oe: [erases objection to analysis of Grumpy as grumpy(e,x) thing(x)]
Dan: The Grumpy I met yesterday: need Nbar -> A rule, which then interacts with the proper NP->Nbar rule.
oe: Want to look at interactions with other phenomena that might force different interpretations.
Woodley: He is Grumpy will have two structures.
Dan: But in the new proposal He is President will have just one structure (as opposed to two now).
Dan: Adding the entity_rel when converting the adjective to a noun?
Emily: Seems like the time to do it.
Francis: Ann’s proposal [over email] was a direct reply to my attempt to take us down the path that since we’re abandoning the common/proper distinction for nouns could be adapted for idioms.
Francis: The concept of proper name is that we want to give New York the referent of the whole of New York at some point. We want to also keep the internal structure.
Emily: I thought we were assimilating proper names to idioms and now you’re saying we have to assimilate idioms to proper names. Why the second step?
Francis: We’re expanding the domain of idioms to include named entities.
[Textual interpretation … maybe should wait for Ann.]
[Switching to generation issues.]
Dan: What is the expected input we might imagine for someone who wants to generate Modern Gardening. Probably come in with a named entity.
Woodley: Where are they coming from?
Dan: Outside…
Woodley: If it’s MT, might come in with the Japanese translation of a title.
Dan: I was assuming it might be a named entity that’s Modern_Gardening. If we have that long list with the surface/canonical string (in the gazetteer) can look up the semantics we want to put into the input MRS.
oe: Making the assumption that the gazetteer is a list of surface string to MRS fragment pairs (some with big long _n_rel press), though there may be some noise (parse selection)
Francis: Two possible inputs, one from another gazetteer in another language, so we might have a direct mapping, but we might also have something external that just gives us Modern_Gardening +NE.
Woodley: If the generator input has told you what string it wants, then it seems a step in the wrong direction to take that back to the semantic structure.
oe: What about those cases where the things are coordinated.
Woodley: If the input includes coordinations, then they didn’t give us the string they parse.
Woodley: I’d like the generator and the parser to handle things that aren’t in the gazetteer, and for the generator, that can be done, even if it’s harder for the parser. If on a Winter’s Night a Traveller… especially if that’s not all capitalized, can’t necessarily handle it. But if the generator already has the information that that string is wanted, why balk?
oe: Want to preserve the ability to generate unknown names where we put the CARG into the output. I think we always want that option.
Francis: because it might be not in the gazetteer.
Woodley: If there are cases where going to MRS representation allows us to generate more outputs, I’m in favor of that. Examples?
Francis: Secretary of State and Secretary of War/Secretaries of State and War.
Dan/oe/Woodley: With a suitable paraphrase rule.
Dan: Also trying to think of a case where we have something sg in the input and we need plural.
Woodley: Or in languages with case marking, can certainly imagine that.
Dan: Want to avoid saying The Andes is tall. Need at least the plural marking info from the gazetteer. Not really sure that plural marking is a property of the named entity in the semantics. (Though this might be a systematic fact of mountain range.) Want to know if it’s true in the same language over.
Emily: Why care about another language, if that’s a feature of English?
Dan: MT. How do I put it in?
Emily: The gazetteer. It’ll have that info in the MRSs, if it’s in the grammar.
Dan: Expect Ann to wrinkle her nose at the claim that The Andes is plural.
Woodley: How to do that in parsing (block The Andes is small).
Dan: In the lexicon I can record both syntactic and semantic information in the lexicon. We’re talking about having an idiom storehouse of these entities, and so I won’t have them in the lexicon.
Woodley: I don’t see how you can get away with not having Andes or something like it in the lexicon.
Dan: Or at least have the info that this is a mountain range, in case they’re not in the lexicon.
Woodley: That is at least a fact about English, which has to be encoded somewhere.
Dan: Even the requirement for the is an idiosyncrasy about syntax.
Woodley: CARG _the_Alps in generator input, how do we make sure that you pull in the constraints form the grammar?
Dan: Want to avoid the French the Alps, while allowing the lowland Rockies, the foothill Rockies
Emily: The gazetteer.
Dan: The gazetteer will say that this is just Alps, and the lexical entry from the grammar will have the constraints needed (as an instance of a mountain range).
Francis: cf my dissertation
Dan: So needed a sharper example.
Emily: Woodley’s point about case marking is very apropos, you can’t just take the string from the generator input and use it directly if it needs to inflect based on grammatical function.
[Die die the the]
Woodley: But surely Secretary of State changes with different case marking in German?
oe: Yes, but it’s one long morphological compound.
Dan: And The current Secretary and former Treasury of State?
oe: [works out how to say that in German]
Woodley: In English, that sounds like one person to me.
Dan: Can always make it one person, but … had an affair?
Woodley: With who?
Dan: … had an affair with each other.
Dan: And now let’s get back on topic. When we get an input which is string based, it shouldn’t take long to remember that strings don’t always correspond to underlying representations, so we should remember to do processing of that English string to figure out what the idiom is that corresponds to that string. So we do parsing before generate?
Emily: No, we just look it up in the preparsed gazetteer?
Dan: But they might give us a different case form or something else weird.
oe: I think Dan’s assuming that’s the string in what’s currently the CARG.
oe/Woodley: We don’t want to allow them to give long coordinated strings.
Dan: What about plural/singular?
Woodley: One canonical input for each entity.
Dan: The Clintons arrived. ??
oe: Clinton_n(x) NUM pl.
Dan: That’s not in our lexicon. Can’t have both sg and pl versions for each.
Emily: Leave it underspecified.
Dan: But we just said that Alps has to be plural.
Emily: But Alps are special.
Dan: How do I create the four million entry lexicon? How do I know that Alps has to be plural, but Clinton can be either?
Emily: The linguist says that Alps are special, the other ones (ordinary) start of as ordinary.
Francis: Will have to be some work by hand in the early stages.
oe: There are names that have surprising properties, and these have to be known in the grammar.
Emily: But the claim is that these are the minority of cases.
Francis: _Alps_n_rel/_Clinton_n_rel
Woodley: But Dan says he doesn’t have a theory about whether Alps is semantically plural.
Emily: But we know that it’s syntactically plural, since this is an idiosyncrasy that has been discovered and can be written down in the grammar so it does the right thing.
Dan: Might be able to handle this via a generic entry for “name of mountain range”, don’t need to record that info in the idiom lexicon.
Emily: … because that’s a syntactic fact.
Woodley: Doesn’t need to interact with the gazetteer/idiom dictionary.
Dan: New example: star configurations like Cassiopeia, the Pleadies. Where is it recorded that one is plural and one is singular.
Woodley: Like the Alps.
Dan: Not the same, because with the Alps it’s predictable, based on what’s in the world. But probably not true for astronomical features.
Emily: It’s going to come down to someone writing it down, either gazetteer side or grammar side.
Francis: Not yet clear to me which side it will be, and I’m interested in talking about this more.
Woodley: Is this a special problem for named entities?
Francis: If we can put this kind of info into the grammar rather than the gazetteer, it gives us a wider range of gazetteer it gives us a wider range of gazetteers we can use.
Woodley: It’s not information that’s likely to be already available annotated out there somewhere, and will have to be manually done at a lexical level.
Francis: Or in lexicons we can’t get access to. It’s in Alt-JE for 240,000 named entities.
Dan: The piece that’s still a little blurry to me because I was thinking of doing some curation of that lexicon and you all sneered at me because it would take a lifetime. So what’s the back off? Emily suggests punt, live with error.
Francis: I was thinking of the curation of the internal structure, and it seemed to me there was rather a lot of that. But the number distinction, I’m expecting a small minority of named entities to have number distinctions, so that seems more manageable and doable.
Woodley: It’s also this case something you could do with a relatively straightforward test over unannotated corpora.
Emily: Isn’t this really a precision problem. Can we say give us number if you don’t want The Timberlakes make an appearance for the input for Timberlake makes an appearance?
Dan: I was looking for cases that must be plural (syntactically), so The Alp never comes out.
Emily: The claim is that these are few and far between and can be noted as they come up.
Francis: Wikipedia has this info for several of these.
[Emily leaves room momentarily]
Francis: What is the desired output?
Dan: If we think there is an instance of an entity of the same logical real-world type out there, then yes I think we should be consistent about what the number is, or we should have no opinion.
Emily: Why not just over generate?
Francis: Alps
Dan: What if someone tells me that they think the Alps is singular.
Emily: Constrain it in the grammar as a syntactic feature only. That’s a fact about English.
Woodley: That will fail the generation subsumption test … but that’s a separate issue.
Dan: Kim ran is going to be marked singular, not unmarked. But Andes has an entry in the lexicon so it gets special handling where I know not to strip off the s and not to put another one on.
oe: Don’t need to decide today about the semantic number of mountain ranges.
Dan/Francis: But in the gazetteer it will be unmarked for number.
Dan: Okay, that’s clear; doesn’t lead to any new burden to the lexicon that we had before we added the gazetteer.
Francis: But now you may want to look at a larger number of these.
Dan: Will want to be careful to strip number off of every entity that we put in — as we are producing the MRSs for the MRS side of the gazetteer. Prune them to make them suitable for generation.
Emily: So if someone gives us Alps-singular, then can’t the gazetteer head that off?
Dan: I expect that the gazetteer will protect me from over specification in the input.
Emily: But you don’t want that for “the Clintons” plural.
Dan: …
Emily: Two aspects to this: Grammar says syntactically plural, or grammar says underspecified. Input gives a specific number, or input is underspecified. Which is the problem?
Dan: If the grammar says that the thing is semantically singular, and the input says plural, then we don’t generate.
Woodley: Then it’s not a valid generator input.
Francis: So we’re not as precise as we could be…
oe: There is a ground truth about what the number for these is semantically, but we don’t know it now.
Francis: Can document that we want to things to be underspecified, or we can strip info from the input.
Emily: Don’t want to strip because of the Clintons.
Emily/oe: If they can’t be bothered to tell you the number then giving them both is valid.
Dan: So you want to treat proper names and common nouns as the same?
Emily: Yes, realization ranking.
Dan: But that’s lots more work — extra analyses for each proper noun.
Francis: Who is this source of input.
Dan: David B-P is going to give a set of constants.
Woodley: There’s some mechanism that takes his LFs and turns them into MRS that already does some accommodations. Can put in the singular then.
Dan: Okay, I have a babysitter in that case, but not for Chinese->English translation. Chinese doesn’t mark number.
Woodley: So have the transfer put in plural if there’s a mark (men, equiv of tachi –FCB), and stamp in singular otherwise.
oe: Francis and I put a rule in for that for proper names in Jaen (with a comment that it break on the Andes.
Francis: Want to try this as an empirical problem in Jaen and see what it looks like both ways. Also: I care about computer expense, not Dan space.
Dan: I wouldn’t care either if we didn’t have sentences with lots of proper names.
Francis: Don’t we pack in generation too?
Emily: Is it really a problem in parsing?
Dan: Okay, let’s talk about purity, not efficiency.
Dan: So what does our idiom lexicon have a number? Nothing at all?
Emily: We’re taking the position that the idiom lexicon knows nothing about number, but the grammar’s lexicon can constrain syntactic number.
Francis: The Japan and Northern Japan Alps. I have climbed five of the Alps. I have climbed five Alps.
Dan: The French and Swiss Alps.
oe: So it’s a common noun?
Dan: No the name of two mountain ranges.
Woodley: I know three Clintons.
Dan: That’s not hard. The alps is more vexed because it’s already syntactically plural. I don’t think it’s a problem, because we can treat Alps as semantically underspecified, and it can go through the plural rule.
[ Ann joins ]
[ Woodley summarizes discussion so far ]
Francis: And one of the unfinished things is cases where you want to represent something as both the internal structure and a single named entity at the same time.
Emily: And that’s where we thought your email was referring to Francis’s proposal.
Ann: I have been thinking about that for quite some time.
Dan: In the same soup with the semi-transparency: New York Stock Exchange. It’s pretty clear that I want to know that New York is part of that. Sometimes even contrasted with Chicago: New York and Chicago Stock Exchanges. But it’s annoying to have to reassert that that’s [New York] [Stock Exchange] and not [New] [York Stock Exchange]. Would be nice to be able to look it up.
Ann: Yes.
Dan: Generation direction. We were trying to decide about whether that idiom-like lexicon should contain information about further properties, like number information?
Ann: What’s in that lexicon thing. There’s an MRS structure, and what else?
Dan: That resource is going to be a pairing of an MRS structure and something like a sequence of tokens, probably stemmed. Modern Gardening entity. That might or might not decompose in the MRS we’re going to store for it. In this case, probably three ep semantics. But we also want the surface sequence, so I can say which thing is typically capitalized (e.g., not the of in Secretary of State). So I think we need something like the surface tokens including their case. Going to have to have a pair of something surface plus the MRS.
Ann: I’m worrying about treating all these MRSs as compositional, in some sense.
Dan: We have the fall-back strategy for those that don’t sustain a compositional analysis. Extreme default is calling it essentially a word with spaces that is a named entity/NP.
Ann: The case which is problematic is the one where one of the words is something you want to be able to interpret, and the other is a fw or something, and they’re both part of a named entity.
Dan: The two end points seem reasonably okay. Maybe can do something more sophisticated in the middle.
Ann: Someone using this MRS will see Secretary of State as completely compositional?
Woodley: At least that, plus also the info that we recognized it as a named entity. We haven’t discussed how that info is added to the semantics (as another layer or what).
Ann: Any user of the semantics will want the named entity information — not Home Secretary as compositional. I think that also relates to what you’re thinking about for the generation perspective too.
Woodley: Likely that in the generation direction, would only get the opaque form.
Emily: So we can look up the MRS we want from the gazetteer.
oe: Or generate as unknown words.
W/Dan: But then what do you do with Secretaries of State; how do we know where to put the plural?
[ Side conversation on how to talk about words with spaces in generator input, but at any rate, it’s a single ep.]
Dan: We have to go through the unpacking in the named entity lexicon so that we can generate the singular or plural form of that named entity.
Ann: We have a mechanism for putting the morphology in surprising places, could use that.
Emily/Woodley: But unpacking would let the grammar do that for us; that’s exactly what the grammar is good about.
Ann: Sounds like a mechanism for creating named entities semi-automatically without the grammar writer having to do too much. Sounds reasonable.
Ann: Then in some sense internally what you want the semantics to be is the compositional semantics, but externally you want a single entity so some user doesn’t have to decompose (parsing)/compose (generation).
Ann: The sort of general issue relates to conversations decades ago about how to degenerate from a mixture of word sequences and proper semantics. In certain circumstances you just parse the word sequences and put that into the generator.
Emily: Publications?
Ann: No, it was a major factor in thinking about MRS in the way we did. In principle the chart generator will allow you to do a mixure of bits of words and bits of semantics, but in practice we never got past small-scale experiments (sometime in the mid-90s).
Dan: Bumped into it again in discussions of names of books and silly hotel names in Verbmobil, where the phrase itself because a named entity. I think that On Nomnialization is a better paper than [name of some other paper. Want a uniform treatment of that that says that despite the internal structure their function in the larger sentence is just an entity.
Ann: Were also thinking about this in terms of AAC applications where you take a chunk of text as input and need to make a valid, proper sentence. Seems plausible that what you want to do is the same in hybrid MT — got some bits of text that are good, know they’re good, and want to stitch them together. If we did do this in a very general way, could be useful more broadly than named entities.
Dan: One we have a mechanism that allows to take whatever source a chunk of validated MRS and we have instructions on how to realize it in the syntax… If you give me something that looks like it’s a PP but is marked as a named entity, then I know how to proceed. Might need coercions like knocking out number information.
[Summarizing previous discussion about number and astronomical entities.]
Dan/Ann: Dialect variation: IBM are releasing a new product.
Ann: That’s got to be a general coercion in BrE, not just that IBM takes plural agreement.
Ann: Trying to think of examples that are sometimes sg and sometimes pl according dialect, or maybe how well you know the domain.
Dan: If we’re dealing with a generator input where someone has tried to give us number information. Would be nice to generate regardless of what they did.
Ann: You’re thinking of them putting in a mixture of an MRS and a string.
Francis: _Modern+Gardening_n_rel is I think what we should expect from them.
Dan: Especially in MT input when this comes from a language when they just keep the English name. There is an entity of the MRS there, but it’s just a single ep.
Ann: I don’t think there’s any way they’re going to be putting in sg or pl, and they could only be putting in sg or pl wrt to the denotation of the thing, not actual number agreement.
Dan: We were thinking along those lines too, with the distinction between syntactic and semantic number, where we already allow mismatches. Were thinking about having these be underspecified for number, but if someone hands me _Clinton_n_rel: num plural, they and we both know that it’s plural.
Ann: The Clintons isn’t a named entity.
Dan: Clintons always run for office after a Bush has been in power.
oe: Clintons always run after Bushes
Emily: We’re moving to a world where there’s less of a distinction between common and proper nouns.
Ann: I’m thinking about cases where someone gives you a complex string they’ve asserted is a named entity, and then you parse that and then use the MRS, as opposed to saying this string is a named entity and here’s how it combines with the rest of the string.
Emily: That was the point of the German examples, and how we got to the number question in English. [See notes above.]
Ann: So you’re going to take in hints about the form of the string, but not fully believe those hints.
Emily: Right use the hints to look things up in the gazetteer (or maybe on the fly) and the build up the string according to how the grammar says it should be.
Ann: Sort of worried about finding rules to integrate the thing in to the MRS, but I can see the point.
Dan: Can you think of another approach to Clintons always like to run for office.?
Woodley: Going off of Ann’s email, don’t have to get rid of named_rel, just understand that named_rel Kim is always optional.
Ann: I like the idea of someone being able to give you a string packaged up in the CARG or something, rather than having to take it through the rel.
Dan: But then that’s directly in competition with my hope of not having to decide whether Secretary and State are opaque names. I’d like to say, that’s just _secretary_n_rel and _state_n_rel, and I know one other thing: that they both occurred in mixed case.
Woodley: Not convinced it’s in competition.
Ann: I’m still thinking of some set up in which you have both. Also thinking of the more general MWE problem. I don’t mind so much about the exact representation of NEs (aside from general concerns about MRS usability issues). Still sort of going on this idea that you have dual annotations. If when you’re treebanking, and you come across something that’s not in the gazetteer you’d want to add it, right?
Woodley: We were assuming an exhaustive gazetteer form external sources. (pretending)
Ann: Thinking more in principle than practical: this is lexical information; in some sense the gazetteer is a way to expand our information.
Dan: Would be nice to apply that info after the analysis has been chosen so that I don’t have to contemplate that division of labor…
Ann: The advantage of doing it before is that you want that thing to be a constituent.
Emily; A constituent, or a connected subgraph in the MRS?
Ann: Normally you don’t want internal modification of a named entity. At least you want to disprefer certain kinds of modification.
Emily: The former Secretary of State?
Ann: former should take scope over Secretary of State
Dan: The new Secretary of State
Ann: Whatever entity the Secretary of State refers to, I want new to apply to that same entity. You don’t want it to be applying to just part of that named entity. The fact that it is a named entity gives you that information.
Dan: In some sense it seems like we’re imaging in that we’ve build up the structure Secretary of State, and say “aha, my idiom lexicon tells us we can coerce that into an entity”. In some cases like I saw Grumpy yesterday or Old Angry just came in it’s obvious because without the coercion it doesn’t parse. Maybe we need something a little more generalized that lets me do a similar decoration of the MRS for Secretary of State.
Ann: Suppose you had a word with spaces type semantics of Secreatry of State. Thinking of that as an analysis that’s available, can be chosen in tree banking, as opposed to something you do on the MRS as a secondary step.
Emily: Wasn’t that exactly the decision Dan was hoping to avoid making?
Dan: The Chicago Stock Exchange/The Chicago stock exchange/The Chiacgo Stock exchange… not always clear if they mean a named entity.
Ann: I get it, but I don’t know that stopping the possibility of you annotating something as a named entity is the way to go. At the moment the problem is that you are required to make a decision between named entity and other, maybe you should have the ability to make it if you can, but not be forced to.
Dan: So I could chose it if I want to, but then could leave it underspecified, by not using the coercion rule.
Ann: Slightly worried by the use of coercion, because I want to think of this in some sense independently of the implementation of the effect. But: There is an indeterminacy sometimes between whether something is an NE or not, but when you know, it’s useful to be able to record that fact. (Cuts out lots of parses, if the internal structure is complex.) Also want modification to apply to that thing as the named entity —might otherwise get incorrect analyses. Long term solution is in terms of supporting a dual analysis approach.
Emily: Does it matter where the dual analyses come from—both from grammar, vs. one post-treebanking from gazetteer?
Ann: Not thinking in terms of order of events. There are cases where only the NE analysis is available, no compositional analysis (e.g. Grumpy) so that has to come from the grammar. I guess that implies that if you did it as a two stage process, could think of the treebank leaving that as a compositional analysis where it worked…
Dan: You (pl) remind us that if we do it in a two-stage way, that’s going to require us to have an inventory of rules that allow us to get it through the parser for the case when it doesn’t look like an NP (e.g. On Gardening).
[ Emily gone briefly ]
[ Topic picked up again towards the end of the day ]
oe: Simpler examples to try to clarify where we are:
- Bill arrived: only the name
- The bill arrived: only common noun
- The Bill Arrived: could be either
- Old Navy opened a new store.
- Very Old Navy opened a new store. – star?
Dan: Not star. There’s an entity named Navy (if it’s capitalized), and I’m telling you it’s very old. If not capitalized, then won’t parse because it’s a singular count noun (leaving aside the color name). Taking the crucial property of the mixed case as something I get to see, and it is involved in the licensing of that N’ as an NP.
Francis: And if we want to parse annoying text like Twitter?
Dan: It’s too expensive.
Woodley: You might have to start a new project: The Twee RG.
Francis: But they’re short. Can we turn it on?
Dan: It’s an open-source grammar.
oe: I heard that very Old Navy opened a new store. Should be star?
Dan: very sunny Pittsburgh/very Sunny Pittsburgh: both impeccable, want to be able to parse that. What I wouldn’t get was an entity called Old Navy that was very something.
Emily/Woodley: Only Old Navy is in the gazetteer, but we build very Old Navy, and then only the Old Navy part of that subgraph gets projected out into the Schrödinger MRS.
Dan: The incoming arc might be a way to block the named entity from firing.
Emily: What about very sunny New York?
Dan: Difference is that you can modify the main index, but not the modifier.
Woodley: How does that work with coordination.
Emily: Same way.
Woodley: At least if it’s top-level coordination, but what about Secretaries of State and War?
Emily: We already have to do something special for those.
Last update: 2014-02-17 by EmilyBender [edit]