[For link to slides, see SingaporeSchedule]
Representation of DELPH-IN Pragmatics
- Discourse can improve language processing
- info str, honorification, questions, rhetoric qs, anaphora resolution, ...
Kim has two children that study linguistics (Kim can have >2 children)
Kim has two chilren, who study linguistics (Kim only has 2 children)
- Assertion/Denial (e.g. Chinese ma-questions vs A-not-A); rhetorical questions are different
- Speaker/hearer and inner/outer frame (w.r.t. quotes and things)
see JacyPragmatics for implementations in Jacy
- some unresolved issues that, if resolved, could aid in things like MT and IE
- two questions regarding implementation: what is the purpose of "ACTIVATED" feature? and the PRESUP feature?
Emily: this is code untouched in a long time (maybe needs to be revised?); I'm pretty sure I just copied it over from the ERG in 2001 when creating the initial version of matrix.tdl.
Dan: Used ACTIVATED to block It, I like.
Emily: The "ACTIVATED" name suggests it's to do with information structure, but it's in CTXT which means the connection to the index involved is lost.
Dan: I think these were added by Malouf from Pollard & Sag 1994, but hasn't been touched since
Sanghoun: PRESUP is not used in any current grammars, right? (Dan and Emily: true) I added "SPEAKER" and "HEARER" keys in C-INDICES, and 3 new HOOK features (ICONS-KEY, SPEAKER-KEY, HEARER-KEY). (Sanghoun explains implementation of ICONS for showing "addressor" and "addressee" in quotes)
- "'You have been cruelly used,' said Holmes."
Emily: This looks like progress, but maybe we should use the INDEX instead of the TOP handle?
Sanghoun: yes, that's one of the problems
(SS gives example of Japanese honorifics "ojiisan ga oyasumi ni narimasu": this example yields 4 ICONS: addressor, addressee, higher x 2)
Guy: asks for clarification about the honorification
SS: there's 2 honorifications here, toward the old man (ojiisan, oyasumi ni naru) and toward the hearer (narimasu)
EMB: I'm not too worried about not have the speaker in the RELS, but rather the IARG2 pointing to the INDEX vs LTOP
Dan: [defends the current state]
Dan: concerning MSG, we had trouble deciding how many to put in; I'm concerned ICONS may be on a similar path
EMB: there's a limit: one for utterance, and one for verbs that embed the utterance (direct quotation), there's a small number of verbs that have these constraints
Dan: where do I find this list of verbs?
EMB: This is in the syntax literature, I think, in discussions of root phenomena. Also could look in FrameNet for a list of "saying" verbs?
Dan is skeptical of an exhaustive list or reliable diagnostic to find such verbs, Emily thinks that such verbs that don't involve quotation marks are of a small number, Dan points out that no one uses punctuation consistently.
Emily: Still, there's a distinction between direct and indirect quotation.
Dan: is this a cross-linguistic claim?
Emily: yes, I think this is like negation in that all languages do it in some way
EMB: I don't think this is the slippery slope of MSGs
Guy: Whats the slope?
Dan: the constructs that yielded MSGs would explode, so you'd have an MRS with many MSGs floating around, and nobody liked them. Instead we have a SF (sentential force) feature on events.
Petter: in the Norwegian grammar we have ~114 of these verbs.
SS: for 2nd person pronouns (honorific vs non-), I use the "higher" relation, where the HEARER is coindexed with the higher
Melanie: I object to the term "higher" because in, say, Japanese, it's more about social distance than hierarchical relations; it's more formal. E.g. with my students I'd use "-masu" form, even though they are not higher.
EMB: it's also stylistic, rather than hard constraint, logical assignment of social standing
Guy: can't we add a "higher-or-lower" as opposed to the "int" (intimate), in order to get, say, the German "sie"? (SS: I'm doing that for Korean currently)
SS: I wonder:
- if the representation I've given is plausible for MT?
- if ACTIVATED and PRESUP are still necessary for our grammars?
- if (something... didn't catch)
- how to represent pragmatic information in grammars?
EMB: I don't know, but I hope, we can represent things solely in the MRS, but maybe we also need the syntax?
Dan: I don't think we do enough to contrast (1a) and (1b), the ERG fails to do this. I want to represent the intersective semantics, but also include new information (e.g. the presupposition).
EMB: Can we keep a separate list for these?
Dan: I'd rather not keep other stuff outside of the RELS list...
Woodley: Remember the example about camper vans? There was a proposal about how to do this. (see: AbbeyApposition)
Dan: but it didn't involve presupposition.
EMB: do you think having a separate diff-list is better than stamping some EPs as "presupposed" rather than "asserted"? The tricky thing is (something about) different kinds of presupposition. How much inference do we need to do to get this info?
Dan: It's not that you can't say (1b) where Kim has more than 2 children...
Hans: if you want this for MT, don't you need to keep the parallelism? [ scribe couldn't follow ].. if one does something here, you need to make sure it's useful for other things...
Dan: This might help for apposition, like "My brother, the doctor, makes a lot of money". Right now we don't do this well; it hardly scopes. If this works out, it makes a case the presupposition may work for other langs.
David Inman: Can't you have a RELS value "presuppose_rel" with ARG1 and ARG2?
Dan: you may be misreading the presupposition as part of the syntax; it's not a sign or anything like that; it's context.
FCB: what is it we're presupposing? That Kim exists. We're asserting she has 2 children, and that they study linguistics.
Dan: if I say "My brother, the doctor, is rich" and you contend "he's not a doctor", that's a strange thing to contend...
EMB: I don't know if these are the most relevant example... maybe "When did you stop smoking", if you never smoked, there's a presupp failure.
MWG: how about "Kim's two children study linguistics"? (others: yes, that'd be presupp)
EMB: if I say "both of Kim's children..." and you say "no, Kim has 3 children", there's a presupp failure; the goal is to capture this presupp. The goal is to calculate all of these presuppositions by looking at the semantics.
Guy: for relative clauses, could you have a diff in the restriction of "two", change it's scoping. So "children who study linguistics" is in the restriction of the "two" quantifier.
Dan: then you say it's like anaphora, as if you'd said it in two different sentences
Woodley: I think it could be in BODY instead of RSTR
SS: (changing subject) what about different usage of questions? Chinese -ma vs A-not-A, rhetorical..
EMB: this connects to conventional implicatures vs ... something..
Dan: I'm pretty sure we don't want to include things in the semantics.. e.g. if I say "it's cold in here" to mean "turn off the A/C", that kind of implicature is not something we should directly model.
EMB directs back to -ma vs A-not-A
Wenjie: it has to do with speaker's bias. if I ask "you like dogs-ma", I assume you like dogs. With A-not-A, I don't have the bias (simple asking).
Zhenzhen: it also has to do with the focus. If you say "xihuan bu xihuan" you are focusing on the verbs... "do you LIKE dogs?", or if you say "xihuan gou haishi xihuan mao", you're focusing on the object of the liking.
Dan: this is like tag questions, "you like dogs, don't you?". So if I ask "you like dogs ma?" can you respond "As you know, I do" (others: no)
Zhenzhen: more like a tag question is -ba ("ni xihuan gou ba?")
EMB: I think the A-not-A is harder; I think the difference is a change in the focus rather than a strong presupposition. So it's not your problem. (SS: *phew*)
EMB: what about the ne_rel (in German ne, in Japanese ne; (Dan: in English, "you'll do that, no?")? ... that was a message rel. What is it now?
Dan: Just an ordinary EP.
Zhenzhen: (slide 7) I think similar info is represented in CONT and in ICONS... is it necessary to be so redundant?
EMB: I don't think SPEAKER and HEARER is doing much.
SS: So I think to remove CTXT (Dan: just ignore it, like other grammars)
EMB: I challenge you to do it without the HEARER-KEY and SPEAKER-KEY. You'll have the relevant info in the ICONS.
EMB: [Added afterwards, I think this comment was about here in the sequence:] The CTXT feature reflects an older conception of how pragmatic information is gathered during linguistic processing, namely, that it is done on the backbone of the syntax tree as the semantics is. I wonder if it would be possible instead to use the MRS (i.e. the compositionally constructed semantics) as the basis for computing the semantics (together with lexical information, i.e. about predicates). Or is there some info required from the syntax that is not reflected in the MRS.
SS: If I drop CTXT, I don't need HEARER-KEY and SPEAKER-KEY.
EMB: there's also the question of 2nd person pronouns... 1st person is tied to the addressor, 2nd to the addressee... but we shouldn't be trying to do that in the syntax; the info you'd need would be there with the variable properties on the pronoun's indices plus the info about which clause they belong to to link them up to the right addressor/addressee in post-processing. (NB: also not identity in the case of pronouns.)
Hans: it seems dangerous to include presupposition, because the outside world needs to learn how to deal with it. It seems like an appeal to involve other people in coming up with a uniform solution, rather than implementing your own solution that others have to deal with.