All Transitives are Derived
Lushootseed is a Central Salishan language spoken in the Puget Sound region of what is now Washington State, USA. One hypothesis about Salishan languages is that all verb roots are intransitive (Intransitivity Hypothesis; Davis and Matthewson, 2009), the corollary is that all transitive verbs are derived. This phenomenon brings up questions about how to perspicaciously model the semantic reflexion of such transitivization processes: should we provide structures which build up transitive predicates hand-in-hand with the morphological rules which derive them from intransitive bases? Or should we "swap out" one predicate for another in the application of transitivizing lexical rules? RMRS (Copestake n.d.,Copestake 2009) may provide a way to model the compositional structure of transitivizing lexical rules in a typed-feature structure based grammar. The purpose of this session is to present some data which motivate the use of RMRS and to discuss the application of RMRS principles in this context.
Quoting Beck (2009, 1--2):
- What are transitive verbs in most languages are derived from a large set of monovalent patient-oriented (Hess 1995) radicals whose syntactic subject expresses the semantic PATIENT or ENDPOINT of an event rather than the AGENT. Consider (1):
1. a. ʔuɬič̓ čəd. ʔu-ɬič̓ čəd PFV-be.cut 1SG.SUB `I got cut with a knife.' b. ʔuɬič̓id čəd tə sqʷiqʷali ʔu-ɬič̓i-t čəd tə sqʷiqʷali PFV-be.cut-ICS 1SG.SUB INDEF hay ‘I started to cut hay (with a blade)’
(Bates, Hess & Hilbert 1994: 146)
√ɬic̓ ‘be cut (with a knife)’—in spite of expressing an event high on the scale of semantic transitivity (Hopper & Thompson 1980)—can take only a single syntactic argument, a subject expressing the PATIENT of the event
- to express an AGENT, it is necessary to apply a valency-increasing suffix such as the internal causative -t
So, the Lushootseed lexicon has many such pairs of alternants where an underived verb has a single argument which seems to correspond most closely with what Dowty called proto-patient, and a derived transitive where both an argument in correspondence with proto-agent appears as well as one corresponding to proto-patient. This leads to questions about representations in MRS. One choice is to list the two verbs separately in a lexicon which is based on stems after derviational suffixes have applied:
ɬič̓=: intransitive-lex-item & [ STEM < "ɬič̓">, SYNSEM.LKEYS.KEYREL.PRED "_ɬič̓_v_be.cut_rel" ]. ɬič̓id =: transitive-lex-item & [ STEM < "ɬič̓id">, SYNSEM.LKEYS.KEYREL.PRED "_ɬič̓id_v_be.cut_rel" ].
intransitive-lex-item := basic-one-arg-no-hcons & basic-icons-lex-item & [ ARG-ST < [ LOCAL.CONT.HOOK [ INDEX ref-ind & #ind, ICONS-KEY.IARG1 #clause ] ] >, SYNSEM [ LKEYS.KEYREL.ARG1 #ind, LOCAL.CONT.HOOK.CLAUSE-KEY #clause ] ]. transitive-lex-item := basic-two-arg-no-hcons & basic-icons-lex-item & [ ARG-ST < [ LOCAL.CONT.HOOK [ INDEX ref-ind & #ind1, ICONS-KEY.IARG1 #clause ] ], [ LOCAL.CONT.HOOK [ INDEX ref-ind & #ind2, ICONS-KEY.IARG1 #clause ] ] >, SYNSEM [ LKEYS.KEYREL [ ARG1 #ind1, ARG2 #ind2 ], LOCAL.CONT.HOOK.CLAUSE-KEY #clause ] ].
But this misses the generalization that the event-type denoted by _ɬič̓_v_be.cut_rel and _ɬič̓id_v_be.cut_rel are linguistically the same, we have two predicate strings which refer to the same event type but with different argument structures, ie, we've packed a syntactic dependency into a semantic predicate name.
Another option would be to treat the transitivizer in a way similar to the ERG's treatment of periphrastic causatives. In this scenario, the transitivizing lex rules add a _cause_x_rel. Traditionally, this _cause_x_rel has its own characteristic variable, and then takes two further args, one corresponding to the CAUSER (perhaps something like a proto-agent), the other will be the characteristic (event) variable of the root. Under this sort of analysis, a verb like ɬič̓id would have semantics:
- [ _ɬič̓_v_be.cut_rel,
- ARG0 e0 ARG1 x0 ],
- ARG0 e1 ARG1 e0 ARG2 x1 ]
We want this to mean something like x1 cut x0 with a knife.
One thing that stands out to me about this analysis is the "extra" e variable on the cause rel. That is, I don't yet see the motivation for this. I think one of the characteristics of the AGENT prototype is acting as a causer of an event. For example, the English verb "fell" means to cut down, but someone might argue that this is a morphological causative variant of "fall". The ERG demo gives _fell_v_rel(e,x0,x1) for fell in "Sandy felled the tree". If we are happy to analyze English transitives as denoting a single event, I'm not convinced that the presence of a derivational affix in Lushootseed is motivation enough to warrant a proliferation of event variables in the Lushootseed grammar.
A third idea for representing these structures comes from using Parsons-style decomposition of the arity such that:
- RELS which introduce event variables are always monadic, they predicate a string-label for the event and nothing more
- RELS for AGENT,PATIENT,GOAL,etc are bivalent, their signatures being something like this: REL-NAME(e,x).
Under this sort of system, (1a) looks like this:
- [ _ɬič̓_v_rel,
- ARG0 e0 ],
- ARG0 e0, ARG1 x0 [ PNG 1sg ] ]
And (1b) looks almost exactly the same, but with the x0 being 3sg (some hay), and with one further REL added:
- [ _agent_rel
- ARG0 e0, ARG1 x1 [ PNG 1sg ] ]
In fact, our DELPH-IN formalism already has a precedent for treating argument relations as predicates a la Parsons, this is called R(obust)MRS. In RMRS, thematic role names aren't used such as _agent_rel, but instead the predicate names which are used are the familiar ARGN types, a la Dowty's ordered-argument/proto-role (Dowty 1989, 1990). Here I hasten to comment that in interpreting these ordered argument relations, the lexical class of the verb has to be taken into account, Lushootseed also has unergative monovalent roots (Hess' agent-oriented radicals) in addition to the unaccusative ones like ɬič̓. There are few more examples down below. EDIT: the more I think about this, the less sure I am of the previous comment---perhaps independent of verb class, ARG1,2,3 should be proto-agent, proto-patient, proto-goal/location, etc. In this case, there will be verb which can only select for an ARG2, for example.
Copestake motivates RMRS (Copestake n.d., 2009) as an MRS variant which is amenable to encoding the output of shallow processing (such as POS tagging) in a notion which is compatible with information from a deep grammar. The idea is that if you can write default rules which map POS tags onto semantic types, but one of the principle things you're missing in that scenario is the arity of the predicate for a given item. Parsons-style reification of argument relations as predicate names allows bare predicate names to be monads, predicated of an event variable and nothing more. RMRS then allows for syntactic dependencies to further predicate syntactic argument relations on that event, just in case you have a tool for finding them. Essentially, RMRS allows for underspecification of predicate-arity in this way. Although the motivation is different here, for Lushootseed, an underspecification of predicate arity allows that intransitive roots can their derived transitive variants share a predicate type.
Beyond the decomposition of predicate arity, yet connected to it, one thing which is different between the "standard" MRS and RMRS is that labels are required to be unique for each relation. Instead of using labels for conjunction at a particular node of the scope tree, RMRS adds explicit conjunction constraints to the representation. Each label, then, is used as the first argument to the ARGN_REL predicates. Keeping with the idea that argNs are to be interpreted with respect to verb class (so arg1 on unaccusatives refers to a proto-patient) and continuing the example above using ɬič̓id, it seems like we'd have something like the following for "ɬič̓id čəd tə sqʷiqʷali" (I cut the hay (with a knife).)
a0:_ɬič̓_v_rel(e1) a2:_arg1_rel(a0, x3 [PNG 3sg]) a4:_arg2_rel(a0, x5 [PNG 1sg])
One question which pops out to me here, is why should I use the label a0, as the thing passed into the first argument of the ARGN rels rather than the actual event variable 'e1' in the example directly above? It would seem that using the e variable would allow us to bring back in the explicit coordination using labels, as it's the fact that each label has to stand in for its predicate's argument as a reference point which forces the requirement that labels be unique.
Beyond the descriptions of the RASP system in the Copestake papers and references therein, I would like to know if there are typed-feature structure grammars which have implemented RMRS. I've heard talk of an ERG branch which was using RMRS. Is this just an apocryphal tale? Are there versions of the LKB or PET which know what to do with in-g constraints when calculating the scope-machinery? For me, seeing some of this stuff in use would be very interesting and useful.
Summary of Questions for this Session
- What collection of previous DELPH-IN work should I download and look at for previous RMRS work?
- What DELPH-IN tools are compatible with RMRS format?
- What subfeatures of particular DELPH-IN tools do or don't work when using RMRS in Typed Feature Structures
- Depending on the answers to the above, how do we actually write down RMRS in TFS using TDL?
- Is there any motivation to construct an MRS representation which doesn't enforce label uniqueness (perhaps by passing around eNs rather than labels as predicate arguments)?
- What's the *semantic* difference between an "e" and an "x"? Is a tacit assumption about the relationship of morphosyntax to semantics taken when saying that words of a given morphosyntactic class introduce xs vs introducing es.
- Given different classes of intransitive roots (unergatives whose single arg is something like a proto-agent, and unacccusatives whose single arg is something like proto-patient) does ARGN always map to the same proto-role type or is interpretation verb-stem dependent? If the former, ARGN seems like a shorthand for the proto-role-labels, if the latter ARGN starts to seem like a stand-in for the grammatical relation of SUBJ vs OBJ (or COMPS.FIRST).
Further morphosyntax of Lushootseed transitives
In addition to the "patient-oriented" roots discussed above, Lushootseed has agent-oriented roots. These can be used a base for deriving transitives with applicative morphology. Transitives derived from either type of root can only take a single full NP argument (termed the "direct argument" in Salishan studies), Lushootseed is notable for completely disallowing two direct NP arguments even for transitives. Thus:
- one argument or the other of a transitive is either null marked or marked via morphology (object-marking suffixes) or marked via a subject-clitic (the čəd seen in the examples above is such a clitic)
Some further facts of the system are that
- third person singular is null in the subject-clitic series and in the object-marking suffix series.
- if there is a full NP argument, and there is no object marking, the NP argument is interpreted as the more PATIENT-like of two args
- if there is a full NP argument, and there is object marking, the NP argument is interpreted as the more AGENT-like of the two args
- passivization reverses the interpretation
Thus there are no sentences such as (two full NPs, even in a transitive verb):
*ʔuʔux̌ʷ-c ti č̓ač̓as tsi č̓ač̓as (intended: the boy went after the girl)
But there are (subject marked with clitic):
ʔuʔux̌ʷ-c čəxʷ ti č̓ač̓as I went after the girl
And (object marked with morphology):
ʔuʔux̌ʷ-c-bš ti č̓ač̓as The girl went after me
Here are some further examples from Hess which show something of how causatives and applicatives are built from a couple of underlying root classes.
From Hess (1973, 89--90)
ʔuʔux̌ʷ ti č̓ač̓as The boy went ʔugʷədil ti č̓ač̓as The boy sat. -- ʔuǰiq̓ ti č̓ač̓as The boy drowned. ʔubəč ti č̓ač̓as The boy fell. ʔuč̓axʷ ti č̓ač̓as The boy got hit (by a branch while going through thick brush). -- ʔuʔux̌ʷ-c He went after someone. ʔugʷədil-s He sat next to someone -- ʔuǰiq̓i-t He immersed something. ʔubəča-t He set something down -- ʔuč̓axʷa-t ti č̓ač̓as He clubbed the boy. ʔuʔux̌ʷ-c ti č̓ač̓as He went after the boy. ʔuʔux̌ʷ-c-bš The boy went after me.
A further note is that there are quite a few more types of causatives and applicatives which aren't discussed here. The language has a robust system in its lexicon for building out types of transitives in what seems to be an obviously grammaticised way. The fact that this system stands out so readily when looking at Lushootseed verb structure is one of the principle motivations for finding a semantic representation which can capture the combinatorics exhibited here.
- Bates, Dawn and Hess, Thom and Hilbert, Vi (1994). Lushootseed Dictionary
Beck, David. (2009). A taxonomy and typology of Lushootseed valency-increasing suffixes. International Journal of American Linguistics 75, 533–569.
Beck, David. (2000). Semantic agents, syntactic subjects, and discourse topics: How to locate Lushootseed sentences in space and time. Studies in Language 24:2, 277–317.
Beck, David. (1996). Transitivity and causation in Lushootseed morphology. Canadian Journal of Linguistics 41, 109–140.
Copestake, Ann (n.d.) Draft on website:([http://www.cl.cam.ac.uk/~aac10/papers/rmrsdraft.pdf], accessed 19 Oct 2015.
Copestake, Ann (2009). Invited Talk: Slacker Semantics: Why Superficiality, Dependency and Avoidance of Commitment can be the Right Way to Go. In: Proceedings of the 12th Conference of the European Chapter of the ACL (EACL 2009), pages 1-9. Athens, Greece.
Davidson, Donald. (1967). "The Logical Form of Action Sentences". In Nicholas Rescher (ed.), The Logic of Decision and Action. University of Pittsburgh Press, pp. 81--95.
- Davis, Henry and Matthewson, Lisa. (2009). TITLE OF ARTICLE HERE. 'Language and Linguistics Compass' REST OF CITATION HERE
Dowty, David (1989). "On the Semantic Content of the Notion of 'Thematic Role'" In Gennaro Chierchia, Barbara H. Partee, and Raymond Tuner (eds.), Properties, Types and Meaning, II. pp 69--129
Dowty, David (1990). 1990. "Thematic Proto-Roles and Argument Selection". Language. Vol. 67, No. 3, pp. 547--619
- Hess, Thom (1973) "Agent in a Coast Salish Language". IJAL. Volume 39, No. 2.
- Hess, Thom. (1993) "A schema for the presentation of Lushootseed verb stems". 'American Indian Linguistics and Ethnography in Honor of Laurence C. Thompson', University of Montana Occasional Papers on Linguistics no. 10, ed. by Anthony Mattina, and Timothy Montler, 113–27. Missoula, MT: University of Montana.
Parsons, Parsons, Terrence (1995). Thematic Relations and Arguments. Linguistic Inquiry. Vol. 26, No. 4 (Autumn, 1995), pp. 635-662
Notes from VLAD
Joshua: How do I know how many separate es to posit?
Glenn: What is the rule of thumb in HPSG for when you want more than one event in a verb complex?
Joshua: Everything should have its own ARG0 --- just for preserving translatability into RMRS?
Joshua: What does it mean for translation if one verb is a single event, and one, with causative is two event structures? One linguistic test would be if you can find cases where aspect attaches to only one?
Francis: Causative/inchoative --- wanted to do it analytically, but Dan talked me out of it because it's not always productive. So current solution is _open_v_1_rel and _open_v_cause_rel to be able to identify. Is that the case in Lushootseed? Do you ever get underived transitives?
Joshua: Maybe just one or two in the whole lexicon. What we do have is transitive verbs where the morphology is still visible, but the bare root itself is unattested. Maybe evidence of fossilization. Very few where we can say it doesn't look like anything transitivizing has ever happened there.
Francis/Joshua: Roots are more lexemes than lexical items.
[Awkward pause while we switched over to Skype because it turns out that Google Hangouts only supports up to 10 participants and we actually had closer to 15.]
Dan: You raised about 17 very good questions, and we probably don't have time to work through all of them to any real satisfaction. Will try to jump to the ones that seem most central. The labels on arg1 type baby predications are not handles. Just a bookkeeping attribute saying that these pieces fit together to make a normal MRS predication. Don't engage in scope relations, etc. You are right to be wary of adding extra events --- our neo-Davidsonian universe already gives us lots. The anchors in RMRS are neither handles nor events. Just a bookkeeping technique to allow us to navigate back and forth.
Joshua: So to clarify: We use handles (the h sense of labels) as a way to build a scope tree -- finding which nodes are conjoined. Why is it different to pass around an anchor than to pass around an e?
Dan: You might think that the e would be a sufficient way of connecting things up. That may be true as we've moved from RMRS (14, 12 years ago) to DMRS, we have come to find it important to ensure uniqueness of ARG0s. That didn't used to be true. ARG0s used to be more promiscuous, and some EPs didn't have any (e.g. degree specifiers). It may be now that if we make the additional assumption (not built into MRS itself) of ARG0 uniqueness that there's enough info to use the event to tie RMRS units together.
Joshua: So use the es….
Dan: Be careful because the ARG0 isn't always an e. x for noun predications, degree specifiers use the underspecified type between x and e.
Joshua: Let's say I define a causative predicate with its own unique ARG0 (probably an e), and then it has another argument which is the causer and a third argument which is a caused situation. There's some trivial mapping between that representation and one in which the cause relation doesn't have its own e.
Dan: Agree that there is a close connection between the two choices, but they do have consequences.
Emily: We have examples in Japanese, from Manning, Sag and Iida 1998 where adverbial modifiers can be interpreted as either modifying the cause relation or the embedded predicate. (And the Japanese causative is arguably morphological, not periphrastic, though in Jacy we treat it as a type of auxiliary.)
紀子 が 勝 に 学校 で 走ら せ た Noriko ga Masaru ni gakkou de hashira-se-ta Noriko NOM Masaru DAT school LOC run-CAUS-PST ``Masaru made Noriko run at school.''
Dan: If you have two separate events, then you expect modifiers to be able to modify either of the events (ambiguity). If that's the case, then you can't just have a single event.
Joshua: In a world where you were allowed to share ARG0, you're making an hypothesis about items which share ARG0, that there's no linguistic process that needs to individuate these things.
Dan: I think share ARG0 is different---a third possibility. (1) Parsons-style with shared anchors, but not a separate event; (2) Two predications, _x_v_1 and _x_v_cause; (3) …
Joshua: Parsons treats causatives separately e.g. fell the tree, where he does include the cause predicate, which is different from native transitive.
Dan: Right. The Parsons representation doesn't preclude the decision to do further decomposition. The most interesting part is the ability to dissociate the assertion of a role in a predication from talking about the predication as a full entity.
Emily: Did you ever build RMRS in the ERG? Or did I make that up?
Dan: You didn't completely make it up.I tried a little experiment way back when, intrigued by the idea of doing RMRS natively and then computing the MRS from that. Temptations in terms of cleaning up the redundancy in the lexicon and cleaning up causative. I think Ann has an RMRS baby grammar, maybe with the LKB. I think she went a little further down that road during the DeepThought project.
Glenn: Why would that facilitate what you described?
Dan: The convenience is that I can create under specifications between ARG1 and ARG2, the syntax might help me differentiate later. For MRS I don't have a way to underspecify between which of two features something is linked to.
Joshua: Another question regarding Dowty's ordered argument theory of roles v. Parsons's notion which uses labels. If I am going to go this route of decomposition of predicates of arg1, etc. Is there a constant interpretation for ARG1 or does that depend on the lexical class?
Dan: Our long-standing view of MRS or RMRS is that the notion of ARG1, ARG2 is not very informative. It's always the pair of the predicate plus the argument number that tells you some. ARG1, ARG2 mean something different with _resemble_v_1 than with _chase_v_1. There's nothing inherently agent-like or patient-like about ARG1 or ARG2. Have to do a little computation to figure out the proto-agent/proto-patient classification because you have to know something about what kind of predicate you're talking about. Even in the Parsons-style there's nothing meaningful about ARG1. There's just too much variation in the world.
Joshua: Does that give me a very strange notion of the idea of predicate---because they're not constant in the model.
Emily: I think that ARG1 isn't a predicate the same way that chase is, in RMRS. The RMRS is an exploded view of the MRS, that is perhaps more convenient for composition (esp underspecification).
Joshua: So two kinds of predicates?
Dan: I think we need to refine our terminology. A predicate is a predicate symbol, plus its ARGs. That's not true of these building blocks. Likewise, just _chase_v_rel(e) isn't something to interpret until it has its ARGs, too. We need to be careful about what we're mapping to full entities in our models --- it's the EPs with their arguments, perhaps with some underspecified.
Joshua: Just yesterday the term 'elementary predication' started to mean more to me. Is there such thing as 'complex predication'? Should I think anything more into that name?
Dan: I think we were conscious that the word 'predication' is a generic/widely used, where people might talk about the predication chased the cat applied to the dog in the dog chased the cat, but we want to talk about the smaller pieces inside there. There are settings in which it's useful to talk about chase the cat as a predication on something else, so we added 'elementary predication' to talk about the smallest atomic units that are semantically coherent.
Emily: Interesting that you put it that way --- keep waiting for Joshua to ask what is special about that level, as opposed to the subatomic pieces we see in RMRS?
Dan: Ann might have more to say, but, I'm assuming that there's inference and things I can do on the EP level that don't make sense/would be category errors if tried on things like ARG1(a,e,x). I would expect these to be sharply differentiated. Maybe Ann has some tests that we can apply to differentiate.
Woodley: The ERG at least doesn't make any attempt to have those be meaningful, and you say that you do that because you think it can't be done consistently, but the folks in Haifa are trying to find something consistent there.
Dan: I would expect it to be the dominant world view that God decreed she would give us 15 or 26 thematic roles that are useful. I have watched enough people trying to do that crash against the cliffs of despair (in the AI world in the 1980s) as they try to scale that. But that's a matter of religion---I'm denying the existence of a thing, can't prove that. Haifa is trying an experiment to make that scale over a broad-coverage grammar…
Woodley: Joshua is in a position where he has some morphology that tells him something at least syntactically uniform about those types of arguments. Right?
Joshua: Right --- and the morphology is more differentiated than what we've talked about so far.
Dan: Yes, that looks very cool. A very different way of trying to differentiate than looking at Levin verb-class style alternations. Might lead to a way to classify sets of ARGns. 4-way distinction? 20-way distinction?
Joshua: Just because you might have an out-of-control causer v. a completely in control causer, are those really different semantic roles? Maybe we're mapping not into specific roles, but into something like proto-roles. A cluster of properties gives a proto-causer.
Emily: Could have the morphology also put in a modifier EP (_voluntary_a_rel etc).
Joshua: Or a feature on the event? What's the difference?
Dan: Feature predicts reentrancy.
Joshua: Like agreement?
Dan: Yes. More a heuristic than a true principle. We often have the choice between types and features. Features are compact and you have the feature composition principles…
Emily: Another difference is that once we put a feature into the semantics, we're stuck with it, but what goes in as an EP or similar can then have yet another EP "wrapped around it" that coerces something. See Francisco Costa's work on aspect.
Francis: Now that Dan's gone I can say it. I also have a hankering for semantic roles, I must admit. It would be nice to be able to access the difference/relation between _break_v_1 and _break_v_cause more directly in the grammar. I think it ties in with nominalizations. The breaking event also has similar things. Having also worked in a group where we did this back in the day in Japan. There are many very difficult edge cases, but there are also lots of easy cases, and it's very useful. We still don't have the balance right and it would be nice to find a workable subset.
David I: You're not okay of saying this in front of Dan, but you are okay of saying it front of Emily.
Francis: Emily has been less vocal in her opposition…
Emily: Maybe at the Summits, but not at UW. Definitely skeptical of a small set of consistently defined semantic roles.
Francis: Just because you're skeptical of a small set of roles doesn't mean you're skeptical of the same roles being shared between verbs and nominalizations?
Emily: Show me the data. What are the cases where we really need access to this in the grammar?
Joshua: This language does a lot with nominalization, that's for sure. What's the difference between e and x? (In terms of the model.)
Francis: xs have to be quantified.
Joshua: Emily and I were talking the other day about quantification over events.
Woodley: es are also variables, and assumed to be existentially quantified.
Joshua: Variable means something that can stand in for different things in the model?
Joshua: So something like raccoon introduces an x, but be a raccoon introduces an e? What's the dividing line, aside from morphology?
Mike G: Francis was saying that one's an individual an one's an event . More clearly: referential index v. eventuality. Ref inds need to be quantified. Also different in variable properties (PNG v TAM).
Joshua: That's exactly what Emily said when I asked her… I guess I would say is that just a reflection of the morpho-syntax then? It's true in Lushootseed that I can aspect on verbs, not on nouns, and possession on nouns but not on verbs. I've got that in the syntax, but what does it correspond to in the model?
Mike G: Not an answer to that, but something that might help: In the ERG, colors are ambiguous between e and x --- the color blue v. the sky is blue. I don't know why Dan doesn't do it for nominalizations of verbs, but there's an instance of something where a single predicate can take either an x or an e.
Woodley: You keep asking about what is the difference in the model. THe idea of a model that defines how we think about MRSs is a nice one but not a true one. No one has a model-theoretic interpretation of MRSs beyond a toy one. And MRS is not an interlingua. THere is no reason to assume that the model for ERG RMSs should be the same as the one for Lushootseed MRSs.
Joshua: I think it makes sense that the model has to be different. I'm trying to figure out even what kinds of things go in these models. The English and Lushootseed models have to be different in the collection of individual and eventualities … but shouldn't the be the same kind of things.
Emily: I'm surprised at the notion of language-specific models. Isn't the model a model of (speaker's conceptualizations of) the world? I could see the interpretation functions being language-specific… Also, I think formal semanticists think of their work as something that applies across languages.
Woodley: Might be ivory-tower syndrome. Do semanticists look at different languages?
[EMB later --- yes, many do.]
Joshua: Collection of predicates & individuals different, but I'd like to think the kind of model holds the same types of entities as the one I could build for Lushootseed.
Emily: I think it's wrong to talk about models including predicates.
Mike G: I wish Josh Cason was here… We have reality and then there's a model. I don't think the model is the real world itself. If you think about categorical theory (or something). You have things made up of metal and rubber that I call cars that might include what others call trucks. There's some set of things in my mind that fill this set. And I might have a predicate that links to this set, although it could also link to other kinds of set. Different layers of abstraction…
Joshua: I like the idea that the model is full of individuals, or that there's some way to divide the world into items and that the predicates pick out sets of individuals. Is that the right way to think about the model?
Emily: Like Woodley says, I think don't think it makes sense to talk about 'the model' as one known thing. Cd. Ann's talk at IWCS (unfortunately not recorded) where she talks about a subset of MRS (ERS?) that maps onto model-theoretically interpretable predicate logic. Outside of that, we work by analogy. Are there other sources of evidence we can use to choose between e.g. ARG0 e or ARG0 x for a particular predicate?
Woodley (in text chat): For the record: I believe models do include predicates (aka relations)
David I (in text chat): Yes, I have an a priori understanding of a predicate as a relation, too.
Francis: I teach this in Intro Semantics. Model includes sets, which might accidentally be the same, [scribe fell behind], denotation function, … from that point of view, the predicates are part of the model. But then this is the simplest possible model, and it's known not to describe things very well. No idea of time, properties… Even if we had a model it wouldn't tell us everything.
Woodley: Taking the case of I'm hungry and I have hunger. Clearly expressed differently in different languages. If they get mapped to the same thing in the model, then you're taking the position that you have a semantic interlingua, which you can for a toy example.
Francis: If at the end of the day, all you have is individual and sets of individuals, then you'd expect have hunger and be hungry to map to the same set of individuals?
Joshua: In a simple model, might we have predicates that don't map to anything?
Francis: If that predicate is not in the denotation assignment function, then it's not part of the model. Turns out that to define singing as the set of things that sings not possible for anything beyond a very small world.
Joshua: Thanks -- it was helpful to have that starting point POV.
Francis: Not sure if that's the same kind of model Ann talks about (disclaimer).
Woodley: Translate that to scenario where being a raccoon naturally described as an event in one language and another language where it's more natural to talk about raccoon as a set of individuals. Not clear that you'll find the same model that you can map into from both sides (compared to have hunger v. be hungry where you're tucking the index from hunger in the former under the rug).
Francis: Isn't the normal interpretation of being a raccoon being the set of raccoons?
Joshua: In English, the be makes that intuitive, but plenty of languages don't have copulas. In Lushootseed sentences not described in terms of verbs, but in terms of main predicates. Nominal predicates are intransitive, verbal predicates can derive transitives, but both predicates. So maybe it's the determiner that picks out xs v. sets?
Emily: Time to wrap up...