Should long vowels be counted for Xaracuu / xara1243 (and other New Caledonian languages)

[antipodite]

So Angela asked in the CoOL channel if the vowel count of 30 for Xaracuu / xara1243 were correct given that its sister languages in the dataset have much lower numbers of vowels reported.

This seems to be an issue of differing phonemic analyses. I looked at some of the references for these languages, and it is claimed by several authors such as Moyse-Faurie 1989, Osumi 1995 and Zetterberg 2021 that these languages have a phonemic long vowel counterpart to each short oral and nasal vowel which is listed for them.

I looked at McCracken 2019's discussion of this issue in her grammar of Belep (a northern language). She notes that the traditional approach to NC languages treats long vowels as separate phonemes instead of sequences of short vowels as has been done for Xaracuu. However in Belep she says that although there are minimal pairs that seem to bear this out, there is always a syllable boundary between the two "halves" of a long vowel and that speakers have no intuition regarding vowel length, but they reliably segment words such as /aawa/ into a.a.wa for example. Like in Xaracuu, Aijie etc Belep displays hiatus in sequences of short vowels, i.e. they are realised as heterosyllabic sequences of vowels.

Russell also pointed out in the chat that when there is reduplication of a syllable containing a long vowel in Xaracuu, only "half" of the long vowel is reduplicated, indicating that "long" vowels are not necessarily perceived as a single unit. This also appears to be the case in the closely related language Tiri / Tinrin / tiri1258 (Osumi 1995).

Zetterberg says of Aijie, following de la Fontinelle, that "Vowel length is phonemically distinct in Aijie" and that each oral and nasal vowel can appear as short or long, but then goes on to say that "this is analysed as an aspect of syllable structure", which seems to imply something similar to what is reported for Belep above.

Looking at Angela's comment then it seems like the issue is different authors have chosen different analyses, long vowels or vowel sequences (and that there is seemingly more justification to choose the latter), and different sources for different languages are showing up as a discrepancy in the number of vowels in the data. It seems like the best thing to do is not to count long vowels separately for the languages at issue, but it would possibly also be worth going through the other NC languages in the dataset to make sure the same issue isn't hiding in some of the languages with lower vowel counts. I get the impression that this issue is relevant for NC as a whole from McCracken's comments

[angela-mc]

I will add couple of figures to this issue. I went through the Oceanic tree pruned for the languages for which we have data on phoneme inventories and extracted pairs of sister languages (sharing the same parent node). Then I looked at the difference in #vowels between them, and I noticed that, Canala (Canala(glottocode xara1244)) & Aragur (abvdoceanic-Aragur, glottocode xara1243) had a huge difference in #vowels compared to all other sister-languages (Canala is listed with 30 vowels and Aragur with 12), despite the branch length connecting them not being particularly long. Of course this can happen, but I thought I would flag the example in case it is a scenario of discrepancies between how #vowels are counted. VowelDifferences.pdf

[SimonGreenhill]

ok, so Angela's analysis here makes it seem like (Canala, Aragur) are a pathological pair. Can we get a histogram on the overall number of vowels and number of consonants and see if we can spot other outliers?

[angela-mc]

VowelDiff.pdf VowelNumber.pdf

These are histograms for the #vowels: total number & differences between sister-languages. I've spelled out the right tails on each plot (I've been very generous, I would say for differences in #vowels, only 18 & 10 are outliers, but I thought might as well flag generously for illustration/exploration); for differences I've written the pair and then in brackets the number of vowels for each language.

[angela-mc]

ConsonantDifferences.pdf ConsonantNumber.pdf (same as above, but with number of consonants)

[angela-mc]

(I will attach the .csv file with the numbers too, the columns are: Language1 and Language2 (sister-languages), BranchLength (length of branches connecting the parent to each of the languages; the tree I am using is ultrametric, so these branches have the same length), vowel1 and vowel2 (number of vowels for each of the languages), DiffVowels (difference in number of vowels), consonants1 and consonants2 (number of consonants for each of the languages), DiffConsonants (difference in number of consonants), DiffVowelsTime (difference in number of vowels between the two sister languages divided by branch length), DiffConsonantsTime (difference in number of consonants between the two sister languages divided by branch length). surori_phoneme_inventories.csv

Looking at difference in number of consonants or vowels divided by branch length is also indicative I think, in the sense that we could expect sister languages who had more time to diverge to be more different in their number of vowels/consonants. I am attaching two histograms corresponding to DiffVowelsTime and DiffConsonantsTime from the dataframe. The clear outlier for DiffVowelsTime is the pair: Mortlockese_Lukunosh_923 and Satawalese_343 (outlier for the difference in vowels between the two languages as well with 10 vowels difference). DiffPerTime.pdf

[SimonGreenhill]

Hmm, it looks like the outlier Ulithian should have 19 consonants not 45.

It also worries me that Nimoa and Sudest have a difference of 15 despite being sisters.

I think we need to have some formal evaluation of how close we are to something like Phoible. Angela -- can you do a comparison between N Consonants & Vowels in our data and that listed in Phoible? If you get stuck I can do it but have a full month coming up.

[angela-mc]

Sure! Do we have some expectations about the differences in #consonants or #vowels between sister-species?

I am asking because this is helpful for setting the priors for the Fabric Model, there is a prior for the differences. It is by default set to Weibull distribution, with a shape and scale that translate to a left-tailed distribution (scale = 1.1, shape = 1.5). I think Weitbull left-tailed makes sense (ie: we expect small changes in general from ancestor to descendant, but we allow for some bigger changes to happen as well). I am trial-ing priors that simply mirror the distribution of differences in the contemporary languages. but I was wondering if there are any intuitions we might have based on linguistic knowledge?

[HedvigS]

I found 17 languages that are both in abvdoceanic and phoible. Here are scatterplots comparing their number of consonants and vowels. color reflects absolute diff

I matched over language level ID. Languages in abvdoceanic or PHOIBLE that didn't have glottocodes at all were ignored.

code for making the comparison is here. https://github.com/angela-mc/LinguisticDisparity/blob/hedvig_help_phoible/Rscripts/compare_abvdoceanic_phoible.R

If you know a way of better matching PHOIBLE entries to glottocodes or any other way of increasing the overlap, I'm all ears.

[HedvigS]

The highest diff is motl1237, where PHOIBLE says 15 consonants and abvdoceanic says 25.

[HedvigS]

Table for your browsing.

Language_level_ID	Consonants_PHOIBLE	Vowels_PHOIBLE	Consonants_abvdoceanic	Vowels_abvdoceanic	vowels_diff	consonant_diff
maor1246	10	10	11	10	0	1
hawa1245	9	10	9	6	4	0
koko1269	22	5	18	6	1	4
mana1295	13	5	22	6	1	9
motl1237	15	7	25	8	1	10
dobu1241	19	5	20	8	3	1
matb1237	15	5	22	7	2	7
seim1238	11	17	12	10	7	1
siar1238	15	7	19	7	0	4
sout2856	15	5	15	5	0	0
sata1237	29	13	29	12	1	0
yape1248	28	16	26	14	2	2
woge1237	15	5	22	5	0	7
loni1238	16	7	20	6	1	4
bari1286	12	5	14	6	1	2
biak1248	14	10	17	7	3	3
hoav1238	16	5	17	7	2	1

[HedvigS]

For what it's worth, I think @angela-mc 's approach of comparing sisters is really smart, especially since there are so few matches between PHOIBLE and abvdoceanic. I don't know if what should be done is re-evaluate the underlying data or the way that the phoneme counts are extracted or the orthography profiles, or all three. I'll leave that up to you all.

I alos thought we already talked about this very same issue several months ago? Haha or am I dreaming again? What was the solution proposed then?

[maryewal]

RE: Xaracuu, according to Clarie Moyse Faurie, length is definitely phonemic without question (this was also the analysis of Jean Claude Rivierre when he went to check on her work apparently). She has lots of minimal pairs to support this. Essentially, there are 11 plain short vowels + 6 nasal vowels, all with length distinction (as we know is reported in the literature). The choice to write long vowels as 2 short vowels was due to the overwhelming diacritic marking already on vowels in the language - she says orthography in this case should not be considered as indicative of long vowels actually being underlyingly short. There are a few rare cases when a seemingly long vowel is actually 2 short vowels, but these instances have clear historical trajectories (consonant loss) and are separate from the issue of phonemic length.

[maryewal]

RE: "It seems like the best thing to do is not to count long vowels separately for the languages at issue, but it would possibly also be worth going through the other NC languages in the dataset to make sure the same issue isn't hiding in some of the languages with lower vowel counts. I get the impression that this issue is relevant for NC as a whole from McCracken's comments" I disagree with the first part of this. I think we should count long vowels as phonemic when they are reported as such. New Caledonian scholars agree that these languages are unusual in their unusually large vowel inventories, and we don't have enough to counter their wealth of expertise and peer-reviewed publication on the matter. The issue of phonemic length here may indeed be worth questioning further and would be an excellent study that I am certainly interested in pursuing together, but until we do that, we have to rely on what is published/reported by experts. I agree with the second part, though, that we should make sure all the NC languages are accounted for as consensus in the available literature suggests. My conversations with Claire today make me think that we should actually be seeing higher counts in other nearby NC languages as well. The Northern languages may indeed be different.

[maryewal]

Regarding the stark difference between Aragur et Canala: The issue may be due to source. Leenhardt is our source for Aragur, which has a red flag on it concerning reliability. Canala is coming from George Grace's work so is going to be more complete. But another question, we also have sets from CMF for both of these languages, both with decent coverage. 1576 | Xârâcùù | xara1244 | Claire MOYSE-FAURIE & Marie-Adèle NECHERÖ-JOREDIE | 250 1609 | Xârâgurè | xara1243 | Claire Moyse-Faurie | 198 Are these showing the same discrepancy? Or have we chosen the Leenhardt/Grace sets over CMF's for the analysis? @angela-mc @SimonGreenhill

[maryewal]

Hmm, it looks like the outlier Ulithian should have 19 consonants not 45.

@SimonGreenhill We dealt quite a bit with Ulithian already (https://github.com/lexibank/abvdoceanic/issues/16), but maybe we need to check it through again to be sure the changes we made stuck? @antipodite should we take another look?

[maryewal]

Sure! Do we have some expectations about the differences in #consonants or #vowels between sister-species?

@angela-mc This is an excellent question. My thinking is that we would expect very closely related languages to resemble each other in numbers of both consonants and vowels (among Oceanic languages) @SimonGreenhill, thoughts?

[SimonGreenhill]

Great, can we get @antipodite to look at the mismatches in the languages that overlap between phoible and abvd -- what are we getting wrong and why/

Angela: don't think there are any intuitions, but could we use phoible to construct an empirical weibull prior e.g. get a distribution of differences between sisters in phoible and construct a weibull to mimic that?

[antipodite]

Hmm, it looks like the outlier Ulithian should have 19 consonants not 45.

@SimonGreenhill We dealt quite a bit with Ulithian already (#16), but maybe we need to check it through again to be sure the changes we made stuck? @antipodite should we take another look?

IIRC the issue with Ulithian was that the data contained words from two sources with different orthographies. I thought we had changed the orthography profile already to correct this but I'll have another look.

I'll investigate the mismatches between phoible and ABVD, I also have my own data on inventories for New Caledonia so can compare against this also

[antipodite]

Alright, I ran Simon's phoible to abvdoceanic comparison script again, comparing phoible inventories with inventories for the latest version of this repository.

(sorry, accidentally closed the issue) So we have some languages that have more phonemes in Phoible than ABVD and some that have more. These languages have more phonemes in ABVD than in Phoible, ordered by difference:

Of these:

• Marquesan looks pretty straightforward, the inventory generation is recording it as having geminate consonants for some reason. This can be corrected in the orthography profile.
• The Vanuatu languages Malo, Anejom and Mwotlap have many more phonemes generated from ABVD than are in Phoible. I'm not so familiar with these languages, but looking at the ABVD it looks as if it's maybe an issue of the orthography used in the ABVD entries being more phonetic than phonemic? Someone more familiar with Vanuatu should check these. I had a look at Lynch's (2000) grammar of Anejom and the phonemes in the Phoible list don't seem to 100% correspond to those in his grammar either, for example Phoible has "cç" which doesn't correspond to anything on p.13 and Phoible doesn't have long vowels, which Lynch says are phonemic, though "derived from underlying geminate sequences" so this could be why
• For Lenakel, Roro, Wogeo, Kwaio, and Ponapean it also looks like there is a mismatch between orthography and the actual phoneme inventory as recorded by Phoible, like the ABVD entries are in a more phonetic orthography so the profiles are generating more "phonemes" for the inventory than Phoible says exists. Again this can be corrected in the profiles but we need to know how the orthography maps to phonemes
• Then we have a bunch from Hoava to Mafea where there is only a difference of -3 to 0 between the ABVD and Phoible in terms of numbers of phonemes, but some of the IPA symbols used are different, again probably an orthography thing.
• Hawaiian ABVD isn't counting long vowels as separate phonemes

Here is the data from the Phoible/ABVDOceanic comparison script: name	glottocode	total_abvd	total_phoible	delta	tokens_abvd	tokens_phoible
Malo (North)	malo1243	35	21	-14	a b bʷ c d e i j k l m n o p pʰ pʷ r s sʷ t ts tʃ tʰ u w x ŋ ɔ ɛ ɣ β βʷ ⁿb ⁿbʷ ⁿd	a e i k l m mb mbʷ mʷ n nd nɟ o r s t u x ŋ β βʷ
Anejom (Aneityum)	anei1239	37	26	-11	a aː b dʒ dʒʱ e f h hʷ i iː j k kʷ l lʷ m n nʰ nʷ o oː p pː r s t ts tʰ u v w ŋ ɣ ɣʰ ɲ θ	a cç f h j k l m mʷ n p pʷ s t v w ŋ ɔ ɛ ɣ ɪ ɲ ɾ ʊ ʔ θ
Mwotlap	motl1237	32	22	-10	a b d h i j k kʰ l lʷ m mʷ n o p pʰ pʷ s t tʰ tʷ u w ŋ ɔ ɛ ɣ ɣʰ ɪ ɰ ʊ β	a e h i j kpʷ l m mb n nd o s t u v w ŋ ŋmʷ ɣ ɪ ʊ
Manam	mana1295	27	18	-9	a b bʷ d dʒ e g gʷ i j k l m mʷ n o p pʷ r s t u w z ŋ ɪ ʔ	a b d e i k l m n o p s t u z ŋ ɡ ɾ
Marquesan	nort2845	30	22	-8	a aː d e eː f g h hː i iː ì j k kː m n nː o oː p pʰ r t tː u uː v w ʔ	a aː e eː f h i iː k m n o oː p r s t u uː v ç ʔ
Lenakel	lena1238	29	21	-8	a aː e eː h i iː k l lʰ m mʰ mʷ n nʰ o p pʷ r s t u uː v vʰ w ŋ ŋʰ ə	a e̞ f h i k l m mʷ n o̞ p pʷ s t u w ŋ ə ɰ ɾ
Roro	waim1251	22	14	-8	a ã b e h i iː ĩ j k l m n o p r s t u v w ʔ	a b e̞ h i k m n̪ o̞ p t̪ u ɾ̪ ʔ
Wogeo	woge1237	27	20	-7	a b bʷ d dʒ e f g i j k kʷ l m mʷ n o p r s t u v w x ŋ ɲ	a b d̪ e f i j k l̪ m n̪ o r s t̪ u v ŋ ɡ ɲ
Kwaio	kwai1243	26	21	-5	a aː b d e eː f g gʷ i iː k kʷ l m n o oː r s t u uː w ŋ ʔ	a e̞ i l m mb n nd o̞ s t u w x xʷ ŋ ŋɡ ŋɡʷ ŋʷ ɸ ʔ
Ponapean	pohn1238	24	20	-4	a d e eː g h i j k l m mʷ mː n o p pʷ r s t u w ŋ ɛ	a j k l̪ m mʷ n̪ o p pʷ r s t̪ u w ŋ ɔ ɛ ɪ ʈʂ
Hoava	hoav1238	24	21	-3	a b d dʒ e g h i k l m n o p r s t u v z ŋ ɔ ɛ ɣ	a b d d̠ʒ h i k l m n p r s t u ŋ ɔ ɛ ɡ ɣ β
Siar	siar1238	25	22	-3	a b d e f g h i j k l m n nː o p r rʷ s t u v w ŋ ə	b d e e̝ i j k l m n o o̝ p r s t u w ŋ ɑ ɡ ɸ
Bariai	bari1286	20	17	-3	a b d e g i k l m n o p r s t u uː v w ŋ	b d e i k l m n o p r s t u ŋ ɑ ɡ
Sursurunga	surs1246	24	21	-3	a b d e g h i j k l lʰ lʷ m n o p r s t u v w ŋ ə	b d e h i j k l m n o p r s t u w ŋ ɐ ə ɡ
Dobu	dobu1241	27	24	-3	a b bʷ d e g gʷ i j k kʷ l m mʷ n o oː p pʷ r s t u uː w ʔ ʔʷ	a b bʷ d i j k kʷ l m mʷ n p pʷ s t u w ɔ ɛ ɡ ɡʷ ʔ ʔʷ
Poai	fwai1237	38	35	-3	a aː c dʒ e ẽ f g i k l m n o oː p r sʰ ts t̪ u v w ø ŋ ɛ ɣʷ ɲ ʔ ʝ ʰk ʰm ʰn ʰp ʰv ʰʝ ⁿb ⁿd	a aɨ ă e f h i ĭ j k kʰ l̪|l m n̪|n ŏ p pʰ t̠ʃ t̠ʃʰ t̪|t t̪ʰ|tʰ u ŭ v ŋ ɔ ɔ̆ ə ɛ ɛ̆ ɨ ɨ̆ ɬ̪|ɬ ʃ ʔ
Loniu	loni1238	25	23	-2	a c d e h i j k l m mʷ n o p pʷ r s t ts u w ŋ ɪ ɲ ʔ	a e h i j k l m mʷ n o p pʷ r s t t̠ʃ u w ŋ ɔ ɛ ɲ
Tongan	tong1325	19	17	-2	a aː e eː f h i k l m n o p s t u v ŋ ʔ	a e f h i k m n̪ o p s t̪ u v ŋ ɺ ʔ
Maori	maor1246	21	20	-1	a aː e eː g h i iː k m n o oː p r t u uː w wʰ wʱ	f h i iː k m n oː o̞ p t u uː w ŋ ɑ ɑː ɛ ɛː ɾ
Mafea	mafe1237	25	24	-1	a e f h i k l m mʰ mʰː n nʰ o p r s t u v vʱ w ŋ ʈ ʔ β	e i k l lː m mː m̼ n nː o p p̼ r rː s t u v v̼ ŋ ɑ ɖ ɰ

Just checking some other ones, more soon