Action, not ritual

It is achingly apparent that an overwhelming amount of research in speech and language technologies considers exactly one human language: English. This is done so unthinkingly that some researchers seem to see the use of English data (and only English) as obvious, so obvious as to require no comment. This is unfortunate in part because English is, typologically speaking, a bit of an outlier. For instance, it has uncommonly impoverished inflectional morphology, a particularly rigid word order, and rather large vowel inventory. It is not hard to imagine how lessons learned designing for—or evaluating on—English data might not generalize to the rest of the world’s languages. In an influential paper, Bender (2009) encourages researchers to be more explicit about the languages studied, and this, framed as an imperative, is has come to be called the Bender Rule.

This “rule”, and the aforementioned observations underlying it, have taken on an almost mythical interpretation. They can easily be seen as a ritual granting the authors a dispensation to continue their monolingual English research. But this is a mistake. English hegemony is not merely bad science, nor is it a mere scientific inconvenience—a threat to validity.

It is no accident of history that the scientific world is in some sense an English colony. Perhaps you live in a country that owes an enormous debt to a foreign bank, and the bankers are demanding cuts to social services or reduction of tariffs: then there’s an excellent chance the bankers’ first language is English and that your first language is something else. Or maybe, fleeing the chaos of austerity and intervention, you find yourself and your children in cages in a foreign land: chances are you in Yankee hands. And, it is no accident that the first large-scale treebank is a corpus of English rather than of Delaware or Nahuatl or Powhatan or even Spanish, nor that the entire boondoggle was paid for by the largest military apparatus the world has ever known.

Such material facts respond to just one thing: concrete actions. Rituals, indulgences, or dispensations will not do. We must not confuse the act of perceiving and naming the hegemon with the far more challenging act of actually combating it. It is tempting to see the material conditions dualistically, as a sin we can never fully cleanse ourselves of. But they are the past and a more equitable world is only to be found in the future, a future of our own creation. It is imperative that we—as a community of scientists—take steps to build the future we want.

References

Bender, Emily M. 2009. Linguistically naïve != language independent: why NLP needs linguistic typology. In EACL Workshop on the Interaction Between Linguistics and Computational Linguistics, pages 26-32.

Is formal phonology in trouble?

I recently attended the 50th meeting of the North East Linguistics Society (NELS), which is not much of a society as a prestigious generative linguistics conference. In recognition of the golden jubilee, Paul Kiparsky gave a keynote in which he managed to reconstruct nearly all of the NELS 1 schedule, complete with at least one handout, from a talk by Anthony Kroch and Howard Lasnik. Back then, apparently, handouts were just examples: no prose.

In his talk, Paul showed a graph showing that phonology accounts for an increasingly small number of paper at NELS, and in fact the gap has actually gotten worse over the last few decades. Paul proposed something of an explanation: that the introduction of Optimality Theory (OT) and its rejection of “derivational” explanations has forever introduced a schism between phonology and other subareas, and that syntacticians and semanticists are simply uncomfortable with the non-derivational nature of modern phonological theorizing.

With all due respect, I do not find this explanation probable. As he admits, most OT theorizing (including his own) now actually rejects the earlier rejection of derivational explanations. And on the other hand, modern syntactic theories are a heady brew of derivational (phases, copy theory, etc.) and non-derivational (move α, uninterpretable feature matching, etc.) thinking. And finally it’s not really clear why the aesthetic preferences of syntacticians (if that’s all they are) should produce the data, i.e., fewer phonology papers at NELS.

But I do agree that OT is the elephant in the room, responsible for an enormous amount of fragmentation in phonological theorizing.

I would liken Prince & Smolensky’s “founding document” (1993) to Martin Luther’s Ninety-five Theses. Scholars believe that Luther wished to start a scholarly theological debate rather than a popular revolution, and I suspect the founders of OT were similarly surprised with the enormous impact their proposal had on the field. Luther’s magnificient heresy may have failed to move the Church in the directions he wished, but he is the father of hundreds if not thousands of Protestant sects, each with their own new and vibrant “heresies”. The founders of OT, I think, are similarly unable to put the cat back into the bag (if they wish to at all).

In my opinion, OT’s early rejection of derivationalism has been an enormous empirical failure, and the full-blown functionalistic-externalist thinking—one of the first post-OT heresies (let’s liken it to Calvinism)—is, in my opinion, ontologically incoherent. That said, I would encourage OT believers to try more theory-comparison. The article on “Christian denominations” in Diderot’s & d’Alembert’s Encyclopédie begins with the obviously insincere suggestion that someone ought to study which of the various Protestant sects is most likely to lead to salvation. But I would sincerely love to find out which variant of OT is in fact most optimal.

[Thanks to Charles Reiss for discussion.]

Should Noam Chomsky retire?

Somebody said he should. I don’t want to put them on blast. I don’t know who they are, really. Their bio says they’re faculty at a public university in the States, so they probably know how things go around here about as well as me. Why should he retire? They suggested that were he to retire his position at the University of Arizona, that it would open up a tenure line for “ECRs”.¹

Let me begin by saying I do not have a particularly strong emotional connection to Noam. Like many linguists, my academic family tree has many roots at MIT, where Noam taught until quite recently. I have met him in person once or twice, and I found him polite and unassuming. This is a surprise to me. The Times once wrote that Noam is “arguably the most important intellectual alive today”, and important people are mostly assholes.

But I do have very strong intellectual commitments to Noam’s ideas. I think that the first chapter of his Aspects of the Theory of Syntax (1965) is the best statement of the problem of language acquisition. I believe that those who have taken issue with the Aspects idealization of the “ideal speaker-listener” betray a profound ignorance of the role that idealizations play in the history of science.

I think The Sound Pattern of English (SPE), which Noam cowrote with Morris Halle, is the most important work in the theory of phonology and morphology. I believe that the critics who took issue with the “abstract” and “decompositional” nature of SPE have largely been proven wrong.

I even admire the so-called “minimalist program” for syntactic theory Noam has outlined since the 1990s.

It is impossible to deny Noam’s influence on linguistics and cognitive science. We who study language are all pro- or anti-Chomskyians, for better or worse. (And I have much more respect for the “true haters” than the reflexive anti-Chomskyians.) I don’t think Noam should apologize for his critiques of “usage-based” linguistics. I don’t think Noam can fairly be called an “arm-chair” theorist. I think generative grammar has made untold contributions to even areas like language documentation and sociolinguistics, which might seem to be excluded by a strict reading of Aspects.

And, I admire Noam’s outspoken critique of US imperialism. While Noam may have some critics from the left, his detractors (including many scientists of language!) are loud defenders of the West’s blood-soaked imperial adventures.

As a colleague said: “I like Noam Chomsky. I think his theories are interesting, and he seems like a decent guy.” He is a great example of what one can, and ought to, do with tenure.

None of this really matters, though. I do not think he “deserves” a job any more than any other academic does. So, could Noam clear up a “tenure line” simply by retiring? The answer is probably not. Please allow me an anecdote, one that will be familiar to many of you. I teach in a rather-large and robust graduate linguistics program at a publicly-funded college in one of the richest cities in the world (“at the end of history”). Two of our senior faculty are retiring this year, and as of yet the administration has not approved our request to begin a search for a replacement for either of them. Declining to replace tenure lines after retirement is one of the primary mechanisms of casualization in the academy.

Even if you disagree with my assessment of Noam’s legacy, the availability of tenure is not directly conditioned on retirements (though perhaps it should be). Noam bears no moral burden for simply not retiring. If you’d like to fight back against casualization of labor, take the fight to the administration (and to the state houses who set the budgets), don’t blame senior faculty for simply continuing to exist in the system.

PS: If you enjoyed this, you should read The Responsibility of Intellectuals.

¹: I had to look up this acronym. It stands for “early-career researchers”, though I’m not quite sure when one’s “early career” starts or ends. I find that an unfortunate ambiguity.

A first impression of Rome

Rome is a body, freshly exhumed, straddled on all sides by leering police.

Rome is dominated by ruins: the forum is a boneyard, the colosseum a headstone, the Aurelian walls fence in the hallowed ground. Few visible works predate the emperor Trajan, who lived to see Rome reach its territorial maximum, saw it begin its long decline. The archaelogical mode in Rome is to unearth, to lay out before us whatever can be found, as it was found. Reconstruction is mostly confined to pathetic medieval attempts to extract from the pagan monumenta a bit of glory for Christ. There are the shrines—and later a church—built on the floor of the colosseum, at a time when the inner corridors were turned into barns for sheep—and temples rededicated as shrines or churches—though these lack their fine marble facings and bronze ceilings they had in an earlier era.

The ongoing excavation of the Domus Aurea is the most striking Roman corpse. Nero was not so much a madman as the first disaster capitalist. After the great fire of 64 C—a fire that may have been started by the emperor’s confederates—Nero seized a full third of Rome for a pleasure palace, a ‘golden hall’. Years later, Trajan, by comparison a great liberalizer and homo populi, razed Nero’s colossal artificial lake and set to construct an amphitheatre that later became the colosseum. As for the halls of the Domus, Trajan stripped them of their marble facings and filled them with rubble, and they became the foundations, and the sewers, for a great public bath. The only hint of their one-time splendor are the fine frescoes that can be seen on weekend tours offered while the excavators rest.

Standing over the ancient Roman corpse are a huge mass of police, of which there are both far too many and far too many types. There is the esercito—the army—of whom it is said, “at least they are competent”. There are carabinieri, who use the iconography of musketeers but whose portfolio includes roles played by state troopers and sheriffs, the DEA and the FBI, the Pinkertons and second-world paramilitaries. And there are also polizia and guardia. As far as one can tell, small bands of the various armed forces have been camping at their chosen corners for years, doing little more than smoking cigarettes and talking amongst themselves. The duplication of effort is exquisite, and not without a bit of apparent dispute over turf. One will not uncommonly find an entrance to a church guarded over by the esercito and the exit by polizia. The effect would be chilling were the patient not so long dead, so long in the ground.

Latin vowel-glide alternations

Post-war structuralist phonology greatly emphasized phonemics and largely ignored morphophonemics. But in 1959, Morris Halle’s Sound Pattern of Russian argued that the distinction between allophony and alternation has little cognitive importance, and in fact the distinction leads to an unnecessary duplication of effort. As a result of Halle’s forceful arguments, the contrast between phonemic and morphophonemic processes plays little role in modern phonological theory. I would like to go one step further and suggest that patterns of alternation are actually more principled facts than those of allophony. Simply put, a speaker must command the pattern of alternation in their language; but it is not at all clear whether they exploit allophony when constructing their lexical entries. This is highlighted most clearly by the notions of lexicon optimization, Stampean occultation, and richness of the base in Optimality Theory, though as Hale et al. (1998) note, similar points apply to rule-based theories.

In writing the Romans did not draw distinctions between the high monophthongs [i, u, iː, uː] and glides [j, w], respectively. This naturally led structuralist linguists (e.g., Hall 1946) to suggest that the glides are allophones of the high monophthongs. There are some apparent problems with this suggestion, though not all of them are fatal. One point that has largely been ignored in this discussion is that Classical Latin has at least four types of plausible alternations between high monophthongs and the corresponding glides. In this squib I review these alternations.

Deverbal -u- derivatives

There are a large number of adjectival derivatives formed from verbal stems by the addition of -u- and the appropriate agreement suffixes, e.g., masculine nominative singular (masc. nom.sg.) -u-us, feminine nom.sg. -u-a, and neuter nom.sg. -u-um, and so on. These derivatives have a similar semantics to past participles (“having been Xed”) but in some cases have a secondary meaning “able to be Xed”. For example, the masc. nom.sg. form dīuiduus [diːwi.du.us] means ‘divided’ (cf. dīuidō [diːwi.doː] ‘I divide’) but also ‘divisible’. This is a fairly productive process, as the following examples show. (I have taken the liberty of leaving off certain further productive derivatives, such as intensified adjectives in per-.)

(1) assiduus ‘constant, ambiguus ‘hither and thither’, annuus ‘annual, arduus ‘elevated’, cernuus ‘bowed forward’, circumfluus ‘flowing around’ (refluus ‘ebbing’), cōnspicuus ‘visible’, contiguus ‘neighboring’, continuus ‘continuous’, dīuiduus ‘divided; divisible’ (indīuiduus ‘undivided; indivisible’), exiguus ‘strict’, fatuus ‘foolish’, incaeduus ‘uncut’, ingenuus ‘indigenous’, irriguus ‘irrigated’, mēnstruus ‘monthly’, mortuus ‘dead’ (dēmortuus ‘departed’, intermortuus ‘decayed’, praemortuus ‘prematurely dead’), mūtuus ‘borrowed’ (prōmūtuus ‘paid in advance’), nocuus ‘harmful’ (innocuus ‘harmless’), occiduus ‘westerly’, pāscuus ‘for pasturing’, perpetuus ‘perpetual’, perspicuus ‘transparent’, praecipuus ‘particular’, prōmiscuus ‘indiscriminate’, residuus ‘remaining’, riguus ‘irrigated’, strēnuus ‘brisk’, succiduus ‘sinking’, superuacuus ‘superfluous’, uacuus ’empty’, uiduus ‘destitute’

In all the above cases …uus is read [u.us]. However, when the stem ends in a liquid [l, r] …uus is read [wus], indicating that the deadjectival affix is realized as [w].

(2)
a. caluus ‘bald’, fuluus ‘reddish-yellow, tawny’, giluus ‘pale yellow’, heluus ‘honey yellow’
b. aruus ‘arable’, curuus ‘bent’ (incuruus ‘bent’), furuus ‘dark, swarthy’, paruus ‘small’, prōteruus ‘violent’, toruus ‘savage’

It is interesting to note that the contexts where -u- is realized as [w] align with a well-known allophonic generalization (Devine & Stephens 1977: 61., 134f.): a u preceded by a (tautomorphemic) coda liquid or front glide, and followed by a vowel, is realized as [w], as in silua [sil.wa] ‘forest’ or ceruus [ker.wus] ‘deer’, but is realized as a vowel when the preceding consonant is either a nasal, an obstruent, or part of a consonant cluster, as in lituus [li.tu.us] ‘trumpet’ or patruus [pa.tru.us] ‘paternal uncle’.

Two residual issues remain. First, when the verbal stem end in qu [kw], the adjectival derivative is spelled …quus. By the normal rules of spelling this would be read as [kwus], which would suggest that a zero allomorph of the adjectival suffix is selected for here.

(3) aequus ‘equal’, antīquus ‘old’, fallāciloquus ‘falsely speaking’ (fātiloquus ‘prophetic’, flexiloquus ‘ambiguous’, grandiloquus ‘grandiloquent’, magniloquus ‘boastful’, uāniloquus ‘lying’, uersūtiloquus ‘slyly speaking’), inīquus ‘unjust’, longinquus ‘distant’, oblīquus ‘slanting, oblique’, pedisequus ‘following on foot’, propinquus ‘near’, reliquus ‘remaining’

This is consistent with the metrical evidence. For instance in the following verse, aequus must be read as bisyllabic.

Secondly, there are a number of deverbal derivatives in -u-us where the verb form also has a stem-final [w]. In this case we also observe [wus].

(5)
a. cauus [ka.wus] ‘hollowed; hollow’ (concauus ‘hollow’); cf. cauō [ka.woː] ‘I excavate’
b. flāuus [flaː.wus] ‘yellow, gold, blonde’ (sufflāuus ‘yellowish’); cf. flāueō [flaː.we.oː] ‘I am yellow’
c. (g)nāuus [naː.wus] ‘active’ (īgnāuus ‘lazy’); cf. nāuō [naː.woː] ‘I do s.t. enthusiastically’
d. nouus [no.wus] ‘new’; cf. nouō [no.woː] ‘I renew’
e. saluus [sal.wus] ‘safe; well’; cf. salueō [sal.we.oː] ‘I am well’
f. uīuus [wːi.wus] ‘living’ (rediuīuus ‘restored to life’); cf. uīuō [wiː.woː] ‘I live’

This may be another context where the adjectival suffix has a zero allomorph, though it is not clear whether we are looking at the same derivational process as above.

The foregoing discussion leads me to posit a deverbal adjective-forming suffix /-u-/ with two phonologically-predictable allomorphs: [w] before liquids, and zero before [kw] and possibly, [w].

The “third stem”

Schoolchildren learning Latin memorize four forms (or principal parts) of each verb: the first person singular (1sg.) present active indicative (e.g., amō ‘I love’), the present infinitive (amāre ‘to love’), the 1sg. perfect active (amāvī ‘I loved’), and the perfect passive participle (amātus masc. nom.sg. ‘loved). The first two principal parts effectively index the so-called “present stem” of the verb, and the third principal part gives the so-called “perfect stem”. The relationship between the present and perfect stem is often unpredictable. Some perfect stems lengthen a monophthong in the final syllable of the present stem (e.g., legō/lēgī ‘I choose/chose’); some perfect stems omit a post-vocalic nasal in the final syllablem with comcomitant lengthening (uincō/uīcī ‘I win/won’); some are mutated by the addition of a -s- perfect suffix (dīcō/dīxī [diː.koː, diːk.siː] ‘I say/said’); others bear a CV-reduplication prefix, and so on. This has lead some to suggest that the latter two stems are essentially “listed” or “stored” for all verbs. This is, for instance, the position of Lieber (1980:141f., 152f.), but has been disputed by Aronoff (1994: chap. 2) and Steriade (2012), among others, who claim there are many productive regularities in both cases.

The majority of verbs have perfects that consist of the bare verb root, the theme vowel, a high back vocoid perfect suffix, and the appropriate person-number agreement suffixes (e.g., 1sg. -ī-). The perfect suffix is preceded by a theme vowel and as the appropriate agreement suffixes are all vowel-initial, it is always intervocalic. Allophonically, this is a context where [u] is never found but [w] is, and this is what we find here: amāuī [a.maː.wiː] ‘I loved’. This type of perfect is in fact found in all conjugations, and found in the overwhelming majority of 1st (-ā- theme vowel) and 4th conjugation (-ī-) verbs (Aronoff 1994:43f.).

(6)
a. cōnsōlāuī [kon.soː.laː.wiː], portāuī [por.taː.wiː] ‘I carried’
b. dēlēuī [deː.leː.wiː] ‘I destroyed’, plēuī [pleː.wiː] ‘I filled up’
c. cupīuī [ku.piː.wiː] ‘I desired’, petīuī [pe.tiː.wiː] ‘I sought’
d. audīuī [aw.diː.wiː] ‘I listened to’, mūnīuī [muː.niː.wiː] ‘I fortified’

However, there is an alternative formulation in which the theme vowel is omitted, placing the perfect suffix to the right of a consonant, and in this context it is instead realized as [u]. This type of perfect is also found in all conjugations but is most common in the 2nd (-ē-) conjugation.

(7)
a. domuī [do.mu.iː] ‘I tamed’, uetuī [we.tu.iː] ‘I forbid’
b. docuī [do.ku.iː] ‘I taught’, tenuī [te.nu.iː] ‘I held’
c. rapuī [rap.u.iː] ‘I snatched’, texuī [tek.su.iː] ‘I wove’
d. aperuī [a.pe.ru.iː] ‘I opened’, saluī [sa.lu.iː] ‘I leapt’

Together the patterns in (6-7) account for the vast majority of perfects in all conjugations except the 3rd (itself a grab-bag of etymologically dissimilar verbs).

I propose that the default perfect suffix is /-u-/ and that it undergoes glide formation to [w] in (6), in intervocalic position, a generalization consistent with the allophonic facts. In (7), when adjacent to the verb root, glide formation is blocked. However, the examples in (7) cannot take a “free ride” on any allophonic generalization. As can be seen in (7d), the perfect suffix does not form [l.w, r.w] syllable contact clusters, unlike the adjectival suffix in (5). There is a surfeit of possible analyses for the failure of glide formation in this context: it might be an effect specific to the perfect suffix or to the category of verb, or the result of cyclicity or phase-based spellout. We leave the question open for now.

The “fourth stem”

The form of the perfect passive participle, the fourth principal part, similarly problematic. For many verbs, the perfect passive participle is formed by adding to the verb root a -t- suffix and the appropriate agreement suffixes (e.g., in citation form, the masc. nom.sg. -us), once again sometimes accompanied by lengthening of the stem-final vowel and/or leftward voice assimilation (an exception-less rule of Latin) triggered by the -t- as in (8b).

(8)
a. docuī [do.ku.iː] ‘I teach’, doctus [dok.tus] masc. nom.sg ‘taught’
b. tegō [te.goː] ‘I clothe’, tēctus [tek.tus] masc. nom.sg. ‘clothed’

Two verb roots which end in consonant followed by a high back vocoid and form a -t- perfect passive participle: soluō [solwoː] ‘I loosen; I explain’ and uoluō [wolwoː] ‘I roll’. This places the root-final high back vocoid, by hypothesis /u/, between two consonants, a context where glides are forbidden. The result is solūtus [soluːtus] and uolūtus [woluːtus]. However, it should be noted that this particular pattern is limited to these two verbs and their derivatives, and that the long ū is unexpected unless it reflects stem vowel lengthening (cf. tēctus above).

Synizesis and diaeresis

Latin poetry exhibits variation in glide formation. (The following examples are all drawn from Lehmann 2005). Synizesis, the unexpected overapplication of glide formation in response to the meter, can be seen in the following verse.

In this verse, tenuis ‘thin’ occurs initially, which requires that the first syllable be heavy. The only way to accomplish this is to read it as the bisyllabic [ten.wis] rather than the expected trisyllabic [te.nu.is]. Similarly, in another verse (Verg., Aen. 8.599), abiēte, the ablative singular of abiēs ‘silver fir’, must be read as trisyllabic [ab.jeː.te] rather than the expected [ab.i.eː.te].

On the other hand, the poets also make use of diaeresis, or apparent underapplication of glide formation. For example, siluae, the genitive singular of silua ‘forest’, is in one verse (Hor., Carm. 1.23.4) read as trisyllabic [si.lu.aj] rather than as the expected bisyllabic [sil.waj]. The conditions governing synizesis and diaeresis are not yet well understood, but they constitute further evidence for the close grammatical relationship between [i ~ j] and [u ~ w] in Classical Latin.

Conclusion

We have seen four ways in which the Latin high vocoids alternate between vowels and glides. Together, these four patterns provide indirect evidence for the hypothesis that Latin glides are allophones of the corresponding high vowels, though there are some minor dissociations between patterns of allophony and alternations.

[Earlier writing about Latin glides: Latin glides and the case of “belua”]

References

Aronoff, Mark. 1994. Morphology by itself: stems and inflectional classes. Cambridge: MIT Press.
Devine, Andrew M., and Stephens, Laurence D. 1977. Two studies in Latin phonology. Saratoga: Anma Libri.
Hall, Robert A. 1946. Classical Latin noun inflection. Classical Philology 41(2): 84-90.
Hale, Mark and Kissock, Madelyn, and Reiss, Charles. 1998. Output-output correspondence in Optimality Theory. In Proceedings of WCCFL, pages 223-236.
Halle, Morris. 1959. The sound pattern of Russian. The Hague: Mouton.
Lehmann, Christian. 2005. La structure de la syllabe latine. In Touratier, Christian (ed.), Essais de phonologie latine, pages 157-206. Aix-en-Provence: Publications de l’Université de Provence.
Lieber, Rochelle. 1980. On the organization of the lexicon. Doctoral dissertation, MIT.
Steriade, Donca. 2012. The cycle without containment: Latin perfect stems. Ms., MIT.

Latin glides and the case of “belua”

Latin texts leave the distinction between high monophthongs [i, u, ī, ū] and glides [j, w] unspecified. This has lead some to suggest that the glides are allophones of the monophthongs. For instance, Steriade (1984) implies that the syllabicity of [+high, +vocalic] segments in Latin is largely predictable. Steriade points out two contexts where high vocoids are (almost) always glides: initially before a vowel (# __ V) and intervocalically (V __ V). In these two contexts, the only complications I am aware of arise from competition between generalizations. For instance, in ūua [uː.wa] ‘grape’ and ūuidus [uː.wi.dus] ‘damp’, intervocalic glide formation appears to bleed word-initial glide formation. (Or it could be the case that ū is ineligible for glide formation by virtue of its length.) And the behavior of two adjacent high vocoids flanked by vowels is somewhat idiosyncratic: compare naevus [naj.wus] ‘birthmark’ and saeuiō [saj.wi.oː] ‘I am furious’, where (by hypothesis) /ViuV/ surfaces as [j.w], to dēuius [deː.wi.us] ‘devious’ and pauiō [pa.wi.oː] ‘I beat’, where (by hypothesis) /VuiV/ surfaces as [.wi] but never as *[w.j]. And so on.

However, Cser (2012) claims that syllabicity of high vocoids is not at all predictable after a consonant and before a vowel, i.e., in the context C __ V. Here we usually observe [w] when the preceding consonant is coda [j, l, r], as in the aforementioned naevus or silua [sil.wa] ‘forest’. Cser contrasts this latter form with belua ‘wild beast’, which is trisyllabic rather than bisyllabic. However, it is not clear this is a good near-minimal pair. The word was clearly not pronounced as [be.lu.a] because the first syllable scans heavy. In the following hexameter verse, the word comprises the fifth foot, a dactyl:

et centumgeminus Briareus, ac belua Lernae (Verg., Aen. 6.287)

Lewis & Short and the Oxford Latin Dictionary both give this word as bēlua [beː.lu.a]. However, it seems much more likely that the word is in fact bellua [bel.lu.a], as it was sometimes written. (Note also that tautomorphemic geminate ll is robustly attested in Latin.) In this case we would expect glide formation to be blocked because the [lw] complex onset is totally unattested, just as Cser predicts from general principles of sonority sequencing. Thus the above verse is:

As Cser notes, many of the remaining near-minimal pairs occur at morphological boundaries⁠—and thus look to someone with my theoretical commitments as evidence for the phonological cycle—or relate to the complex onsets qu [kw] and su [sw], which might be treated as contour segments underlyingly. But much work will be needed to show that these apparent exceptions follow from the grammar of Latin.

References

Cser, András. 2012. The role of sonority in the phonology of Latin. In Parker, Steve (ed.), The sonority controversy, pages 39-64. Berlin: Mouton de Gruyter.
Steriade, Donca. 1984. Glides and vowels in Romanian. In Proceedings of the Berkeley Lingusitics Society, pages 47-64.

Exceptions to reduplication in Kinande

Mutaka & Hyman’s (1990) study of reduplication in Kinande, a Bantu language spoken in “Eastern Zaire” (now the Democratic Republic of the Congo), is the sort of phonology study one doesn’t see much of anymore. The authors begin by noting the recent interest in reduplication phenomena, but note that most of the major work has completely ignored Bantu, an enormous language family in which nearly every language has one or more type of reduplication. Mutaka & Hyman (MH) proceed to describe Kindande reduplication in detail, with only occasional reference to other languages.

Nouns that undergo reduplication have the semantics of roughly ‘the real X’. Most Kinande verbs also undergo reduplication, with the semantics of roughly ‘to hurriedly X’ or ‘to repetitively X’. Verbal reduplication is somewhat more interesting because certain other verbal suffixes (or “extensions”, as they’re sometimes called in Bantu) may also be found in the reduplicant, argued to be a roughly-bisyllabic prefix. For instance, the passive suffix is argued to be underlyingly /u/ but surfaces as [w], and is copied over in reduplication. Thus for the verb hum ‘beat’ the passive e-ri-hum-w-a ‘to be beaten’ reduplicates as erihumwahumwa. However, larger vowel-consonant verbal suffixes are not copied; the applied (-ir-) passive infinitive e-ri-hum-ir-w-a ‘to be beaten for’ has a reduplicated form erihumahumirwa, and for the verb tum ‘send’ the applied passive reciprocal (-an-) infinitive e-rí-tum-ir-an-w-a ‘to be sent to each other’ has a reduplicated form erítumatumiranwa (MH, 56).

What’s even more interesting to me is the behavior of verb stems with what MH call ‘unproductive’ extensions (all of which appear to be vowel-consonant). MH report that for only a small minority of these verb stems is there any plausible etymological relationship to a verb without the extension. One example is luh-uk-a ‘take a rest’ which is plausibly related to luh-a ‘be tired’ (MH, 73e), but there is no *bát-a paired with bát-uk-a ‘move’ (MH, 74d). Verb stems bearing unproductive suffixes may have one of three behaviors with respect to reduplication. For some such stems, reduplication is forbidden: eríbugula ‘to find’. For others, reduplication occurs but the ‘unproductive’ extension is stranded (the same behavior as the ‘productive’ extensions): e-rí-banguk-a ‘to jump about’ reduplicates as eríbangabanguka. Finally, some such stems (roughly half) unexpectedly build a trisyllabic (rather than bisyllabic) reduplicant consisting of the verb root and the unproductive extension: e-ri-hurut-a ‘to snore’ reduplicates as erihurutahuruta (MH, 75). This entire distribution poses a fascinating puzzle. How is the failure of reduplication encoded in the first case? What licenses the trisyllabic reduplicant in the last case?

References

Mutaka, Ngessimo and Hyman, Larry M. 1990. Syllables and morpheme integrity in Kinande reduplication. Phonology 7: 73-119.

Libfix report for June 2019

You may be familiar with fatberg, a mass of non-biodegradable solids and fats found in sewers, which suggests -berg has been innovated (presumably via iceberg). And now London is also haunted by a concreteberg.

Late great tech unicorn Theranos made use of a proprietary blood-collection device they called the nanotainer (via container), and I recently found out about vacutainer and a security software package called Cryptainer. So -tainer has been liberated.

The other day in Queens I saw a sign for a Mathnasium, presumably extracted from gymnasium, and the Corpus of Contemporary American English also has a token of jamnasium (a space for jam seshes), suggesting a nascent -nasium.

In a recent, widely-derided ad campaign, Applebee’s coined sizzletonin on analogy with the neurotransmitter seratonin and the hormone melatonin, but as far as I know that’s the end of the line for -tonin.

The revolution will be at Starbucks

One of the biggest shocks about life in The Zone (11/9/2016-present) is how often Starbucks makes the news. Just a couple days after the election, a group of patriots, organizing around the hashtag #TrumpCup, decided to show solidarity with their big wet boy, subverting the sacred ordering ritual to trick baristas to shout “Iced Frappucino for Trump”. Then, there was the everyday-in-America story of a few young men thrown out—by the police—of a Philadelphia Starbucks for the mere act of being black in public. And now gloriously quixotic former CEO Howard Schultz is considering a third-party run for president. How has Trumpism turned America’s top coffee chain into a battleground?

I think I know. Starbucks is a looking glass, and when we gaze into it, we see what we want to. Allow me to explain.

We yet again live in a time where the public commons is contracting. It is not so much being “enclosed” (as it was in Georgian England) as neglected by the inexorable logic of austerity (as it happens, the key plank of Schultz’s platform). Even public libraries—a radical, and incredibly impactful, experiment in architecture and government—are at risk; President Trump has sought to eliminate federal spending on libraries, and they are under threat both in communities small and large. Faced with disappearing public commons, we turn increasingly to private simulcra of the park, the library, the school or university, and for some, a busy Starbucks will have to do.

Starbucks has another thing going for it. The product is really not bad, and of surprisingly uniform quality. While coffee snobs turn their noses up at the burnt-tasting drip coffee, the espresso drinks are quite good if not always great. The production of a large menu of high-quality, complex, labor-intensive goods, daily, at 14,000 locations across the US, is an incredible feat of logistics. US social welfare programs, increasingly administered by a patchwork of hostile state governments, do not come off well in comparison to the fungible, always-available Starbucks latte. It is easy to see why. Starbucks is embedded in an all-encompassing matrix of market capitalism, but internally, it is a command economy, one in which no store can be left behind. It is hard to even imagine living in an America where say, welfare or health care services are provided to citizens with the same efficiency of Starbucks manager requisitioning a case of oat milk.

At least that’s what I see when I look at Starbucks. But, as #TrumpCup shows, others see something different: the masses of Americans not moved—if not outright repelled—by the mixture of petty grievances and white identity politics that animates President’s Trump’s base. The libs (as we’ll call them) are a diverse group, better defined by exemplars—sometimes, right-wing media caricatures—than prototypes, and one key lib exemplar is the Starbucks barista. The barista is probably young, and possibly urban. Perhaps they have a college education and have taken the job for the health care benefits the state does not provide. Maybe they even share former CEO Schultz’s tepid opposition to President Trump.

If this wasn’t enough to forever code the barista as the Other, there is also a whole new language, not quite English, to learn. A small coffee is unexpectedly “tall”; a large is a “venti”; a “macchiato” is something else entirely. Mastering this language gives the customer the power to summon strange and fantastic beasts: the “blonde espresso”, or if the stars are properly aligned, the “spiced sweet cream nariño 70 cold brew”.

And, perhaps most importantly, the barista is a captive audience. The barista has a manager, and yes, you really can ask to speak to them. For the #TrumpCup Republican, this is a potent brew, a hierarchy in which they stand above the Other, the perfect victim for a bit of everyday cruelty and meaningless self-gratification.

It was probably inevitable that the of the most ubiquitous corporations in American life was going to ultimately come to index something, and where I see the state’s abdication of responsibilities inherent in the social contract, others just see a snot-nosed, underemployed 25-year-old who would rather not be working this job forever. In conclusion, Starbucks is a land of contrasts, and will remain so until we resolve the contradictions inherent in American society.

Using a fixed training-development-test split in sklearn

The scikit-learn machine learning library has good support for various forms of model selection and hyperparameter tuning. For setting regularization hyperparameters, there are model-specific cross-validation tools, and there are also tools for both grid (e.g., exhaustive) hyperparameter tuning with the sklearn.model_selection.GridSearchCV and random hyperparameter tuning (in the sense of Bergstra & Bengio 2012) with sklearn.model_selection.RandomizedSearchCV, respectively. While you could probably could implement these yourself, the sklearn developers have enabled just about every feature you could want, including multiprocessing support.

One apparent limitation of these classes is that, as their names suggest, they are designed for use in a cross-validation setting. In the speech & language technology, however, standard practice is to use a fixed partition of the data into training, development (i.e., validation), and test (i.e., evaluation) sets, and to select hyperparameters which maximize performance on the development set. This is in part an artifact of limited computing resources of the Penn Treebank era and I’ve long suspected it has serious repercussions for model evaluation. But tuning and evaluating with a standard split is faster than cross-validation and can make exact replication much easier. And, there are also some concerns about whether cross-validation is the best way to set hyperparameters anyways. So what can we do?

The GridSearchCV and RandomSearchCV classes take an optional cv keyword argument, which can be, among other things, an object implementing the cross-validation iterator interface. At first I thought I would create an object which allowed me to use a fixed development set for hyperparameter tuning, but then I realized that I could do this with one of the existing iterator classes, namely one called sklearn.model_selection.PredefinedSplit. The constructor for this class takes a single argument test_fold, an array of integers of the same size as the data passed to the fitting method. As the documentation explains “…when using a validation set, set the test_fold to 0 for all samples that are part of the validation set, and to -1 for all other samples.” That we can do. Suppose that we have training data x_train and y_train and development data x_dev and y_dev laid out as NumPy arrays. We then create a training-and-development set like so:

x = numpy.concatenate([x_train, x_dev])
y = numpy.concatenate([y_train, y_dev])

Then, we create the iterator object:

test_fold = numpy.concatenate([
    # The training data.
    numpy.full(-1, x_train.shape[1], dtype=numpy.int8),
    # The development data.
    numpy.zeros(x_dev.shape[1], dtype=numpy.int8)
])
cv = sklearn.model_selection.PredefinedSplit(test_fold)

Finally, we provide cv as a keyword argument to the grid or random search constructor, and then train. For instance, similar to this example we might do something like:

base = sklearn.ensemble.RandomForestClassifier()
grid = {"bootstrap": [True, False], 
        "max_features": [1, 3, 5, 7, 9, 10]}
model = sklearn.model_select.GridSearchCV(base, grid, cv=cv)
model.fit(x, y)

Now just add n_jobs=-1 to the constructor for model and to spread the work across all your logical cores.

References

Bergstra, J., and Bengio, Y. 2012. Random search for hyperparameter optimization. Journal of Machine Learning Research 13: 281-305.