[Lingtyp] wordhood: bonded vs. bound

Sat Nov 18 16:38:08 UTC 2017

Dear Volker,

  I think most typologists are aware that (i) defining categories for coding is very hard, especially across languages -- hence all the discussions about comparative concepts on Lingtyp (some of which have subsequently been published in some form in Linguistic Typology), of which this discussion of ‘word’ is only the latest; and (ii) that typologists must usually operationalize those criteria and make the operationalizations as explicit as possible. I think that (i) and (ii) are fairly common practice in typology, despite my previous comments about essentialism and methodological opportunism (cherry-picking of criteria).

   On the other hand, your point about mono-annotator annotation is well taken. Nevertheless, the operational factor is this one:
> And I'm not saying that mono-annotator projects are useless, sometimes you just don't have the manpower for multi-annotator projects 
> 
  I have recently been working on computational projects that involve annotation, and even there, where there is a lot more large-scale funding than in typology, it is very expensive to hire and train annotators, and in the end there are maybe two annotators and a third person acting as adjudicator for a pilot annotation at most. (In fact, most of the effort in computational linguistics is towards training classifiers to do the annotation automatically on large corpora, and in my small experience those are often worse than mono-annotator annotations.)

    In typology, there is virtually no funding for any sort of multi-annotator annotation whatsoever. This is especially true for graduate students doing typological dissertations, but also for faculty doing typological research. I would guess that many typologists are aware that multi-annotator annotation is preferable, but impractical. But we don’t normally add a statement like “We are aware that engaging multiple annotators would improve the reliability of our coding and hence of the results of our crosslinguistic study; but due to lack of funding, all annotation of the data was performed by the author.” Perhaps we typologists should starting adding such statements.

Best wishes,
Bill

> On Nov 18, 2017, at 6:32 AM, Volker Gast <volker.gast at uni-jena.de> wrote:
> 
> 
> Hi Johanna, even if I could do this diplomatically, I wouldn't, and I think it wouldn't make much sense, as my point is not about specific publications or authors; it's about common practice (and common practice is reflected in the publications of 'major authorities'). But I think I get your point; so let me be a bit more specific.
> 
> A lot of (quantitative) typological work relies on 'coding': Information is extracted from grammars and transformed into a data matrix. Now, it is common practice (and I'm not excluding myself here) for the coding to be done by the analyst him/herself, and by no one else. But that's considered bad practice in other fields. Ideally, you'd need a team of annotators coding independently, on the basis of annotation guidelines. The team codes a sample, determines inter-annotator agreement, and adjusts/specifies the annotation guidelines where necessary. This is done until the inter-annotator agreement is satisfactory. And then you can start with the actual coding. Ideally, the analyst shouldn't be involved in the coding process, as her annotation decisions might be (subconsciously) influenced by her working hypotheses. (Note that this might be a viable solution to the question of how comparative concepts can reliably be defined, for a given study; you can just measure how much inter-annotator variation there is; whether or not the operationalizations make sense is a different question, of course, one of validity. When you use a set of criteria disjunctively, the question is what exactly your operationalizations are intended to represent.)
> Note that I'm not saying that there are no multi-annotator projects in typology (I'm actually involved in two such projects, though one of them is actually a comparative corpus linguistics project); but as far as I can tell, it is 'basically' comon practice for analysts to code the data themselves. And I'm not saying that mono-annotator projects are useless, sometimes you just don't have the manpower for multi-annotator projects (and one of the multi-annotator projects I'm involved in was really painful; but it was instructive to see that even for categories that we thought we had defined rather clearly, inter-annotator agreement was rather low in some cases). But as I said earlier, it would be nice to have some standards or at least general guidelines for coding typological data. Minimally, I think, the data should be published, along with at least some information on the operational tests that were applied, even if done by a single annotator.
> I hope this clarifies my (too general) remarks in my previous post.
> Volker
> 
> Am 18.11.2017 um 13:27 schrieb Johanna NICHOLS:
>> Volker,  
>> 
>> If there's a way to do this diplomatically, could you cite an example or two of  "important publications by major authorities of the field where these criteria are simply not applied"?   In linguistics we don't have as much technical comment on publications as some other fields do, and maybe we should.  In journals where I see technical comments sections those comments are refereed, edited, brief, and focused on factual and methodological matters, i.e. about empirical fundamentals and not debate on theoretical frameworks.
>> 
>> If there's no way to do it diplomatically, never mind.
>> 
>> Johanna
>> 
>> On Sat, Nov 18, 2017 at 12:37 PM, Volker Gast <volker.gast at uni-jena.de <mailto:volker.gast at uni-jena.de>> wrote:
>> Matthew -- are you saying that "one cannot rule out disjunctively defined comparative concept" because this is what you did?
>> 
>> I am not convinced by "disjunctive comparative concepts". Now, that's nothing for you to worry about -- I'm just one reader (actually, audience of your ALT/2015-talk) who doesn't buy your conclusions because he doesn't accept your operationalizations.
>> 
>> But if we want "to talk TO each other (not only PAST each other)", as Martin writes, it would be good to have what other fields call "standards of empirical research". We have copied a lot of statistical methods from fields such as the social sciences and biology. I think it would also be beneficial to take a look at their standards at the "lower" level -- for instances, wrt how data is gathered, processed and classified, how hypotheses are operationalized, etc., to make sure that the results obtained by somebody are also accepted by others (just think of the 5%-threshold for statistical significance, which is just a matter of convention).
>> I'm aware that this type of remark is annoying for some of you. I teach both corpus linguistics and typology. In corpus linguistics our students deal with very basic questions of empirical research -- like the traditional 'quality criteria' -- e.g. (external, internal) validity, objectivity, reliability -- and then, in typology, we read important publications by major authorities of the field where these criteria are simply not applied, sometimes the statistics are faulty, and students do enquire about this. What can I say? There are no research standards in typology? There is an ongoing discussion about "arbitrary/subjective/random/disjunctive comparative concepts" on the Lingtype-list? I'm afraid it wouldn't convince them. What I say is that typology still has some way to go to in terms of research methods. There are many non-trivial problems, as we have seen in various discussions on this list, and we should be aware that linguistic data is sui generis (for instance, I think we can't adopt just any method/software package from genetics). But we shouldn't use "authority" as a criterion in our methodological choices, and the choices shouldn't be made in such a way to legitimize our own research 'ex post'.
>> Volker
>> 
>> Am 18.11.2017 um 07:36 schrieb Dryer, Matthew:
>>> With respect to Martin’s comment
>>> 
>>>  
>>> “It is my impression that such ortho-affixes (= forms written as affixes) are perhaps even more common than “phonologically weak” ortho-affixes, but this is an empirical question (in his 2015 ALT abstract, Matthew mentions 248 languages with weak affixes, but 308 languages with only affixes of the Tauya type, apparently confirming my impression).”
>>> 
>>>  
>>> I realize that this is a reasonable inference from my abstract, but one often has to simplify things for the purposes of an abstract. My definition of a weak affix is very narrow and many if not most affixes that are not weak affixes by my narrow criteria can still be shown to be attached phonologically by broader criteria. Furthermore, I also treat a morpheme as an affix for the purposes of this study if it triggers phonologically conditioned allomorphy in stems it attaches to and it is clear from Macdonald’s description of Tauya that some of the ortho-affixes Martin mentions do trigger phonologically conditioned allomorphy in stems they attach to (pp 54, 72, 74, 79).
>>> 
>>>  
>>> I counted an affix as weak for the purposes of the study in my 2015 ALT talk only if the description of it in a grammar makes clear that it is nonsyllabic (or has nonsyllabic allomorphs) or that it exhibits phonologically allomorphy or triggers phonologically conditioned allomorphy in adjacent stems. But in many grammars, it is only in the discussion of phonology that it becomes clear that a given affix exhibits phonologically conditioned allomorphy or that it triggers phonologically conditioned allomorphy in adjacent stems. But because I wanted to include a large sample of languages and because it is often unclear from discussions of phonology whether particular rules apply to particular affixes or stems such affixes combine with, I adopted the procedure of not consulting the discussions of phonology in classifying ortho-affixes as weak. This made sense for my 2015 ALT talk since I was examining whether there is a suffixing preference and restricting attention to weak affixes so defined applies equally to prefixes and suffixes. For a different type of typological study, this would have been inappropriate. This illustrates how comparative concepts are specific to particular typological studies.
>>> 
>>>  
>>> Furthermore, there are other factors that I did not examine that are relevant to whether a given ortho-affix is attached phonologically. There may be clear evidence from allophonic rules, but it is often very unclear from grammatical descriptions whether particular allophonic rules apply to particular ortho-affixes or stems to which ortho-affixes are attached. And even if the information is there in the grammatical description, it may take a lot of work to see whether they apply to a particular affix. For example, careful examination of Macdonald’s description of Tauya implies that the benefactive ortho-affix -pe that Martin mentions is attached phonologically, since she gives examples of phonetic representations of forms containing this morpheme where it takes the form [-be] after /m/ ([tembe] on page 54).
>>> 
>>>  
>>> There might also be evidence from stress, but still be unclear how stress is assigned to forms including ortho-affixes. For example, Tauya has word-final stress, but it is not clear from Macdonald’s description whether this means that nouns bearing the ortho-affixes that Martin mentions take stress on the ortho-affix.
>>> 
>>>  
>>> Some of you may have noticed that what I say here contradicts what I said in my earlier email about comparative concepts needing to be exhaustive. The comparative concept I used in my 2015 ALT talk was not exhaustive and was in fact disjunctive. Since that seemed appropriate for that study, this suggests that one cannot rule out disjunctively defined comparative concepts. I sympathize with Martin’s objecting to disjunctive comparative concepts as a way to continue to use confusing and ambiguous terms and I agree that there is something odd about arbitrary disjunctive comparative concepts, but it is a mistake to simply rule out disjunctive comparative concepts.
>>> 
>>>  
>>> I should note finally that while it is clear that the ortho-affixes that Martin mentions are attached phonologically, they are actually not affixes by either his criteria or mine since they are clitics that attach to postnominal modifiers. [Martin has written about problems with the use of the term “clitic”. I am in complete agreement with him about this. But I use the term here and elsewhere in my research (including my upcoming ALT talk on the encliticization preference) as a label for a comparative concept for grammatical morphemes that are phonologically attached but attach to stems of more than one stem class.]
>>> 
>>>  
>>> Matthew
>>> 
>>> 
>>> From: Lingtyp <lingtyp-bounces at listserv.linguistlist.org <mailto:lingtyp-bounces at listserv.linguistlist.org>> on behalf of Martin Haspelmath <haspelmath at shh.mpg.de <mailto:haspelmath at shh.mpg.de>>
>>> Date: Thursday, November 16, 2017 at 7:14 PM
>>> To: "lingtyp at listserv.linguistlist.org <mailto:lingtyp at listserv.linguistlist.org>" <lingtyp at listserv.linguistlist.org <mailto:lingtyp at listserv.linguistlist.org>>
>>> Subject: Re: [Lingtyp] wordhood: bonded vs. bound
>>> 
>>> Matthew Dryer thinks that wordhood is generally understood by grammar authors in terms of bondedness (= phonological weakness, as shown by nonsyllabicity and phono-conditioned allomorphy), not in terms of boundness (= inability to occur in isolation).
>>>  
>>> I don’t know if this is true, but Matthew actually recognizes that grammars often describe grammatical markers as “affixes” even when they do not show the two “phonological weakness” (or bondedness) features.
>>>  
>>> For example, Tauya (a language of New Guinea) is said to have (syllabic) case suffixes, but these never show any allomorphy, e.g.
>>>  
>>> fena’a-ni [woman-ERG]
>>> na-pe [you-BEN]
>>> wate-’usa [house-INESS]
>>> Aresa-nani [Aresa-ALL]
>>> Tauya-sami [Tauya-ABL] (MacDonald 1990: 119-126)
>>>  
>>> It is my impression that such ortho-affixes (= forms written as affixes) are perhaps even more common than “phonologically weak” ortho-affixes, but this is an empirical question (in his 2015 ALT abstract, Matthew mentions 248 languages with weak affixes, but 308 languages with only affixes of the Tauya type, apparently confirming my impression).
>>>  
>>> For this reason, I have suggested that the stereotypical “affix” notion should perhaps be captured in terms of boundness together with single-root-class adjacency. Since the Tauya case-markers attach only to nouns, they count as affixes; by contrast, if a bound role marker attaches to both nouns (English “for children”) and adjectives (“for older children”) as well as to other elements (“for many children”), we do not regard it as an affix (but as a preposition), even if it is bound (= does not occur in isolation; English "for" does not).
>>>  
>>> Matthew quite rightly points out that this notion of boundness (which goes back at least to Bloomfield 1933: §10.1) implies that most function words in English are bound, and in fact most function words in most languages are bound – but this is exactly what we want, I feel, because the best way to define a “function word” is as a bound element that is not an affix. Linguists often think of function words (or “functional categories”) as defined semantically, but it is actually very hard to say what is the semantic(-pragmatic) difference between a plural marker and a word like “several”, between a dual marker and the word “two”, between a past-tense marker and the expression “in the past”, or between a comitative marker and the word “accompany”. It seems to me that these distinctions are best characterized in terms of boundness, i.e. inability to occur in isolation.
>>>  
>>> It may be true that occurrence in isolation is a feature of an element that is not easy to elicit from speakers, but in actual language use, there are a very large number of very short utterances, so at least positive evidence for free status (=non-bound status) is not difficult to obtain.
>>>  
>>> In any event, it seems clear to me that some key concepts of grammatical typology such as “flag” (= bound role marker on a nominal) and “person index” (= bound person marker, generally on a verb) require the Bloomfieldian boundness notion, and that these concepts are much easier to work with in typology than the traditional stereotypical notions of “case”, “adposition”, “agreement marker”, and “pronominal clitic”. (For bound person forms, this was a major lesson of Anna Siewierska’s 2004 book “Person”.)
>>>  
>>> Best,
>>> Martin
>>> 
>>> On 14.11.17 07:02, Dryer, Matthew wrote:
>>>> I have a number of problems with Martin’s proposal:
>>>> 
>>>>  
>>>> "Here’s a proposal for defining a notion of “affix”, in such a way that the results do not go too much against our intuitions or stereotypes:
>>>> 
>>>>  
>>>> An affix is a bound form that always occurs together with a root of the same root-class and is never separated from the root by a free form or a non-affixal bound form."
>>>> 
>>>>  
>>>> If one examines the notion of “bound” from his 2013 paper, I believe it implies a comparative concept of affix that differs greatly from what most linguists (at last most non-generative linguists) understand by the term. That’s not a problem for it as a comparative concept, but it is a comparative concept that differs considerably from the stereotype.
>>>> 
>>>>  
>>>> Martin’s definition of “free and “bound” from his 2013 paper is as follows:
>>>> 
>>>>  
>>>> "But distinguishing in a general way between bound elements and free elements is quite straightforward, because there is a single criterion: Free forms are forms that can occur on their own, i.e. in a complete (possibly elliptical) utterance (Bloomfield 1933: 160). This criterion correlates very highly with the criterion of contrastive use: Only free forms can be used contrastively."
>>>> 
>>>>  
>>>> First, I find the notion of complete utterance ambiguous. Does it mean utterances in normal speech or does it include metalinguistic uses (like “What is the last word in the sentence “Who are you going with”? Answer “with”). I would assume that it does not include such metalinguistic uses. But then many if not most so-called function words in English would count as bound since they cannot be used as complete utterances. Perhaps other speakers of English would have different intuitions, but if so that only indicates the lack of clarity in the notion. Furthermore, for many function words in English, I am not sure how to judge whether they can occur alone as utterances. Many such so-called function words would appear to count as bound by Martin’s definition, though they would not count as affixes since they lack other properties in his definition of “affix”.
>>>> 
>>>>  
>>>> Second, many languages have grammatical morphemes that must occur adjacent to an open class word but which behave as separate words phonologically. These would all apparently count as affixes by Martin’s definition. Again, I have no problem with this as a comparative concept, only that it means his notion of affix deviates considerably from the stereotype.
>>>> 
>>>>  
>>>> Third, Martin says that his criterion “correlates very highly with the criterion of contrastive use”. But by my intuitions, the ability to occur as complete utterances does not correlate closely with the criterion of contrastive use, since most so-called function words CAN occur with contrastive use (such as can in this sentence!), as can some morphemes that are conventionally treated as affixes, like un- in “I’m not happy, I’m UNhappy”. Of course, Martin might argue that un- is more like so-called function words and less like morphemes conventionally treated as affixes. But the fact remains that un- is easily the locus of contrast but cannot be used as a complete utterance. I thus see no evidence of a close correlation between the ability to occur as a complete utterance and the ability to be the locus of contrast.
>>>> 
>>>>  
>>>> Finally, it is my experience that languages differ in their conventions regarding what can be a complete utterance. Imagine two closely related languages that differ in their grammatical rules governing what is a complete utterance. By Martin’s definition, there might be a large number of morphemes that count as separate words in one language but as affixes in the other language. This strikes me as odd. It seems odd to have a criterion for what is a word and what is an affix so dependent on the grammatical rules in the language for what constitutes a complete utterance.
>>>> 
>>>>  
>>>> Matthew
>>>> 
>>>> 
>>>> From: Lingtyp <lingtyp-bounces at listserv.linguistlist.org <mailto:lingtyp-bounces at listserv.linguistlist.org>> on behalf of Martin Haspelmath <haspelmath at shh.mpg.de <mailto:haspelmath at shh.mpg.de>>
>>>> Date: Sunday, November 12, 2017 at 10:47 PM
>>>> To: "lingtyp at listserv.linguistlist.org <mailto:lingtyp at listserv.linguistlist.org>" <lingtyp at listserv.linguistlist.org <mailto:lingtyp at listserv.linguistlist.org>>
>>>> Subject: Re: [Lingtyp] wordhood
>>>> 
>>>> Mattis List and Balthasar Bickel rightly emphasize that “word” is not a Platonic entity (a natural kind) that exists in advance of language learning or linguistic analysis – few linguists would disagree here, not even generativists (who otherwise liberally assume natural-kind catgeories).
>>>> 
>>>>  
>>>> But I think many linguists still ACT AS IF there were such a natural kind, because the “word” notion is a crucial ingredient to a number of other notions that linguists use routinely – e.g. “gender”, which is typically defined in terms of “agreement” (which is defined in terms of inflectional marking on targets; and inflection is defined in terms of “word”).
>>>> 
>>>>  
>>>> So is it possible to define a comparative concept ‘word’ that applies to all languages equally, and that accords reasonably with our stereotypes? Note that I didn’t deny this in my 2011 paper, I just said that nobody had come up with a satisfactory definition (that could be used, for instance, in defining “gender” or “polysynthesis”). So I’ll be happy to contribute to a discussion on how to make progress on defining “word”.
>>>> 
>>>>  
>>>> Larry Hyman notes that other notions like “syllable” and “sentence” are also problematic in that they also “leak”. However, I think it is important to distinguish two situations of “slipperiness”:
>>>> 
>>>>  
>>>> (1) “Leakage” of definitions due to vague defining notions
>>>> 
>>>>  
>>>> (2) Incoherence of definitions due to the use of different criteria in different languages
>>>> 
>>>>  
>>>> The first can be addressed by tightening the defining notions, but the second is fatal.
>>>> 
>>>>  
>>>> To take up Östen Dahl’s example of the “family” notion: In one culture, a family might be said to be a set of minimally three living people consisting of two adults (regardless of gender) living in a romantic relationship plus all their descendants. In another culture, a family might be defined as a married couple consisting of a man and a woman plus all their living direct ancestors, all their (great) uncles and (great) aunts, and all the descendants of all of these.
>>>> 
>>>>  
>>>> With two family concepts as different as these, it is obviously not very interesting to ask general cross-cultural questions about “families” (e.g. “How often do all family members have meals together?”). So the use of different criteria for different cultures is fatal here.
>>>> 
>>>>  
>>>> What I find worrying is that linguists often seem to accept incoherent definitions of comparative concepts (this was emphasized especially in my 2015 paper on defining vs. diagnosing categories). Different diagnostics in different languages would not be fatal if “word” were a Platonic (natural-kind) concept, but if we are not born with a “word” category, typologists need to use the SAME criteria for all languages.
>>>> 
>>>>  
>>>> So here’s a proposal for defining a notion of “simple morphosyntactic word”:
>>>> 
>>>>  
>>>> A simple morphosyntactic word is a form that consists of (minimally) a root, plus any affixes.
>>>> 
>>>>  
>>>> Here’s a proposal for defining a notion of “affix”, in such a way that the results do not go too much against our intuitions or stereotypes:
>>>> 
>>>>  
>>>> An affix is a bound form that always occurs together with a root of the same root-class and is never separated from the root by a free form or a non-affixal bound form.
>>>> 
>>>>  
>>>> These definitions make use of the notions of “root” and “root-class” (defined in Haspelmath 2012) and  “bound (form)”                                 vs. “free (form)” (defined in Haspelmath 2013). All these show leakage as in (1) above, but they are equally applicable to all languages, so they are not incoherent. (I thank Harald Hammarström for a helpful discussion that helped me to come up with the above definitions, which I had not envisaged in 2011.)
>>>> 
>>>>  
>>>> (What I don’t know at the moment is how to relate “simple morphosyntactic word” to “morphosyntactic word” in general, because I cannot distinguish compounds from phrases comparatively; and I don’t know what to do with “phonological word”.)
>>>> 
>>>>  
>>>> Crucially, the definitions above make use of a number of basic concepts that apply to ALL languages in the SAME way. David Gil’s proposal, to measure “bond strength” by means of a range of language-particular phenomena, falls short of this requirement (as already hinted by Eitan Grossman). Note that the problem I have with David’s proposal is not that it provides no categorical contrasts (recall my acceptance of vagueness in (1) above), but that there is no way of telling which phenomena should count as measuring bond strength.
>>>> 
>>>>  
>>>> David’s approach resembles Keenan’s (1976) attempt at defining “subject” (perhaps not by accident, because Ed Keenan was David’s PhD supervisor), but I have a similar objection to Keenan: If different criteria are used for different languages, how do we know that we are measuring the same phenomenon across languages? Measuring X by means of Y makes sense only if we know independently that X and Y are very highly correlated. But do we know this, for subjects, or for bond strength?
>>>> 
>>>>  
>>>> Best,
>>>> 
>>>> Martin
>>>> 
>>>>  
>>>> 
>>>> -- 
>>>> Martin Haspelmath (haspelmath at shh.mpg.de <mailto:haspelmath at shh.mpg.de>)
>>>> Max Planck Institute for the Science of Human History
>>>> Kahlaische Strasse 10	
>>>> D-07745 Jena  
>>>> &
>>>> Leipzig University 
>>>> IPF 141199
>>>> Nikolaistrasse 6-10
>>>> D-04109 Leipzig    
>>>> 
>>>> 
>>>> 
>>>> 
>>>> 
>>> 
>>> -- 
>>> Martin Haspelmath (haspelmath at shh.mpg.de <mailto:haspelmath at shh.mpg.de>)
>>> Max Planck Institute for the Science of Human History
>>> Kahlaische Strasse 10	
>>> D-07745 Jena  
>>> &
>>> Leipzig University 
>>> IPF 141199
>>> Nikolaistrasse 6-10
>>> D-04109 Leipzig    
>>> 
>>> 
>>> 
>>> 
>>> 
>>> 
>>> 
>>> _______________________________________________
>>> Lingtyp mailing list
>>> Lingtyp at listserv.linguistlist.org <mailto:Lingtyp at listserv.linguistlist.org>
>>> http://listserv.linguistlist.org/mailman/listinfo/lingtyp <http://listserv.linguistlist.org/mailman/listinfo/lingtyp>
>> 
>> -- 
>> Prof. Volker Gast
>> English and American Studies
>> Ernst-Abbe-PLatz 8
>> D-07743 Jena
>> 
>> Fon: ++49 3641 9-44546
>> Fax: ++49 3641 9-44542
>> 
>> _______________________________________________
>> Lingtyp mailing list
>> Lingtyp at listserv.linguistlist.org <mailto:Lingtyp at listserv.linguistlist.org>
>> http://listserv.linguistlist.org/mailman/listinfo/lingtyp <http://listserv.linguistlist.org/mailman/listinfo/lingtyp>
>> 
>> 
> 
> -- 
> Prof. Volker Gast
> English and American Studies
> Ernst-Abbe-PLatz 8
> D-07743 Jena
> 
> Fon: ++49 3641 9-44546
> Fax: ++49 3641 9-44542
> _______________________________________________
> Lingtyp mailing list
> Lingtyp at listserv.linguistlist.org
> http://listserv.linguistlist.org/mailman/listinfo/lingtyp

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://listserv.linguistlist.org/pipermail/lingtyp/attachments/20171118/fce77279/attachment.htm>