The Reification Problem

Standard phonological methodology would lead us to seek a unified explanation for the vowel/glide realizations in all the contexts above. However, in recent research the analysis is quite commonly narrowed down to the i/j alternation. In this way, the u/w alternation may be disregarded altogether, or may be dismissed as parallel to i/j, the assumption being that anything that is true of i/j will automatically carry over to u/w. Another consequence is that the scope of contexts for i/j alternation is narrowed down to 1. In my view, this is a methodological mistake.

In older research more heed is taken of u/w and other sonorants, for example, Sievers 1878 and Osthoff 1884. When Sievers formulated his generalization regarding vowel/glide alternation in Vedic Sanskrit he included both alternations.

⁵

This much cited quotation is neither comprehensive nor all that well put once you look at it closely, but it is at least clear that Sievers intended his generalization to hold for both vocoids, where their distribution is connected to the preceding syllabic context.

As mentioned above, the i/j alternation is much more extensive than the u/w alternation, in Gothic. Barrack (1998:86) explicitly puts the u/w alternation to the side for this reason, and many scholars do not even mention the u/w alternation. From the point of view of linguistic analysis, this is an opportunity lost, however. Though limited, the u/w alternation is there, unequivocally, and therefore we must ultimately analyze it. It is hard to see how linguistic research methodology could lead to any other conclusion than that u/w alternation must be treated together with i/j alternation (compare the data in 2), and how, in turn, the connection between the data in 1 and 2 could not be made. Furthermore, a simple carryover of conclusions for i/j to u/w will not suffice either, as there are asymmetries between /I/ and /U/ requiring that they not be given completely uniform treatment, as we see in section 4.4.

The disregard of the contexts in 2, as well as other contexts not mentioned above, comes from a reification of Sievers' law in the research tradition. Some researchers have narrowed the scope of morphological contexts studied in Germanic on the basis of the distributions in Vedic Sanskrit. Again, Barrack (1998:87) explicitly delimits his discussion in this way, although he is aware of the phonological overlap. In this way, Sievers' law has come to be understood as a “rule” that could be traced historically in several languages, even though it changes shape every now and then (see, for example, Palviainen 2001). Many other researchers, within and outside of the Neogrammarian tradition, have simply taken over the data set from previous accounts, and so Sievers' law has become the name of an alternation occurring in a subset of the relevant set of data. This is, to my mind, methodologically unsound and linguistically artificial. Consider the following paradigms.

Some assume that a certain paradigm falls under the scope of Sievers' law, for example, for historical reasons (Barrack 1998:86–87). If so, the kuni paradigm is included and the bandi paradigm is not, and then two different explanations will, in principle, be needed for the synchronic i/j alternation in the respective cases.⁶

Some assume that Sievers' law only applies to certain morphological classes (see Palviainen 2001:66), such as nominative and genitive in certain paradigms (and, mutatis mutandi, verbal paradigms). If so, we will need one explanation for the [j] in kunjis, namely Sievers' law, and another for the [j] in kunja. It is hard to see how any of those positions could be reasonably defended, without treating the cases left out. Our duty as linguists is to seek generalizations, as Sievers did, and make them as generally applicable as possible. Of course, morphologically motivated analogy may exempt some contexts from being targeted in the same way as others by the Sievers' law processes, but then those need to be understood, defined, and integrated into the analysis (cf. Kiparsky 2000).⁷

In the following sections, I try to show that the consequences of the unwarranted limits imposed on the research object mentioned above have affected the analysis of Sievers' law as well as the interpretation of orthographic evidence in troubling ways.

Rules of the Game

One rule of the game has already been mentioned, namely Occam's razor, spelled out here in particular (partly overlapping) statements for vowel/glide alternation in Gothic.

The delimitation of the research object to a subset of the relevant data might work if the subset is representative. But if it is not, problems are to be expected. A case in point is the question of underlying phoneme for the vowel/glide alternation. For many, the choice has been between vowel and glide, and as long as one is dealing only with the contexts in 1, the consequences of choosing the one or the other are not very dramatic. It might look arbitrary, but one's rule is just going to either vocalize a glide or glide a vowel, leaving the alternant be in the complementary contexts. However, as soon as one considers a larger data set, including u/w, it becomes clear that one needs conversion both ways.

As shown above, mere inspection of the morphologically more basic forms on the left is not going to help in deciding whether the underlying segment is a vowel or a glide. Also, the form as glide or vowel seems to be dependent to some extent on neighboring segments, one or both of which may be realizations of the other glide/vowel. We may be led to posit the nominative of mawi as reflecting the underlying form /mawi/, but then we would need both a vocalization rule /w/ > [u] and a gliding rule /i/ > [j] to derive the genitive singular form. It might seem unnatural to have different underlying forms for the different vowels/glides, and choosing the one over the other is going to be quite arbitrary (largely as in phonemic theory, compare Moulton 1948). The solution, of course, is to assume that the segments in question are un(der)specified for the feature [±consonantal], and that the phonetic realization of the segment is determined by what position it ends up in in syllable structure. In this way, the desired realizations will not be predetermined by an arbitrary decision on vocalic and consonantal representation.⁸

Gothic Word Division

Orthographic word divisions in Gothic manuscripts clearly follow certain patterns, but the extent to which these are phonologically motivated is not obvious. The most important generalizations are given below.

⁹

I use the terms “line 1” and “line 2” to refer to either side of the word division.¹⁰

¹¹

The decision regarding which pattern is the more regular with respect to Gothic phonology is a matter of interpretation. Presumably, the best analysis will explain why manuscripts vary in this particular regard.

The question now is how we relate these orthographic observations to our understanding of the phonology of Gothic syllabification. To what extent do these observations carry over? Are some observations more “phonological” than others? Several scholars support the strongest claim, namely that word divisions faithfully reflect Gothic syllabifications. For instance, Murray and Vennemann (1983:515) make the following statement:

They then discuss the obstruent+liquid pattern that does not fit the general pattern. It is suggested that that context could be explained either as depending on loan orthography or on morphology. The conclusion at any rate is that the rule above is “the general principle governing Go[thic] syllabification” (Murray and Vennemann 1983:516). This is a really concise description of Gothic orthographic word division, but it cannot be a singular principle of phonology. Rather, it needs to be broken down into components, and in doing so it is possible to see the grouping of obstruent+liquid not as deviant syllabification, but as a phonological regularity, which causes some variation in this particular area (Vennemann 1987). My aim, then, is to try to make clear that some pieces of orthographic information are more reliable than others when it comes to the phonological interpretation.

Implications for Syllabification

First of all, morphology likely plays a role in orthographic word divisions in Gothic (compare Voyles 1992, Barrack 1998), just as in many other systems. This is seen clearly in the treatment of compounds and prefixed forms in 9 above. It is true that the division of compounds like <weina|gardis> could be understood as phonologically motivated by virtue of the reasonable assumption that compound members are syllabified separately, and that they hence constitute prosodic words. Under this view, examples like <inn|ana> and <miþ|ushramidans> in 9 are particularly informative, since the morphemes inn, and perhaps also miþ, are not necessarily their own phonological words. Rather, it seems it is indeed the morphological boundary that is reflected in Gothic word division.¹²

Thus, whether the phonology in normal speech is [in:.ana] or [in.nana] is an open question (until other argumentation is brought to bear), while word division is not.¹³ The influence of morphology in word division generally is widely recognized, but it is usually put to the side once the obvious domains in which it is applicable have been mentioned. This does in no way entail that morphology is not involved in other instances of word division where the divide coincides with lesser morphological boundaries. From a methodological point of view, it is only when we find positive evidence against it that we can rule out the influence of morphology in word division. With suffixes, the role of morphology is less clear; that is obviously part of the problem of interpreting word division as phonological, morphological, or otherwise. The point is made clearly in Vennemann (1987:172), where the Skeireins divisions of <þaþ|ro> and <iupaþ|ro> disrespect what morphology would have dictated (þa|þro and iupa|þro). But on the other hand, there is much to say for morphological influence on word division, too (Barrack 1998:39–40; Riad 1992:86–87).

Second, the well known phonological onset principle (see, for example, Blevins 1995:230ff; Kager 1999:93) is directly motivated in orthographic word division. The onset principle (or the constraint ONSET) requires that syllables begin with a consonant.¹⁴

Examples 10–14 all show this. It seems more important to satisfy this constraint than any other phonological criterion. While in larger consonant clusters there seems to be a prima facie tendency to cluster the consonants on line 1 and to put just one consonant on line 2, it is still the case that when there is only one intervocalic consonant, it goes to line 2; compare 10. Notice in particular that this happens regardless of the length of the preceding vowel: <sei|namma> [si:nam:a], <da|ga> [daga]. The pattern as such is static, however, exhibiting no alternation. In light of this, it is fortunate that the onset principle is so easy to motivate on typological grounds (for example, in processes of stop insertion).

Third, there is a clear tendency to satisfy sonority sequencing, which shows up with the cases where a consonant cluster consists of or ends with an obstruent+liquid sequence; compare 12 and 14. The patterns are somewhat variable here, but if we take first the clusters of three consonants and more, we can see that sonority sequencing of obstruent+liquid plays out against the otherwise general orthographic pattern of maximizing the number of consonants on line 1 (modulo ONSET). The orthographic obstruent+liquid pattern must be motivated by the phonological sonority properties of obstruent+liquid. Murray and Vennemann (1983) downplay the role of sonority sequencing, or rather shift the motivation for this word division to Greek and Latin phonology, from where several of the forms in question are borrowed (<win|trau, sun|dro, Pai|trus>). In a later paper, Vennemann (1987) analyzes the whole set of relevant word division data in terms of sonority sequencing and thus provides a phonological interpretation. Even without appeal to word division, however, it is trivially the case that sonority sequencing is motivated within Gothic. The crucial data are easy to come by, as shown in 15.

Thus, initial sequences of consonants show that consonants are sequenced according to sonority (as expected, of course) and crucially that obstruents and liquids can cooccur in syllable onsets. The simple conclusion for word division is that divisions like <para|kletu> and <sun|dro> obey sonority sequencing, and that this phonological principle manifests itself dynamically by overriding what is otherwise the regular word division pattern.

Fourth, when we turn to sequences of two intervocalic consonants, the tendency is to divide the cluster in the middle. In Ambrosianus B and E this holds even when the consonants are an obstruent followed by a liquid. In Argenteus and Ambrosianus A the picture is more mixed. The Ambrosianus B and E evidence is interesting here because it provides an alternation in the treatment of obstruent+liquid. This alternation tells us that there is some phonological principle that causes the first syllable to sometimes be closed by a consonant. Thus when there are only two intervocalic consonants following a short vowel, one will be claimed by line 1. With an obstruent+liquid sequence this procedure is in conflict with sonority sequencing, as we now know from inspecting trisyllabic consonant clusters. Clearly, whatever the constraint is that claims a consonant to line 1, it is satisfied by one consonant, and when there are enough consonants left, sonority sequencing can rule (the same result is found for Old English, see Lutz 1986). The question is what this other constraint is. The best guess is Prokosch's law (PL). Prokosch's law (Prokosch 1939; Vennemann 1988; Riad 1992, 1995), also known as the Stress-to-Weight principle (McCarthy and Prince 1993, Kager 1999), states that stressed syllables should be heavy. This is again a universal tendency, and it is manifest in several ways in Germanic languages, both in synchrony and diachrony (syncope, apocope, quantity shift). Thus, a disyllabic word like <stib|nai> is divided such that—if this is a true reflection of phonology—Prokosch's law is satisfied by [b] and the onset principle is satisfied by [n]. As always, it is alternation that is truly informative, and Ambrosianus B and E provide it both internally and vis-à-vis Argenteus and Ambrosianus A. The variable behavior of obstruent+liquid in bi- and tri-consonantal clusters, respectively, shows that Prokosch's law (or its equivalent) influences word division (Riad 1992:87). And Prokosch's law is clearly more important than sonority sequencing when there is a conflict, since in the choice of [pe.tro] and [pet.ro], the latter is chosen. This is not to say that all word division reflects Prokosch's law, only that Prokosch's law is motivated within some word division and not contradicted, unless provoked by a stronger constraint (compare 9).¹⁵

There is of course more to say for the variation within and between manuscripts; compare Vennemann 1987. Our point is merely a methodological one, namely that word division does support the relevance of Prokosch's law in Gothic syllabification.

When there is no phonological conflict, however, the pattern is clear, in all manuscripts. If the vowel preceding the division is long as in <hlei|þrai> ‘tent’ and <bai|trei> ‘bitterness’, or unstressed as in <para|kletu>, or followed by three consonants as in <win|trau>, then the sonority sequencing plays out in full, since Prokosch's law is either satisfied (<win|trau>) or not relevant (<para|kletu>).

Before moving on, let us note that if we did not have this alternation in behavior among obstruent+liquid sequences, we would not be able to motivate Prokosch's law within the word division evidence, since the data are otherwise static (compare the situation for the onset principle). Alternation is thus crucial for us to establish from word division that something phonological is going on. If obstruent+liquid invariably split or invariably shifted to line 2, then we would have less of an argument for the visibility of Prokosch's law in orthography.

As our fifth and sixth phonological conclusions, we can list two orders of priority among the constraints that we have found to be motivated within the word division data. They are listed together with examples of data that crucially shows this relation. (Note that “

” should be read as ‘is more harmonic than’.)

¹⁶

The onset principle is more important than Prokosch's law, as a light stressed syllable will be preferred over an onsetless following syllable.

Each of the above conclusions is reflected in the orthographic evidence, but importantly, they are all motivated outside of this domain of evidence. Thus, the onset principle is a well known universal. But since it is a very highly valued constraint in Gothic (as in many other languages), alternations where this constraint asserts itself are hard to come by. However, in the variable realizations of /I/ and /U/ it can be demonstrated that ONSET is in fact active. Thus, /maUI/ comes out [ma.wi], obeying ONSET and violating Prokosch's law, rather than *[mau.i], where Prokosch's law is met and ONSET violated. (In the genitive both constraints can be satisfied: [mau.jo:s]). Beyond diachrony, Prokosch's law is manifest in the synchronic realization of sequences of vowels/glides (/nasIIs/ comes out as [nas.jis] rather than *[na.si:s]). Next, the tendency for obstruents and liquids to tautosyllabify is well known from many languages (for example, Latin and Modern Icelandic). It is just a subcase of sonority sequencing in syllable structure. As we have seen, it asserts itself in Gothic word division through the variable behavior in clusters of two and three consonants, respectively. Finally, the fact that the members of compounds form separate syllabification domains (viz., prosodic words) is clear in the modern languages as well. Nothing leads us to assume that things were different in Gothic.

Coda Maximization

When we have peeled off the last consonant in a cluster and used it as onset on line 2, we may be left with up to 4 consonants at the end of line 1. Prokosch's law has use for at most one of these. If the vowel of the first syllable is stressed and short, then, in accordance with Prokosch's law, that first syllable will claim one consonant in order to become heavy (bimoraic). This yields a residue of consonants, as illustrated below.

In clusters of more than two consonants there are leftovers, and the question is what we should do with them.¹⁸

Orthography clearly puts as many of these as possible on line 1, except in the obstruent+liquid cases, discussed above. Now, for the coda clusters, unlike the situation with obstruent+liquid clusters, we are left with a static pattern in orthography. Moreover, support from other sources such as change or morphophonological alternation is lacking. There is no known particular process of coda maximization in phonology. If anything, it is onsets that tend to be maximized. Furthermore, codas are banned in many languages, while onsets are banned in no language. Thus, it is even phonologically unlikely that some language would have a coda maximization rule. Notice that we need to get away from sonority sequencing in order to isolate a putative maximization rule, whether for onsets or codas. This is so because sonority sequencing often fully determines the syllabification, making a particular maximization principle redundant. Thus, there is really only one possible syllabification of a word like libandans ‘living’, namely li.ban.dans. The n and d cannot belong to the same syllable, because that would violate either sonority sequencing (*li.ba.ndans), or the onset principle (*li.band.ans).¹⁹ Thus, in order to properly test coda maximization we must look at clusters where a medial consonant that is out of reach for both the onset principle and Prokosch's law could go both ways, in terms of sonority.²⁰

The methodological question that arises here is whether orthographic word division alone is reliable as a foundation for a full interpretation as phonological syllabification in Gothic? I think the answer has to be no. There are two a prioristic reasons for this, the first being the fact that morphology appears to influence orthography. We have seen this in compounds and prefixed forms. How can we know that morphology does not affect orthographic word division elsewhere? At the very least, the case against morphological word division outside of those domains must be made. The second reason is the fact that there are orthographic rules in most writing systems. Most of these have their foundation in phonology or morphology, but there is often a residue of sundry rules that are usually called orthographic conventions. For instance, in Modern Swedish there is an orthographic rule that says that when you divide a word you should bring just one consonant to line 2. This rule applies even when it is not phonologically motivated.²¹

Without making any assumptions for the orthographic standards of Gothic written language, we must at least entertain the possibility of the influence of a writing convention in Gothic word division.

Next, it seems to me that any phonological conclusion drawn from orthography should be supported by some type of phonological evidence, too. At the very least, such information will affect the reliability of those conclusions. Nevertheless, it is not uncommon for people to assume that Gothic is a coda maximizing language, as a matter of phonological principle. The formulation of how this maximization is brought about varies, from a straight descriptive principle (Murray and Vennemann 1983, accepted by Palviainen 2001), to rankings of sonority related constraints (Pierce 2001:55). Note, however, that there is no alternation to support this conclusion. Coda maximization qua principle does not assert itself in the face of some other known phonological principle of Gothic. It is merely a static fact about all but the last consonant in clusters (modulo obstruent+liquid sequences). Coda maximization qua sonority-related constraints poses ranking problems, discussed below. We contend then, that coda maximization in Gothic is supported only by this orthographic pattern, and that it therefore is that much less reliable than the other principles. All it would take is for someone to show that Gothic word division does not always reflect phonological syllabification, and the support for coda maximization evaporates. The following discussion of another controversy, which arises from trusting orthographic word division too much, goes some way toward this.

Anti-sonority Treatment of Cluster-final Glides

Sonority sequencing is largely respected in syllabification. It is the guiding principle for how consonants are grouped around a vowel in a syllable. Sonority rises toward the vowel and falls away from the vowel. Everybody accepts the tautosyllabification of obstruent+liquid as a pattern grounded in sonority. The following problem then arises: Glides are even more sonorous than liquids, and so by universal implication, glides should also tautosyllabify with a preceding obstruent, at least in the same environments as liquids.

²²

However, orthography does not give any reflection of tautosyllabification of obstruents and the glide j, and that is the problem. Cast in methodological terms, the problem is that the situation exhibited in orthography means that we must either find some explanation for the unexpected phonology that we take to be reflected in orthography, or we must conclude that orthography is untenable as an indicator of all phonological syllabification. On the whole, the tendency among those concerned has been to try to salvage orthography as a source of evidence on phonological syllabification (beyond the core), and to find reasons not to expect phonologically tautosyllabified obstruents and j in Gothic onsets. The brute force solution is to propose a filter against tautosyllabified consonant and glide sequences. A straight formulation of this constraint as “*Cj” is given in Calabrese (1994:163) and Kiparsky (1998:351). The constraint militates against any consonant followed by the palatal glide within the same onset or coda. Some scholars stop here and use the filter as is in their further analysis, ultimately an unsatisfactory solution, but defensible perhaps as an interim solution.²³

Other scholars look for reasons why a filter like *Cj should hold in Gothic. Murray and Vennemann (1983) formulated the so called glide strengthening theory, further developed in Vennemann 1985, and argued against in Barrack 1998. Murray and Vennemann looked for arguments that would support a view of the glide [j] as a fricative. If the glide can be understood as less sonorous than the liquids, then of course the troubling sonority problem goes away.

Below we take a closer look at the evidence internal to Gothic that can be brought to bear on this discussion. My basic contention is that we are better off sacrificing some of our trust in orthography and thereby salvage sonority sequencing. The argument is made on methodological grounds. Thus I hope to convince the reader that our null hypothesis should be that *Cj is not right for Gothic because it goes against the basic methodologies by which we normally work.

*Cj

Let us look first at the motivation for *Cj. In discussions of orthography and syllabification in Gothic, it is invariably observed that the sequence Cj does not occur word initially in Gothic (Murray and Vennemann 1983, Calabrese 1994, Kiparsky 1998, Palviainen 2001, and Pierce 2001). And when we look at word divisions we find that words very regularly divide directly before the glide.

It would thus seem that the grouping of an obstruent or any other consonant with a following glide is avoided. As mentioned, the two sources of evidence that are brought to bear are word-initial clustering and behavior in word division. The filter *Cj formalizes this constraint and (nominally) accounts for why neither of these contexts displays tautosyllabic consonant+j sequences.

When we consider the value of these sources as cues to phonotactics, there are differences. When two consonants occur tautosyllabically as onsets, then that is good evidence for what can occur in Gothic onsets. But when two consonants do not occur together, then that is not good evidence for what cannot occur in Gothic onsets. Of course, in a case of outlandish sonority sequencing (*nkau), few would be inclined to assume that such sequences might occur in Gothic, but when the sonority sequencing is normal, the evaluation of the evidence—a methodological question—makes a difference. Possibly, the absence of consonant+j in word-initial position could be an inductive gap, due to, say, the fact that the Gothic manuscripts that have been preserved are too short or too genre-bound for that fairly rare Gothic word to occur.

Thus, scholars turn to word division in order to boost the argument for *Cj. However, it is easy to show that word division is unreliable for this purpose, but this fact has not figured in the discussion of orthographic word division as a source to Gothic syllabification. Part of the reason for this is the skewed focus on i/j, as opposed to i/j and u/w, in discussions of Sievers' law. Since the analysis of Sievers' law always involves a discussion of syllable structure, and therefore potential orthographic evidence, the constraint *Cj comes up, but putative *Cw or *Cn do not. Let us examine the argument in detail.

Consider first the sequence k+n. This is an obstruent followed by a nasal. In view of 18, it should be harder for these segments to cluster together than for an obstruent and a liquid, since nasals are less sonorous than liquids. When we look at orthographic word divisions we find that k and n before a vowel are invariably divided: <taik|neiþ> ‘sign’, <swik|nein> ‘cleansing’. Now, does this naturally lead to the assumption that these consonants cannot be grouped together? No, first we need to check the word-initial context and there we find kniu ‘knee’, knussjan ‘fall down’, and several more. Thus we assume that [kn] is a legitimate onset sequence, and for good reason. However, methodologically the consequence of this argumentation is that the regular separation of <k> and <n> by word division in orthography is not allowed as counterargument to phonological tautosyllabification. By the same token, then, the separation of any two consonants by word division should not be taken as evidence against their tautosyllabification.²⁴

One must always check the word-initial position, as well as the sonority sequencing. We find the same situation for example, with the fricative h followed by a sonorant: <hrains> ‘clean’, <hneiwan> ‘incline’, <hlahjan> ‘laugh’ versus <rah|neiþ> ‘count’.

The relevance of these remarks for putative *Cj becomes even clearer when we look at consonant+w, since this sequence parallels consonant+j for which the filter *Cj has been proposed. Again we find that C+w is regularly split in orthographic word division: <iz|war> ‘your’, <waurst|wa> ‘worker’, just like consonant+j sequences. Does this lead anyone to assume that there is a constraint *Cw? No, since when we look at the word-initial context we find that, unlike Cj, Cw does occur: swein ‘pig’, twaddje ‘second’. Again, it is the word-initial context that provides information on the possible clustering of consonants in the onset. Indeed, we expect Cw to tautosyllabify, because of the obstruent+liquid behavior both word-initially and in word division. This means that all we have that speaks against the tautosyllabification of consonant+j is the absence of such sequences word-initially. This is an inductive gap, and at best negative evidence. And it cannot be boosted by any obvious sonority reasons, since it should be a better context for tautosyllabification than obstruent+liquid. Indeed, since the other glide w happily teams up with an obstruent, it seems curious that tautosyllabic Cj should be ruled out in principle. The crucial conclusion to draw so far, then, is that orthographic word division does not bear on the issue of whether Cj is ruled out or not, since word division gives contradictory information in several other contexts. Thereby, word division is compromised as a source for what is a legitimate onset in Gothic.

One reason why C+j may be an inductive gap is the difference in sonority of i/j vis-à-vis u/w, discussed in the next section.

Glide Strengthening and the Asymmetry of /I/ and /U/

Murray and Vennemann (1983) realized the limited value of orthographic evidence with respect to the absence of C+j and looked further for a principled reason why Cj might not occur. They proposed that a strengthening process affected the glide j, turning it into a fricative. This would entail a radical change in sonority relations, moving j to the other side of liquids and nasals on the sonority scale; compare 18. This is a methodologically sound way to proceed, but as it happens, the argument as such runs into two problems relating to methodology that are particularly relevant to the present discussion.²⁵

First, it does not solve the problem. Glide strengthening proposes to explain (among other things) the putative heterosyllabification that the word division is taken to show, but as we have seen above, word division has no bearing—this way or that—on the sequencing of consonants in onsets, since word division facts routinely deviate from word-initial facts. Therefore, whether glides are strong or not cannot be determined or supported by word division. Second, glide strengthening of j creates a new problem with the other glide. The labiovelar glide commonly appears with other consonants on either side of it, and still it is demonstrably less sonorous than j, as discussed below.²⁶

There is a sonority difference between /I/ and /U/, which manifests itself in the phonotactic patterns they take part in. When we look at the realization of these underspecified phonemes in sequence with each other and other segments, we find indications that they differ in terms of sonority. /I/ is more sonorous than /U/, as we can see in three areas. First, when these segments occur in a sequence where only sonority influences their realization, the outcome is not symmetric: wi but not ui, iu but not ju (frequently). We find words like wilþeis ‘wild’ (*uilþeis), winds ‘wind’ (*uinds), but also siuks ‘sick’ (*sjuks), tiuhan ‘pull away’ (*tjuhan), and iupa ‘upper’ (*jupa). This means that if syllable nuclei tend to occur at the relatively more sonorous points in the string, then /I/ is more sonorous than /U/ (Riad 1992:55–56). We have the diphthong /iu/, but not the reverse one */ui/ in Gothic. Second, in combination with other segments, only w can occur outside of a consonant, both in onsets (wlits ‘face’, wrikan ‘pursue’) and in codas (gaidw ‘lack’, waurstw ‘work’). Words like putative +jlits and +gaidj do not occur in Gothic. Third, the glide j cannot occur alone in a coda, a fact we can ascertain by looking at forms where /I/ and /U/ combine with a vowel: hlaiw ‘grave’, but *mauj, and mawi ‘girl’, but *hlaju.

Thus if phonotactics should be understood as controlled in part by sonority, the conclusion must be that the glide w is less sonorous than j. Indeed, the glide w is so strong that it can occur on the outside of other consonants. Nevertheless, this segment happily combines with a preceding consonant in the critical contexts. The question then arises what good it does for the analysis of syllable structure to assume that there is strengthening of j. Why should a strengthened j lead to its inability to occur together with another consonant in a cluster, when the other glide, which is manifestly less sonorous, does occur with another consonant in Gothic consonant clusters? It seems that the only argument for *Cj is negative: C+j is not found initially in Gothic. Argumentation for putative glide strengthening will have to be made elsewhere than in orthographic word division.

The methodological point that I want to make is that word division does not bear on the issue. There is no principled reason to assume that *Cj is a real constraint, and at the very least, word division is no evidence for such a constraint.

Obviously, it remains unsatisfying to just flatly state that the absence of C+j word-initially is due to an inductive gap. However, there is a simple and easily checked diagnostic prediction here. If the absence of C+j is an inductive gap in the Gothic lexicon and the reason for this is chiefly the high sonority of /I/ (as compared to /U/ and all consonants), then the same gap, or at least a tendency toward it, could be expected to occur with a singular j initially, too. To check this expectation we should compare the number of j+V-initial words with the number of w+V-initial words.

This is a crude count of all entries in the two glossaries, but the result is clear enough: There are strikingly few word-initial js in Gothic. When we further look at the word types occurring under j+V and w+V, we find that the j+V group is mostly made up of function words (ja, jabai, jah, jai, jainar, jaind, ju, etc.) and proper names (Jaeirus, Janna, Judas), and that there is only a handful of content words (jiukan ‘to contend’, judaiwisks ‘Jewish’, juggs ‘young’, juk ‘yoke’, and a few derivations of them). The w+V contains mostly content words. This result points in the direction of an inductive gap for word-initial C+j.