decomposition exist, then it is good.
Compare relational adjectives:
But that's a very different sense of ARG2, and since we don't have a way of what we say they mean. So it's nicer to have comp_rel. But if we had _similar_a_comp_rel, that might encode the distinction too. _tall_a_comp_rel? If we were decomposing those strings, maybe, but not easily (if at all) for the periphrastic ones.
That solution also runs up against: _similar_a_to_rel: tells the generator (via trigger rule) which particle to pull out.
comp_rel introduces an ARG0. Is that comparison event something that has temporal-spatial existence? Can be referred to? It is existentially quantified...
Driven to it by trying to make adj & adv parallel, but it has some precedence in the semantics literature. Ann can provide references that back it up. In these cases, the e refers to a property.
Should we come up with a third type that is a property? verbs = event, adj = underspecified, adverb = property. (Sometimes adv are i in current grammar, to try to get at this, but underspecified between x and e isn't really right.)
e := i x := i y := i ;;; properties
Except we might want more intermediate types (placeholder names for now):
e-or-x := i. e-or-y := i. x := e-or-x. y := e-or-y. e := e-or-x & e-or-y.
attributive adjectives underdetermined: can be events in German the brown in Spring lawn (Der im Fruehling gruene Rasen ist jetzt braun und ausgetrocknet.)
That problem is harder than anybody expected.
-> the reference set is sometimes referred to with clause. So the ARG2 of comp_rel is sometimes x and sometimes h, like the ARG1 of _bother_n_rel. Thus in the LF, we'd need to resolve comp_rel to one of two predicates on the basis of the ARG2's type.
If comparison is between two sets along some dimension, why is comp_rel only two-place? The third one is there --- take the adjective rel & comp rel together as a unit, and all three elements are there.
Are we worried that u10 is unbound? That's fine. Is it INI or DNI? Not represented, but this issue does connect to anaphora resolution and whether these can be understood in context.
Does the binary approach to coordination also relate to anaphora? (Aren't we asserting that the smaller group is something that can be referred back to? Maybe with particular intonations that's supported? What about testing gender on plural pronouns?)
John, Mary and Kate went to the park. The girls didn't bring their coats.
John went to the park. Mary went to the park. Kate went to the park. The girls didn't bring their coats.
Definites can introduce new groups more easily than pronouns, so if we could construct an example where the gender on the pronoun required Mary+Kate only (not possible in English), might find acceptability differences.
_and_c_rel: introduces an i ... should it be an x? Probably. Shouldn't the quantifier do it? (And keep using _and_c_rel for both N and V.)
Is electronic in distribute electronic and computer building products N or A?
Don't get A in the middle of N N compounds except if the A-N combination is lexicalized: *camera blue parts, company annual report, so that's possibly a test, though it's not feasible to enter all the lexicalized ones into the lexicon.
... discussion of whether we could block this in generation. Would require having an underspecified representation (between A-N and N-N) so that the generator could deal with the input.
There are lots of cases in English where it's basically impossible to tell whether something is an A or an N.
Was there ever an underspecified A/N gen lex entry? Or some attempt at cleverness regarding using the mass-count gen entry to handle the unknown word's A tag. </ancient history>
Los Angeles listed as a single element (MWE), but plan is to remove them in favor of composing proper names.
Also taking out of the lexicon a lot of the single-word proper names. Except for those in old treebanks parsed without POS-tagging unknown words.
John and Mary Brown bike/*bikes to the store.
Right-headed analysis of compound names is wrong.
Abrams Browne arrived
Two quantifiers, one for each proper name, compound_name_rel joining the two.
When changing the syntactic structure, should the mapping between the ARGs of compound_name_rel and the two names be the same, or opposite?
Can we get down to one entity (one quantifier) for Kim Smith?
Then should Kim and Sandy Brown be Kim Brown and Sandy Brown as the entities? (Out of reach for composition.) But maybe Kim and Sandy Brown being the entities is an okay approximation (not quite right, but compatible.)
How many entities in New York Stock Exchange?
What about University of Chicago and Cambridge University?
That's a productive way to name universities. But how do we know which ones are productive and which ones aren't? "university" is in the lexicon. But so is "gates". (cf. Bill Gates).
ICONS relationship between the two to say they denote the same thing. Doesn't have to be done by the grammar (University of Washington ... the two entities don't refer to the same thing). Some external component could make that call and give the ICONS constraint.
Previous meeting: could we drop quantifiers on pronouns or proper names? Maybe pronouns, but definitely not proper names.
What are we doing with number names? The numbers are adjectives and so have i-type ARG0s, not x-type.
New York Stock Exchange -- grammar provides the analysis where all four are proper names, but the one that Dan treebanks is New York as a proper name and Stock Exchange is an NN compound.
John Robert Charles Smith --- grammar gives ambiguity in the branching structure.
- [Emily [Menon Bender]]
- [[Mary Ann] Smith]
Is the relationship between the parts of a last name different than that between first and last name and between first and middle or two part first names? Can we see what we're doing as an underspecification of that?
Pacific South West Lumber -- can't tell what the company thought the bracketing should have been.
Are there some cases where we can say that one bracketing is preferred to another?
Can we say we're building up a name and make the actual entity turn up at the end? Can we say that the entities for the name parts are referring to the names, and then put in a unary rule at the top to put in the person entity?
Can't do one proper_q and two named_rels with the same ARG0. Breaks DMRS. Would rather that those are anaphora.
name(x1, "Kim") named(x2, x1)
The most popular names this year were Kim and Sandy. <- could model that, but would lead to a terrible proliferation of ambiguity.
name(x1, "Kim") named(x2,x1) proper_q(x2) name_q(x1)
Kim of Avonlea left.
moniker(e-or-y1, x2, "Kim") named(x2, e-or-y1)
Just one entity per complex name. Ambiguity for Kim of Avonlea or University of Cambridge still but not for John Smith Brown, at least, all of the different structures would still have only the quantifier at the top.
moniker is adjective-like; in fact very similar to the card_rel and then we don't need to separate rels.
named(i, x2, "Kim")
... and then only the head gets the I specialized to x.
- Kim Smith
- Smith Kim
semantically different? Yes. Which one gets the x is different.
- Kim Smith Jones
- Kim Jones Smith
will come out looking the same, modulo character positions if the names are just modifying each other. If there is a compound_name_rel, then the order of attachment is recorded.
Do we want to have access to the order of words in the semantics?
We are using it so far for: John slipped and fell (temporal order) and fish and chips != chips and fish.
What about the top-most named-rel --- there is nothing for its ARG1 to point to.
proper_q(x1, h13, h16) named_n(x1,"Kim) named_a(w3, x1, "Smith") sleep(e7, x1)
Names are underspecified as named(ARG0:, CARG:) because when they join the construction always knows which is head and which is modifier, can specialize to named_n and named_a.
named_n := named-rel named_a := named-rel
Abstract type not a string, so can even be done now.