[developers] Realizational v. incremental morphology

Olivier Bonami olivier.bonami at linguist.univ-paris-diderot.fr
Mon Nov 9 20:58:52 CET 2015


> Le 9 nov. 2015 à 20:02, Emily M. Bender <ebender at uw.edu> a écrit :
> 
> Thank you, Olivier.  I think your example helps clarify things somewhat:
> the incremental/realizational distinction is puzzling from an HPSG perspective
> at least in part because unification makes multiple exponence a non-issue, 
> but I still think that the process-neutral stance of HPSG is the deeper reason.
> 
> With apologies for replying without doing my homework (i.e. reading the journal
> paper): It's hard for me to grasp how a system without any intermediate
> representations will scale to languages with long strings of position classes
> housing, among other things, valence changing or other derivational morphology.
> Perhaps I'll be enlightened when I read the paper!

For now our official policy is that derivational morphology *is* about relating signs to signs, but that precisely inflection is not. We take this to be a useful feature of the analysis rather than a bug, as you will see when you read the paper ;-)

Best

Olivier
> 
> Emily
> 
> On Mon, Nov 9, 2015 at 10:45 AM, Olivier Bonami <olivier.bonami at linguist.univ-paris-diderot.fr> wrote:
> Dear Emily, dear all,
> 
> I can only give partial answers to your question, but here is what I can say:
> 
> - In answer to question 2, we do not have an implementation of the Bonami & Crysmann 2013 story or any other version of our story so far. Berthold certainly would be able to speculate about questions 3-5, but the fact of the matter is that as long as we don't have an implementation, it is only a matter of speculation.
> 
> - I should point you to our more recent paper in Journal of Linguistics that supersedes Bonami & Crysmann 2013 and gives many more details on various things: http://dx.doi.org/10.1017/S0022226715000018
> 
> - Now to the incremental/realisational distinction. Let me quote in full the relevant passage from Stump (2001:2):
> 
> > According to INCREMENTAL theories, inflectional morphology is information-increasing; that is, words acquire morphosyntactic properties only as a concomitant of acquiring the inflectional exponents of those properties. On this view, _likes_ acquires the properties ‘sg subject agreement’, ‘present tense’, and ‘indicative mood’ only through the addition of -s (whether this is inserted from the lexicon or is introduced by rule). According to REALIZATIONAL theories, by contrast, a word’s association with a particular set of morphosyntactic properties licenses the introduction of those properties’ inflectional exponents; on this view, the association of the root _like_ with the properties ‘sg subject agreement’, ‘present tense’, and ‘indicative mood’ licenses the attachment of the suffix -s (whether this attachment is effected by lexical insertion or by the application of a morphological rule).
> 
> Clearly, even for Stump himself, this is related to an analysis vs. generation intuition; witness the following formulation from Bonami & Stump (forthcoming)---this can be downloaded from my web page:
> 
> > The definition of an inflectional system is realizational if it is formulated as deducing a word’s form from its content; the opposite of a realizational definition is one that is incremental, which constructs a word’s content on the basis of its form
> 
> However if you look at the theories Stump actually uses to illustrate the difference, it becomes clear (I hope) that this is not the full story. Both classical Item and Arrangement and Item and Process approaches are supposed to be instances of incremental theories. What these approaches share is the assumption that morphologically complex words involve intermediate representations associated with distinct content. For instance under a typical IA or IP analysis, the Spanish adjective "roj-a-s" red-F-PL will involve three separate items with distinct associated content:
> roj     : ‘red’
> roja    : ‘red’, feminine
> rojas   : ‘red’, feminine, plural
> This is what justifies the use of the term "incremental": content is added in increments that correspond to addition of affixes/application of processes.
> 
> By contrast, in a realisational theory like PFM, although the description of a complex word also involves intermediate steps, there is no notion of these steps being associated with increments in content: realization rules are sensitive to morphosyntactic features but only transition from an input phonology to an output phonology.
> Using the same example: the {fem,pl} form of "rojo" is found by taking the lexeme's basic stem "roj" and applying successively two rules:
> Block I rule: X:{fem}->Xa
> Hence roj -> roja because {fem} ⊆ {fem,pl}
> Block II rule: X:{pl}->Xs
> Hence roja -> rojas because {pl} ⊆ {fem,pl}
> 
> Although on these two examples the constrast is rather clear, it is not completely obvious how Stump's distinction applies in general; in particular the general modeling assumptions of HPSG make intuitions about incrementality difficult to make sense of.
> 
> Berthold and I sell our framework as realizational for two reasons:
> (i) We do not have intermediate representations *at all*: it makes no sense to talk of incrementality since the relation between form and content is stated directly for the fully inflected word.
> (ii) Our framework deals seamlessly with the same issues as realizational approaches (e.g. extended exponence)
> Hence if what we do has to be likened with one of the two camps, it is rather realizational than incremental.
> 
> The interesting question is what to think of the kind of approach you are pursuing, which I take to be a version of Item and Process (inflection rules are transitions from signs to signs, but affixes are not signs) but where the modeling assumptions of HPSG make it possible for a rule to monotonously narrow down the morphosyntactic content of both their input and their output. In that context I think both interpretations make sense. Let me take the same example as before, and assume the two schematic rules:
> [PHON [1], HEAD [1]] --> [PHON [1]+a, HEAD [1][GEN fem]]
> [PHON [1], HEAD [1]] --> [PHON [1]+s, HEAD [1][NUM pl]]
> 
> This will license a word like
> 
> |PHON rojas                 |
> |HEAD [1][GEN fem, NUM pl]  |
> |DTRS < |PHON roja        |>|
> |       |HEAD [1]         | |
> |       |DTRS <|PHON roj|>| |
> |       |      |HEAD [1]| | |
> |       |      |DTRS <> | | |
> 
> In a sense this is non-incremental, as the HEAD value of the three signs are identified---there is no more content to the outermost sign than to the two others. In another sense, it could be said to be incremental, in the sense that each inflection rule contributes to a narrowing down of possible HEAD values for both its input and output. I suspect that the second sense is what is relevant for present purposes---that this is how Greg would see it, but in a way this ambiguity shows that the distinction is not as clear as one would like.
> 
> I hope this helps rather than adds more confusion.
> 
> Best
> 
> Olivier
> 
> 
> > Le 9 nov. 2015 à 18:25, Emily M. Bender <ebender at uw.edu> a écrit :
> >
> > Dear Berthold,
> >
> > I hope you'll find time to respond in this conversation! I've also
> > added Olivier to the cc -- Olivier, please see below for context; I hope
> > you'll have time to reply, too.
> >
> > Thanks,
> > Emily
> >
> > On Sat, Oct 31, 2015 at 12:16 PM, David Inman <davinman at uw.edu> wrote:
> > My objection to this during Emily's meeting I think remains: what is the actual difference between these views? It seems like a difference of perspective more than a difference in theoretical predictions, with incremental a generative perspective and relational a set-like perspective. It all seems a bit like wondering whether a set is accurately defined through induction or intersection - it doesn't matter, both describe the same thing, use whatever is most convenient. Unless I am missing something - and I consider this very likely -  are there actual theoretically different claims made between the two?
> >
> >
> > _____________________________
> > From: Guy Emerson <gete2 at cam.ac.uk>
> > Sent: Saturday, October 31, 2015 1:52 AM
> > Subject: Re: [developers] Realizational v. incremental morphology
> > To: Emily M. Bender <ebender at uw.edu>
> > Cc: Berthold Crysmann <berthold.crysmann at gmail.com>, developers <developers at delph-in.net>
> >
> >
> >
> > Here are Stump (2001)'s definitions:
> >
> > "According to INCREMENTAL theories, inflectional morphology is information-increasing; that is, words acquire morphosyntactic properties only as a concomitant of acquiring the inflectional exponents of those properties. On this view, `likes' acquires the properties ‘3sg subject agreement’, ‘present tense’, and ‘indicative mood’ only through the addition of `-s' (whether this is inserted from the lexicon or is introduced by rule). According to REALIZATIONAL theories, by contrast, a word’s association with a particular set of morphosyntactic properties licenses the introduction of those properties’ inflectional exponents; on this view, the association of the root `like' with the properties ‘3sg subject agreement’, ‘present tense’, and ‘indicative mood’ licenses the attachment of the suffix `-s' (whether this attachment is effected by lexical insertion or by the application of a morphological rule)."
> >
> > 2015-10-30 19:55 GMT+00:00 Emily M. Bender <ebender at uw.edu>:
> > Dear Berthold & developers by cc,
> >
> > My students I and were discussing Bonami & Crysmann 2013 (HPSG proceedings) in our group meeting this week, and that led me to a question about the relationship between the characterization of morphological theories (due I think to Stump 2001) and what we do in the implementation. That characterization contrasts "inferential" and "lexical" approaches (where the former relates inflected forms to stems via rules and the  latter has morphemes in the lexicon as separate entries) and "realizational" and "incremental" approaches.
> >
> > I don't fully have a handle on what the "realizational"/"incremental", and I don't have Stump's to hand, but here's how Ida Toivonen summarized the definitions in her review of Stump's book on LINGUIST List:
> >
> > "In incremental theories, morphosyntactic information gets added
> > incrementally as morphemes are added to a stem. In a realizational
> > theory, a word's association with certain morphosyntactic properties
> > licenses the appropriate affixes."
> > https://linguistlist.org/issues/13/13-622.html
> >
> > On one way of reading this, it seems to me that the inferential/realizational dichotomy is itself opposed to the notion of order- and process- independence of constraints which I think is central to HPSG.  That is, it looks like an "incremental" theory is taking an analysis/parsing point of view (morphemes are given; their morphosyntactic effect is calculated) and a "realizational" theory is taking a realization/generation point of view (morphosyntactic features are given; the morphemes that realize them are licensed/generated).
> >
> > Joshua Crowgey pointed out that perhaps the distinction has to do with whether the attachment of morphemes (or application of morphological rules) is treated as something that can have an internal order or not, analogous to how our phrase structure rules can be ordered via constraints on the rules themselves (e.g. head-subj or head-spr refuse a daughter with a non-empty COMPS list, in effect enforcing head-comps to apply lower/'earlier' in the tree).
> >
> > Having laid out (and perhaps spread) my confusion, a few questions:
> >
> > 1. How do you see the incremental/realizational distinction relating to our implementation (within DELPH-IN) of constraint-based grammar?
> > 2. Have you done any implementation based on the theory you lay out in Bonami & Crysmann 2013?
> > 3. How do you associate the presence of something like [ MUD {neg} ] (or analogously for applicatives, say) with the morphosyntactic and morphosemantic consequences?
> > 4. How do you implement a check on whether a form is fully inflected?
> > 5. How do you implement grammaticality that is dependent on comparison across forms (per Panini's Principle)
> >
> > Thanks,
> > Emily
> >
> > --
> > Emily M. Bender
> > Professor, Department of Linguistics
> > Check out CLMS on facebook! http://www.facebook.com/uwclma
> >
> >
> >
> >
> >
> >
> > --
> > Emily M. Bender
> > Professor, Department of Linguistics
> > Check out CLMS on facebook! http://www.facebook.com/uwclma
> 
> 
> 
> 
> -- 
> Emily M. Bender
> Professor, Department of Linguistics
> Check out CLMS on facebook! http://www.facebook.com/uwclma




More information about the developers mailing list