2.1 Database

Both B&K and L&G performed statistical analyses of a subset of the Korean lexicon, viz. CVC words. B&K extracted all CVC entries from a large monolingual dictionary of Korean (Kim et al. Reference Kim, Ko, Im and Lee1991). L&G used a word list obtained from a website of the National Institute of the Korean Language (see Lee Reference Lee2006a: 52f.). The two resulting datasets are of quite different size: 3,001 items in the case of B&K and 939 items in the case of L&G.

If the different size of the datasets is responsible for the divergent results, it would have to be argued that large databases can cover up an effect which is present in smaller databases. Let us uncontroversially assume that a large dictionary contains a higher proportion of infrequent and obsolete entries than a smaller one. If we go along with Treiman et al.'s (Reference Treiman, Brett, Stephanie Knewasser, Bowman, Broe and Pierrehumbert2000) claim that infrequent items are less prone to develop a branching direction asymmetry than frequent ones, we might have an explanation for the null result in B&K and the left-branching bias found by L&G. Although we do not find this line of argument particularly convincing, we believe that it is worth testing whether the size of the database makes a difference. Below we present the results of a new analysis based on a different Korean dictionary, one that is smaller than that originally used by B&K.

We add in parentheses that L&G's word list was not put out in a typical publishing format so that it is not clear whether the compilers adhered to conventional lexicographic standards. We can do little to address this issue here other than to note that previous lexical analyses were usually based on published dictionaries (e.g. Kessler & Treiman Reference Kessler and Treiman1997).

2.2 Data selection criteria

Following Kessler & Treiman (Reference Kessler and Treiman1997), both B&K and L&G compiled datasets consisting of lexical items of immediate relevance to the internal structure of CVC syllables, viz. all monosyllabic CVC words of Korean. In doing so, their selection criteria were in many ways parallel. Both studies counted homophonous lexical entries separately. Neither study excluded loanwords. However, some differences remain. L&G excluded proper nouns whereas B&K did not. More importantly, B&K included inflected verb and adjective roots like mək- ‘eat’ and cak- ‘be small’. In dictionaries, these are listed with the infinitival suffix -ta (e.g. mək-ta ‘eat-inf’, cak-ta ‘be small-inf’). On the other hand, L&G restricted their set of CVC words to roots which are not morphologically bound. This means that their database was essentially limited to nouns and numerals, e.g. mok ‘neck’ and sam ‘three’. This might help to explain some of the difference in the size of the two datasets. However, it cannot be ruled out that bound verb and adjective roots are phonologically unusual and therefore create a special phonotactic bias. This possibility will be addressed below.

2.3 Level of abstractness

In Korean, phonemic representations (as found in dictionaries) often differ substantially from their phonetic realization in speech. One cause of this mismatch is the sweeping phonological process of coda obstruent neutralization (e.g. Park Reference Park1967, Kim-Renaud Reference Kim-Renaud and Kim-Renaud1986, Kim & Jongman Reference Kim and Jongman1996, Kaplan Reference Kaplan2011) whereby the wealth of 15 distinctive stops and fricatives collapses into a simple three-way stop contrast [p, t, k]. To borrow Lee's (Reference Lee2006a: 198) example, nac^h ‘day’, nas ‘scythe’ and nat^h ‘face’ are all realized as [nat]. In addition, there is a large number of other neutralization and assimilation processes affecting the surface form of both obstruents and sonorants (Sohn Reference Sohn1994), which further work to divorce the phonetic and phonemic levels.

B&K conducted independent analyses targeting both levels of representation. Their examination of abstract forms was based on the dictionary data described above, in which all underlying contrasts are maintained. Their analysis of surface forms was based on a collection of 12,500 speech syllables obtained from the Korean Telephone Conversation Speech Corpus (Ko Reference Ko2003a, Reference Kob) and from a recorded Korean radio talk show segment. Syllables were identified through the phonetic surface form in which they appeared in the speech data, i.e. after the application of all phonological processes such as resyllabification, neutralization, and assimilation. In contrast, L&G employed what might be seen as a hybrid approach. They used dictionary entries as their starting point but converted them into the allophones resulting from coda obstruent neutralization (Lee Reference Lee2006a: 198–199; see also pages 220–221, where all 76 VC sequences that entered the calculation are listed). Thus, for example, each of the three words nac^h, nas and nat^h was analyzed as nat.

L&G's hybrid approach is a possible alternative to B&K's separation of levels. However, it is not obvious why a left-branching effect should emerge in one but not the other approach. In order to clarify this question, we test the possible impact of the distinction between neutralized and non-neutralized coda obstruents in our new analysis below.

2.4 Segmental inventory

The Korean vowel and consonant inventories contain a few contrasts that are arguably obsolete. Decisions about the status of these cases may potentially affect phonotactic analyses. For the most part, B&K and L&G made identical assumptions. Neither team recognized the historical contrast between long and short vowels. Both assumed the contrast between the front vowels /e/ and /ε/.

A major complicating issue that was treated differently by B&K and L&G is the phonological status of the prevocalic glides /w/ and /j/. These occur in sequences like /ja/ and /wə/ in which they may be interpreted either as part of a diphthong (in a ‘nucleus analysis’) or as an independent onset consonant (in an ‘onset analysis’). L&G adopted the nucleus view (Kang Reference Kang2003). B&K did not commit themselves on this question and carried out three parallel analyses of their data: one treating glides as parts of diphthongs, another treating them as onsets, and a third one in which words containing prevocalic glides were excluded. Given that L&G's analysis is identical to one of B&K's subanalyses and that this subanalysis failed to bring out a CV effect, the disparate results are not attributable to any decision about the phonological status of glides. However, since we still regard the glide issue as unresolved, our new analysis below follows B&K's three-way approach.

2.5 Statistical analysis

The last point concerns the different statistical procedures adopted by B&K and L&G. To quantify the strength of association between adjacent segments in CVC syllables, both author teams obtained the frequency values a, b, c and d shown in (2) for all CV and VC sequences contained in their respective collections of CVC words.

1. (2)

   

The observed frequency of a `given phoneme sequence (value a) may be unexpectedly low or unexpectedly high in the light of the frequency with which each phoneme occurs in combination with other segments (values b and c). Either of these possibilities constitutes a probabilistic phonotactic effect. Where B&K's and L&G's procedures differ is on the question of how to quantify individual CV and VC effects and how to compare their combined strength.

B&K performed one-tailed Fisher Exact tests over the four frequency values in (2) for every CV and VC combination in order to identify the combinations which occur significantly more or less often than expected by chance. For example, their set of CVC syllables included 39 instances of the CV sequence /cʰə/, as in c^hən ‘thousand’. To determine the probability of this observation, they used (i) the number of times /cʰ/ and /ə/ occur together as onset and nucleus (39 times; value a), (ii) the number of times /cʰ/ is followed by other vowels (97 times; value b), (iii) the number of times /ə/ is preceded by onsets other than /cʰ/ (376 times, value c), and (iv) the number of times onsets other than /cʰ/ combine with nuclei other than /ə/ (2301 times, value d). Fisher's Exact test takes these four values as input and yields an extremely small likelihood of obtaining this particular distribution (p = .000014). Thus, the observed frequency of 39 (value a) is significantly higher than the expected value (20 in this example). B&K used the Fisher Exact test for these calculations because in some cases the numbers per cell were very small, including values of zero. For cell values below 5, the better known Chi-squared test does not yield reliable results, and the non-parametric Fisher Exact test is generally preferred. In addition, given the large number of tests carried out on the same data set, the significance level was appropriately lowered using the Bonferroni correction. In the example above, the obtained probability is below the adjusted significance level of .000139 (.05 divided by 360, where 360 is the number obtained by multiplying 18 onset consonants with 20 vowels).

In a second step, B&K calculated the share of significant CV and VC pairs out of the total number of theoretically possible CV and VC pairs and compared these again using Fisher's Exact test. To illustrate, one of their subanalyses examined 360 theoretically possible CV pairs and 320 theoretically possible VC pairs. In the CV set, they found 16 pairs occurring either significantly more or significantly less often than expected, and 344 which did not differ from chance. In the VC set, 10 pairs showed such exceptional frequencies, while the other 310 did not. The Fisher Exact test reveals a chance distribution (p = .244). B&K found no significant difference between the rate at which significant associations occur in CV and VC sequences in any of their subanalyses. This led them to reject the hypothesis that Korean syllables exhibit a left- or right-branching phonotactic bias.

L&G assessed the strength of association for each attested CV and VC combination by means of Pearson's r _Φ, a correlation statistic which Perruchet & Peereman (Reference Perruchet and Peereman2004) advocated in their investigation of French syllables. The r _Φ formula is given in (3).

1. (3)

   $$r_{\it \Phi} = \displaystyle{{ad - bc} \over {\sqrt {(a + b) \times (c + d) \times(a + c)\times (b + d)}}} $$

To the extent that two given segments are correlated, in the sense that they do or do not tend to occur together, the r _Φ value will positively or negatively diverge from zero. Having determined two sets of r _Φ values, one for all CV associations and one for all VC associations, L&G then used a Mann–Whitney U-test (also known as the Wilcoxon rank sum test) to determine whether there was a significant overall difference in the r _Φ values. Their CV set displayed significantly greater absolute values overall, which they interpreted as supporting a left-branching phonotactic structure.

It is possible that L&G's statistical procedure is more sensitive than B&K's so that the former picks up a relatively weak effect which passes unnoticed in the latter. In B&K's mode of calculation, all sequences which did not reach statistical significance in the first round of calculations were discarded and did not enter the overall CV–VC comparison. In L&G's method, by contrast, the effect of all associations, no matter how strong each one was individually, was compared as a whole, which allows even weak associations to affect the overall distribution. Arguably, this makes L&G's method more sensitive than B&K's.

Given these methodological considerations, we decided to re-calculate the series of analyses reported in the B&K study using L&G's method of calculation. The results are presented at the beginning of the next section. We then go on to present our new Korean analysis, which addresses the other factors discussed above.

6.1 An account of the Korean and Finnish data

We believe that the debate over the phonotactic structure of Korean monosyllables can be put to rest. There is now fairly robust evidence for the lack of a phonotactic asymmetry in Korean CVC words – a result on which all our subanalyses converge. This finding runs counter to the highly consistent experimental evidence arguing for a left-branching analysis of Korean syllables (see Yoon & Derwing Reference Yoon and Derwing2001 and the references cited above), which receives additional support from reduplication patterns (Jun Reference Jun1994).

How can the phonotactic and the experimental evidence be reconciled? As noted at the beginning of this paper, B&K's (Reference Berg and Koops2010: 44–46) earlier discussion of this question suggested that the criteria traditionally used to determine branching direction may contain inherent biases. In particular, the phonotactic criterion was hypothesized to have a universal right-branching bias. If this right-branching bias meets a (phonologically) left-branching language like Korean, and if the two biases are of similar strength, they effectively cancel each other out. The net result is the lack of a CV or VC effect, as we observed in Korean CVC words.

How can this mode of reasoning be extended to Finnish, where the empirical picture is slightly different? Specifically, how do we explain the relative lack of a phonotactic asymmetry in Finnish word-initial syllables and a VC effect in word-final ones? We suggest that the postulated right-branching bias is sensitive to word position: while word-final syllables display a phonotactic VC effect, word-initial syllables are more symmetrical. To see why this should be the case, we need to reconsider the nature of phonotactic constraints in the context of polysyllabic words.

Let us adopt the dynamic perspective of word recognition. Lexical processing begins immediately upon hearing the first segment of a word (Marslen-Wilson Reference Marslen-Wilson1975) and then proceeds in a stepwise fashion whereby the number of mental addresses that match the auditory input quickly diminishes (Marslen-Wilson & Tyler Reference Marslen-Wilson and Tyler1980). In fact, lexical access is often unambiguously possible before the end of a word has been perceived because the incomplete phoneme string already diverges from all others in the lexicon. It follows that a phonotactic constraint placed beyond the uniqueness point implies a loss of dispensable phonological contrasts. Hence, the later in a word phonotactic constraints occur, the lower the threat to the distinctive function of phonemes.

There is another way in which the listeners' perspective favours word-final phonotactic constraints. As the listeners' task is to identify the unfolding word as quickly as possible, it can be assumed that they will exploit every available cue to the identity of the unfolding word, including the immediately upcoming segments. Phonotactic constraints provide such a cue in so far as they increase the predictability of upcoming segments. They do so by decreasing the number of following possibilities. To take a hypothetical example, upon hearing /t/, listeners who know that /t/ is rarely or never followed by /i/ will be able to adjust their expectations of possible continuations and hence be more likely to venture a correct guess at the unfolding word. Note that this strategy is least effective at the word onset and more effective the further advanced the word recognition process is. At an early stage of word recognition, the number of competing lexical candidates is typically still high. The small amount of added predictability that a phonotactic effect provides will be of very limited value in this situation because the cohort of lexical candidates is so large that taking a successful guess at the word is not a realistic option. However, at later stages of hearing a word, especially as the process of lexical access nears completion, a guess is more likely to be correct if the pool of lexical candidates is small.

We have offered two reasons why phonotactic constraints are more effectively placed far from the onset and as close as possible towards the end of words. This is consistent with our results of Finnish phonotactics. In word-initial syllables, the inherently right-branching nature of the phonotactic bias is relatively weak. It is just strong enough to counteract the psycholinguistic left-branching bias. As in Korean monosyllables, the clash of these opposed forces results in a largely balanced distribution of phonotactic effects. At the end of Finnish polysyllables, however, the phonotactic bias exerts itself more forcefully. In fact, it is so strong that it overcomes the language's general left-branching bias. As a consequence, word-final syllables evince a phonotactic VC effect.

Convincing alternatives to the above processing account are hard to find. It is true that there are phonological differences between initial and final syllables, but it is doubtful that these are at the bottom of the phonotactic patterns. In particular, initial syllables have been claimed to be more perceptually salient and to code a higher number of contrasts than final ones (e.g. Beckman Reference Beckman1997). Let us assume, for the sake of the argument, that this is true of Finnish. This assumption ties in well with the fact that Finnish word-initial syllables are always stressed and word-final syllables always unstressed, given that stressed syllables are ‘perceptual anchors’ (e.g. Cole & Jakimik Reference Cole, Jakimik and Underwood1978). A possible hypothesis that emerges from this difference is that the absence of stress leads to a VC-effect and the presence of stress to no such effect. As far as we can tell, there is no evidence for such a claim in Finnish. If anything, the opposite is more likely to be true. It is well-known that segmental structure is more complex in stressed than in unstressed syllables. Analogically, we would expect suprasegmental structure to behave likewise. Since Finnish syllables show the opposite pattern, we are reluctant to hold stress (or contrast, for that matter) accountable for the empirical differences reported above.

6.2 Severing the link between phonotactic constraints and branching direction

The results of our analyses demonstrate a striking disconnect between phonotactic constraints and sub-syllabic constituency.Footnote ⁷ Judging by all the evidence other than the phonotactic criterion, both Korean and Finnish have ‘basically’ left-branching syllables. However, nowhere is this left-branching nature reflected in the phonotactic structure of CVC syllables. At best, it can be observed indirectly where the right-branching phonotactic bias neutralizes a ‘basically’ left-branching bias, resulting in the lack of a phonotactic asymmetry. Thus, the phonotactic criterion fails to predict phonological branching direction. Phonotactics and phonological branching structure should accordingly be viewed as independent phenomena.

The reason why phonotactics has loomed large in previous analyses of branching direction now emerges as a historical accident. The languages on which earlier analyses were performed are all basically right-branching ones. As the inherent right-branching bias of phonotactic constraints happened to match the right-branching syllable structure of languages such as English and German, students of these languages (including ourselves) could be easily tempted into treating phonotactics as valid evidence for branching direction. It is time to break with tradition on this point.

6.3 A phonetic account of sub-syllabic branching direction

Lee & Goldrick's (Reference Lee and Goldrick2008) short-term memory experiments demonstrate that individual CV and VC sequences which are strongly associated phonotactically are recalled more easily than weakly associated ones. This supports the notion that speakers keep track of the probabilities with which segments co-occur within syllables. What is less clear is whether their knowledge of segment-level phonotactic effects can give rise to a general phonological branching structure, for example a body node, as Lee & Goldrick (Reference Lee and Goldrick2008, Reference Lee and Goldrick2011) propose. As long as the phonotactic evidence dovetails with other indicators of branching structure, their account appears convincing. However, when the two phenomena are at odds with each other, as we have shown here for Korean and Finnish, their theory falls apart.

With phonotactics removed as an explanatory factor, what is responsible for phonological left-branching in languages such as Korean and Finnish? In the following, we will discuss and reject one explanation that has been advanced in the relevant literature on Korean and subsequently suggest a novel one. It should be noted from the outset that our discussion of branching direction and its underlying causes is limited to the phonetic-phonological level. That is, we will not consider possible spill-over effects from other levels, such as ‘cross-level harmony’ between branching direction asymmetries in syntax and other levels.

Kim (Reference Kim2007) put forward an explanation of the body-coda preference which is based on the frequency of different syllable types. He argues that because CV syllables occur in Korean far more frequently than CVC, V, or VC syllables, the higher exposure to CV syllables leads Korean speakers to treat CV as a unit. Note that in this proposal language users are implicitly assumed to equate an entire CV syllable with the body of a CVC syllable. While not implausible, this hypothesis cannot be held up. It is well-known that there is a strong asymmetry between CV and VC syllables. Typologically, almost all languages have CV syllables and a sizeable number of them have only CV syllables while only very few have only VC syllables (e.g. Kaye & Lowenstamm Reference Kaye, Lowenstamm, Belletti, Brandi and Rizzi1981, Clements & Keyser Reference Clements and Keyser1983, Breen & Pensalfini Reference Breen and Pensalfini1999). Moreover, research on child language acquisition has shown that CV syllables are acquired prior to VC syllables (e.g. Levelt, Schiller & Levelt Reference Levelt, Schiller and Levelt2000). The attempt of relating the ‘unmarked’ nature of CV syllables to the issue of sub-syllabic constituency would lead us to expect a predominance of left-branching in the phonologies of the world's languages. However, the opposite is true (Berg Reference Berg2009). To give only one example, Spanish has an absolute majority of CV syllables (Lloyd & Schnitzer Reference Lloyd and Schnitzer1967) but shows relatively strong signs of being a phonologically right-branching language. Thus, the predominance of CV syllables does not provide an account of branching direction.

We pursue this line of argument further and focus on CVC syllables in Finnish. The general logic of the frequency argument is that units which are used together fuse together (Bybee & Scheibman Reference Bybee and Scheibman1999: 578). The first subanalysis of Table 4A above was based on 4,084 CVC syllable tokens, yielding 4,084 CV and 4,084 VC tokens. The CV sequences distribute across 172 CV types and the VC sequences across 51 VC types. Hence, the average token frequency of CV types is 23.7 while that of VC types is 80.1. This difference leads us to expect VC to be a more cohesive unit in Finnish than CV is. However, this prediction is incompatible with the left-branching nature of Finnish syllables (see Section 4). We thus cannot help severing the link between token frequency and branching direction.

We propose to shift the burden of explanation from phonotactics to phonetics. Let us return to the essence of branching direction and recall that branching structure is a hierarchical way of representing varying degrees of cohesion. For example, a right-branching syllable structure implies more cohesiveness between V and C than between C and V. At the phonetic level, cohesiveness can be brought about by coarticulation, i.e. the extent to which two adjacent segments overlap in terms of articulatory gestures and their acoustic consequences. If it can be shown that Korean CV sequences evince a particularly high degree of coarticulation, we might have a phonetic basis for the left-branching bias.

Evidence to this effect comes from the literature on the phonetics of the three-way contrast of the Korean voiceless stops, often called tense, lax, and aspirated, which only exists in onset position. Cho, Jun & Ladefoged (Reference Cho, Jun and Ladefoged2002) studied the extent to which the stop contrasts are reflected in the voice quality of the following vowel. They found that coarticulatory cues, specifically short- and mid-range spectral slope as well as fundamental frequency (f0), co-vary systematically with the onset stop type. As Cho et al. discuss in detail, while the existence of CV coarticulation in f0 is not surprising, the temporal extent to which the consonantal information is carried over into the vowel in Korean is unusual. Crucially, the relevant differences are not only seen near the vowel onset, as regularly found in other languages, but spread well into the vowel. Similarly, Kim, Beddor & Horrocks (Reference Kim, Beddor and Horrocks2002) found that phonatory cues contained in the following vowel, in particular f0, are especially strong in Korean, with as much as a 30 Hz difference in f0 distinguishing the lax from the tense and aspirated stops even at the vowel offset. The authors also tested the role of the vowel in the perception of the stop contrasts and found that Korean listeners are able to correctly identify manner of articulation when the stop portion is artificially removed. Moreover, when presented with contradictory information in the form of cross-spliced CV stimuli, the vocalic information usually overrides the consonantal information.

We submit that coarticulation is a better candidate for giving rise to sub-syllabic constituency than is phonotactics. To begin with, coarticulation is a given. Gestural overlap is an inevitable side effect of fluent articulation. The links which coarticulatory effects create between adjacent phonemes are directly observable in the phonetic signal. This means that groupings of segments established by coarticulation, especially where the degree of coarticulation is high, need not be ‘discovered’. They are part and parcel of speech processing. Strongly coarticulated segments can thus be seen as natural processing units. In addition, coarticulatory effects are local in the sense that they can be observed in individual words in ignorance of all others. Remarkably, phonotactic effects are entirely different. Phonotactic constraints can only be defined against the background of the complete lexicon. Nothing at all can be said about the phonotactic structure of an individual word viewed in isolation. In this sense, phonotactic constraints are of a global nature. They represent the outcome of a computational process over all items in a lexicon. This derivative nature of phonotactics contrasts sharply with the local nature of coarticulation. It may thus be argued that coarticulation is the simpler, the more salient, the more reliable and probably also the stronger cue for branching direction than phonotactic constraints.

We therefore tentatively suggest that the strong coarticulation found in Korean CV units may underlie the left-branching nature of CVC syllables. To be sure, such an account is no more than a first attempt at a phonetic explanation. The high degree of coarticulation discussed above has only been demonstrated for CV sequences in which the onset consonant is a stop. While it is true that the 12 Korean stops make up the bulk of the 19 consonants, our case would be strengthened if it could be demonstrated that the same holds for all CV sequences. Also, the phonetic studies cited above only speak to CV coarticulation. A direct comparison between CV and VC coarticulation would be desirable. However, this is impossible because, as mentioned, the relevant stop contrasts are all neutralized in coda position.

The hypothesis that coarticulation is a determinant of sub-syllabic structure generates predictions for other languages. For example, it leads us to expect that coarticulation within VC units is stronger than within CV units in phonologically right-branching languages. There is some evidence to bear this out. As summarized by Browman & Goldstein (Reference Browman, Goldstein, Bell-Berti and Lawrence1995), phonetic research on English sonorants shows that there is more temporal overlap between the vowel and the consonant gesture in nasals and laterals in coda position than in nasals and laterals in onset position (see Sproat & Fujimura Reference Sproat and Fujimura1993 on laterals, Krakow Reference Krakow1999 on nasals).

For obstruents in CV sequences the picture is less clear, however. In the case of stop–vowel sequences, it is well known that the formant transitions near the vowel onset contain crucial information about the stop's place of articulation (e.g. Liberman Reference Liberman1957, Liberman et al. Reference Liberman, Cooper, Shankweiler and Studdert-Kennedy1967). Unfortunately for our purposes, phonetic studies of CV coarticulation have rarely directly compared the relative degree of coarticulation in tautosyllabic CV and VC sequences. A notable exception is the study by Sussman et al. (Reference Sussman, Nicola Bessell and Majors1997), who report less coarticulation between vowels and coda stops than between vowels and onset stops (see also Tabain, Breen & Butcher Reference Tabain, Gavan and Butcher2004). The notion that vowels and onset consonants form a unit is also central to the work of Nam & Saltzman (Reference Nam and Saltzman2003) and Nam, Goldstein & Saltzman (Reference Nam, Hosung, Saltzman, Chitoran, Pellegrino, Marisco and Coupé2010), who have modelled syllables as self-organizing structures in which CV patterns are intrinsically more stable than other patterns of gestural organization. Thus, pending further work on the hypothesized link between coarticulation and syllable structure, our line of argument has to remain highly tentative.

Finally, we hasten to add that the available evidence from coarticulation pertains only to Korean. We are unaware of a comparable study of CV coarticulation in Finnish, let alone a study comparing coarticulation in Finnish CV and VC sequences. We would not even want to commit ourselves to the claim that coarticulation is the only possible basis from which to infer sub-syllabic branching structure. Other phenomena such as the relative timing of consonants and vowels in CV and VC units, as seen in rime isochrony in English, might also play a part. The relevant parameters need not all be phonetic, but could also be phonological or psycholinguistic in nature.