A principled limit on abstractness?

In connection with our first neutralization rule, final devoicing in Russian, we explained the alternation [porok] ‘threshold (nom sg)’ ~ [poroga] ‘threshold (gen sg)’ by saying that underlyingly the stem ends with /g/. The abstract representation /porog/ for [porok] ‘threshold (nom sg)’ is justified by the fact that [porok] and [poroga] have the same root morpheme, and /porog/ is one of the two actually occurring pronunciations of the morpheme. In hypothesizing underlying forms of morphemes, we have repeatedly emphasized the utility of considering any and all of the surface realizations of a given morpheme as candidates for being the underlying form. One might even advance a formal principle regarding abstractness (a principle to this effect was proposed in the theory of Natural Generative Phonology; see Vennemann 1974):

(1) The underlying form of a morpheme must actually be pronounced as such in some surface form containing the morpheme.

The underlying cognitive presupposition of such a principle is that humans only abstract the nature of morphemes by directly selecting from tokens of perceptual experience with that unit.

When you look at a broad range of phonological analyses, it very often turns out that the supposed underlying form of a morpheme is indeed directly observed in some surface form. Nonetheless, such a principle cannot be an absolute condition on the relation between underlying and surface forms, that is, it cannot be a principle in the theory of grammar. Recall that in Palauan, all unstressed vowels become schwa, and underlying forms of roots may contain two full vowels, for example /daŋob/ ‘cover,’ /teʔib/ ‘pull out,’ /ŋetom/ ‘lick.’ We are justified in concluding that the first vowel in /daŋob/ is /a/ because it is actually pronounced as such in [mə-ˈdaŋəb] when the first root vowel is stressed, and we are justified in concluding that the second vowel is /o/ because that is how it is pronounced in [dəˈŋobl]. Although each hypothesized underlying vowel can be pronounced in one surface variant of the root or another, no single surface form actually contains both vowels in their unreduced form: the hypothesized underlying form /daŋob/ is never pronounced as such, thus our analysis of Palauan is a counterexample to the excessively restrictive statement (1).

Similar examples come from English (cf. the underlying stem /tεlεgræf/, which explains the surface vowel qualities in [ˈtɛləgræf ] and [təˈlɛgrəf-ij]) and Tonkawa (cf. /picena/, which is justified based on the surface forms picna-n-oʔ and we-pcen-oʔ). Condition (1) also runs into problems in Yawelmani, which has a rule shortening a long vowel before a cluster of two consonants, and another rule inserting i after the first of three consonants. The two rules apply in stems such as /ʔa:ml/, so that epenthesis turns /ʔa:ml-hin/ into [ʔa:mil-him], and shortening turns /ʔa:ml-al/ into [ʔamlal]. The problem for (1) is that /ʔa:ml/ can never be pronounced as such, since either the vowel is shortened, or else i is inserted.

Rather than abandon the enterprise of doing phonology in these languages out of misguided allegiance to an a priori assumption about the relationship between underlying and surface forms, we might consider a weaker constraint, which allows underlying forms of morphemes to be composed of segments that are actually pronounced in some attestation of the morpheme, but disallows representations that are more abstract.

(2) The underlying form of a word must contain only segments actually pronounced as such in some related word containing the morpheme.

Even this cannot be an absolute requirement. One case that runs afoul of this condition is the case of stem-final voiced stops in Catalan. There is a rule devoicing final obstruents, and another rule spirantizing intervocalic voiced stops. These rules result in alternations such as sεk ‘dry (masc)’ ~ sεkə ‘dry (fem)’ from /sεk/, versus sek ‘blind (masc)’ ~ seγə ‘blind (fem)’ from /seg/. The underlying voiced stop /g/ is not directly attested in any form of the stem /seg/, and thus runs afoul of constraint (2).

Another counterexample to (2) is Hehe. That language has a rule assigning H tone to a penultimate vowel that is not also immediately preceded by an H. This rule accounts for the position of the second H tone in words like kú-kam-íl-a ‘to milk for,’ kú-kam-il-án-a ‘to milk for each other,’ and the lack of H tone in kú-kam-a ‘to milk’ where the penultimate vowel is preceded by an H-toned vowel. Surface forms such as kú-kam-y-á ‘to cause to milk’ and kú-kam-w-á ‘to be milked’ would seem to be exceptions, but actually they follow the general pattern perfectly, as long as we recognize that the underlying forms are /kú-kam-i-a/ and /kú-kam-u-a/. Given those underlying forms, the H is regularly assigned to the penultimate vowel giving kú-kam-í-a and kú-kam-ú-a, and then the high vowels become glides before a vowel, causing the H tone to be transferred to the final vowel. The important point about these examples is that the assumed vowels of the causative and passive never surface as vowels: they appear only as glides, since by quirks of Hehe morphology, the morphemes -i- and -u- are always followed by a vowel suffix, so they always undergo glide formation.