2.1 Variation

As mentioned above, the compound tensification process is highly opaque in Contemporary Korean. Given this, we expect a substantial amount of variation to be observed in the application of sai-sios. Indeed, in our data there are very few compound words for which all 35 speakers agreed in their judgement of tensification (only four words). Figure 1 shows the tensification rate (the ratio of the application of tensification) and the number of compound words. For example, the tensification rate of /kil + pa.tak/ ‘road + surface’ is 84% (32 tense; 6 lax). The tensification rate seems to be quite gradient: some words are tensified very frequently, others less frequently, and some others are rarely tensified. Examples of variation are given in Table II.

Figure 1

Tensification rate for 1171 compound words, divided into deciles.

Table II

Examples of variation. Numbers in parentheses indicate the number of speakers. The total number of speakers is in some cases greater than 35 (the number of subjects in this study), since some speakers reported that both tensified and non-tensified forms are used.

For reasons of space, glosses are not given in this and similar tables. Further examples, with glosses, can be found in the online supplementary materials.Footnote ⁷

Figure 2 shows the correlation between year of birth and tensification rate. In this case, the tensification rate was calculated by dividing the counts of the application of tensification by the total number of responses by one speaker. When an individual speaker offered multiple responses for a single word, both were counted. There does not seem to be any significant correlation between year of birth and tensification rate (p-values for linear/quadratic/cubic models are 0·2667, 0·3174 and 0·3431 respectively), nor, as seen in Table III, is there a clear difference in tensification rate between males and females. Thus variation in tensification rate appears to depend neither on age nor on gender; rather, tensification rate varies for each individual.

Figure 2

Compound tensification rate as a function of year of birth for the 35 speakers.

Table III

Compound tensification rate as a function of gender.

Given this, we can ask to what extent the Yanbian data agrees with the Standard Korean dictionary data. Table IV and Figure 3 show the correlation between Standard Korean and Yanbian. The Standard Korean data is based on Kwuklip kwuke yenkwuwen (1999), in which the application of tensification in compound words is indicated (as a rule, variation is not reported). The Yanbian data in Table IV (tense vs. lax) gives the total number of datapoints for each word with or without sai-sios tensification in Standard Korean. Interestingly, for both tense and lax, the Yanbian data agree with Standard Korean in about 70% of the cases.

Figure 3

Correlation between Standard Korean and Yanbian tensification. The horizontal line in each box indicates the median value, the ends of the vertical lines give the minimum and maximum data values, and the bottom and top of each box are the first quartile (25th percentile) and third quartile (75th percentile) respectively.

Table IV

Correlation between Standard Korean and Yanbian sai-sios tensification. Words that do not appear in the Standard Korean dictionary (mostly dialectal) are not included here. Note that the same word may appear as either tense or lax in Yanbian, depending on the speaker.

These data clearly show that significant variation is observed in compound tensification in Yanbian. This may be partly attributed to the different orthographic conventions in Standard Korean and Yanbian mentioned above: in Standard Korean, sai-sios is marked by the character s, at least when W_A ends in an open syllable, while it is never reflected in the orthographic system utilised in Yanbian. This difference may have resulted in the substantial individual variation seen in Yanbian (how much speaker variation is observed in the compound tensification of Standard Korean is a topic for future research).

2.2 Token frequency (familiarity ratings)

Zuraw (Reference Zuraw2011) points out that compound tensification is more often observed in more frequent lexical items. In order to test this token-frequency effect, familiarity ratings of all compound words in our corpus were judged by four Yanbian native speakers.Footnote ⁸ Familiarity ratings were based on a seven-point scale from 1 (least familiar) to 7 (most familiar). Examples are given in Table V. Figure 4 shows the correlation between the average scores of the four raters and the tensification indexes from the 35 speakers. For example, for items with a familiarity average of 5·5, there are 916 datapoints for tense and 559 datapoints for lax, a tensification rate of 62%. As seen in Fig. 4, even the familiarity-rating data obtained from naive native-speaker judgements correlate with tensification rate: the higher the familiarity rating, the higher the tensification rate. A linear regression analysis with tensification rate as the dependent variable and familiarity rating as the independent variable shows a significant result (est.=0·05, SE=0·00, t=10·2, p<0·0001). Thus our data confirm Zuraw's (Reference Zuraw2011) results.

Figure 4

Compound tensification rate as a function of averaged familiarity rating for 1171 words evaluated by four Yanbian native speakers.

Table V

Examples of compound tensification and averaged familiarity rating.

2.3 Number of syllables

According to Zuraw (Reference Zuraw2011), sai-sios is less probable when W_A and W_B both contain more than one syllable. Our data confirm this, as shown in Table VI and Fig. 5. Moreover, in our data, the probability of tensification correlates more linearly with the stem length of W_A and W_B (except for monosyllabic W_B) than reported by Zuraw (Reference Zuraw2011).Footnote ⁹

Figure 5

Compound tensification rate for first (W_A) and second (W_B) elements as a function of the number of syllables.

Table VI

Compound tensification rate for W_A and W_B elements as a function of the number of syllables.

Examples are given in Table VII.

Table VII

Examples of compound tensification and syllable count in W_A.

I hypothesise that the stem-length effect is correlated with phrase structure. In Yanbian, compound words tend to have only one pitch peak, and its location is as a rule determined by the W_B accent, e.g. /ó.ri + ál/ ‘duck + egg’ → [o.ri.ál] ‘duck's egg’, /pa.tá + ó.ri/ ‘sea + duck’ → [pa.ta.ó.ri] ‘guillemot’. However, when a compound word has four syllables or more, speakers may pronounce these words as two accentual phrases, probably due to a *Lapse-L constraint, which penalises a candidate with two successive syllables without H at the left edge of the word (Ito Reference Ito2013), e.g. /múl + o.cíŋ.ə/ ‘water + cuttlefish’ → {mul.o.cíŋ.ə} ~ {múl}{o.cíŋ.ə} ‘living cuttlefish’. The occurrence of sai-sios (tensification) implies that there is no phrase boundary between W_A and W_B, which is a countervailing condition for longer words. Thus the lower tensification rate for longer words is explained as a result of a conflict between the compounding process and a tendency to divide a long word into shorter accentual phrases.Footnote ¹⁰

2.4 W_A final-segment type

In Standard Korean, tensification applies automatically after an obstruent coda (post-obstruent tensification). However, our investigation of Yanbian compound nouns has found quite a few words that do not tensify even in this environment, as exemplified in Table VIII.

Table VIII

Examples of compound tensification with W_A obstruent codas.

Tonal patterns indicate that these compound words (at least the shorter ones) constitute a single phonological word, and hence satisfy the structural description of the process. The same word can sometimes have tense and lax variants for a single speaker, e.g. /pʰatʰ + pap/ ‘red beans + rice’ → [pʰat.páp] ~ [pʰat.p*áp] ‘rice cooked with red beans’, as seen in Fig. 6.Footnote ¹¹

Figure 6

Sample waveforms, spectrograms and spectra from speaker PKL (female, born in 1985) for (left) [pʰat.páp] and (right) [pʰat.p*áp] ‘rice cooked together with red beans’. CD=closure duration (270 ms for [pʰat.páp]; 294 ms for [pʰat.p*áp]). VD=vowel duration of W_B (214 ms for [pʰat.páp]; 259 ms for [pʰat.p*áp]). H1 – H2 of the vowel onset in the second syllable=5·1 dB for [pʰat.páp] and −3·6 dB for [pʰat.p*áp].

Nevertheless, the probability of tensification is still much higher after an obstruent coda than after a sonorant coda or a vowel, as seen in Table IX, suggesting that it is motivated by post-obstruent tensification.Footnote ¹² A Pearson's χ² test with Yates' continuity correction shows that obstruent and sonorant codas are significantly different in their sai-sios effects (χ²=298·15, p<0·0001). Similarly, a sonorant coda and vowel are significantly different (χ²=348·46, p<0·0001).

Table IX

Compound tensification rate as a function of W_A final-segment type (35 speakers, 1171 words).

Fig. 7 shows W_A final-segment type and tensification rate for all speakers.

Figure 7

Compound tensification rate as a function of W_A final-segment type (35 speakers, 1171 words).

Given that the tensification rate is highest in obstruent codas and post-obstruent tensification is obligatory in Yanbian if the targeted sequence appears within a simplex word (e.g. /sɛk.si/ [sɛk.s*i] ‘maiden’, /nak.ci/ [nak.c*i] ‘octopus’), the observed pattern of variation can be modelled as the interaction of at least three constraints, with different weights: RealiseMorpheme, Post-obstruent Tensification in Compounds (Post-obsTens/Comp) and a faithfulness constraint to the base simplex words (Faith(W_B)), as shown in (2) for /apʰ + ta.ri/ ‘front + leg’ → [ap.ta.ri]~[ap.t*ari] ‘front leg’. Multiple candidates are shown as winners, since there is variation, but the untensified candidate is less probable than it would be when W_A ends with a sonorant, due to the violation of Post-obsTens/Comp in the former. A more detailed weighted constraint analysis of sai-sios distribution is presented in §2.6.

1. (2)

   

The relatively higher tensification rate after a sonorant coda than after a vowel may reflect a typological preference for lenition, such as shortening or voicing of an intermediate consonant: V__V is preferred to R__V for lenition (where R=sonorant), due to the higher sonority or more open vocal tract of vowels. In the case of compound tensification, this typological preference may work against the application of sai-sios, resulting in different tensification rates for sonorant codas and vowels in W_A.

2.5 Laryngeal co-occurrence restrictions

Zuraw (Reference Zuraw2011) shows that Standard Korean sai-sios is blocked if the stem (either W_A or W_B) already contains tense consonants. As noted in §1, this is similar to rendaku in Japanese, in that voicing of a consonant is blocked when the W_B stem contains a voiced obstruent.Footnote ¹³ Our Yanbian data confirm this, but, in addition, aspirated consonants in W_A and W_B tend to block tensification.Footnote ¹⁴

As seen in Table X and Fig. 8, the tensification rate is lower when W_A contains either aspirated or tense consonants. A Pearson's χ² test with Yates' continuity correction shows that aspirated vs. lax and tense vs. lax in W_A are significantly different in their effect on sai-sios, whereas aspirated vs. tense is not (χ²=69·14, p<0·0001; χ²=88·03, p<0·0001; χ²=0·01, p=0·9227).Footnote ¹⁵

Figure 8

Compound tensification rate as a function of W_A consonant type (35 speakers, 1171 words).

Table X

Compound tensification rate as a function of W_A consonant type (35 speakers, 1171 words).

Examples are given in Table XI.

Table XI

Examples of compound tensification as a function of W_A consonant type.

This variation in compound tensification contrasts markedly with the automatic and general post-obstruent tensing process of Yanbian Korean phonology that applies whenever a stem-final obstruent combines with a suffix-initial obstruent, regardless of the laryngeal features in the preceding stem: e.g. /cap-ko/ [cap.k*o] ‘catch + gerund’, /cʰap-ko/ [cʰap.k*o] ‘be cold + gerund’, /c*ap-ko/ [c*ap.k*o] ‘be salty + gerund’; /tək-to/ [tək.t*o] ‘virtue + also’, /tʰək-to/ [tʰək.t*o] ‘jaw + also’, /t*ək-to/ [t*ək.t*o] ‘rice cake + also’.

The compound tensification rate is also lower when W_B contains either aspirated or tense consonants (Table XII and Fig. 9).Footnote ¹⁶ A Pearson's χ² test with Yates' continuity correction shows that aspirated vs. lax and tense vs. lax in W_B are significantly different in their effect on sai-sios, while aspirated vs. tense is not (χ²=262·75, p<0·0001; χ²=138·09, p<0·0001; χ²=0·6, p=0·8102). Examples are given in Table XIII.

Figure 9

Compound tensification rate as a function of W_B consonant type (35 speakers, 1171 words).

Table XII

Compound tensification rate as a function of W_B consonant type (35 speakers, 1171 words).

Table XIII

Examples of compound tensification as a function of W_B consonant type.

Figure 10 shows the combined result of Figs 8 and 9. Except for the case in which both W_A and W_B contain tense consonants (‘tns-tns’), which appears in just a handful of examples, gradient OCP effects are observed.

Figure 10

Compound tensification rate as a function of W_A and W_B consonant type (35 speakers, 1171 words). ‘asp-asp’ means that both W_A and W_B contain aspirated consonants; ‘l/s-tns’ means that W_A contains lax/sonorant consonants and W_B tense consonants.

It is well known that co-occurrence restrictions on laryngeal features (aspiration, glottalisation, etc.) are observed in many languages. For example, MacEachern (Reference MacEachern1999) and Gallagher (Reference Gallagher2010, Reference Gallagher2011, Reference Gallagher2012) report that Peruvian Aymara and Quechua show co-occurrence restrictions on multiple features (aspirated and ejective). The gradient restriction on compound tensification in Yanbian provides further evidence for this typological generalisation.

As seen in Table XIV, it is unclear from our data whether compound tensification is locality-sensitive, i.e. whether or not a laryngeally marked segment in an adjacent syllable blocks tensification more than one which is not in the adjacent syllable. In this table, compound data whose W_A is monosyllabic are excluded, since monosyllabic W_A does not have a non-local case. W_A words that contain both aspirated and tense consonants are also excluded. The tensification rate is higher when the marked feature (aspirated or tense) is adjacent to W_B than when it is not. A Pearson's χ² test with Yates' continuity correction shows that the distribution is not statistically significant for aspirated consonants (χ²=0·38, df=1, p=0·5366), while it is significant for tense consonants (χ²=7·44, df=1, p=0·0064), but the bias is in the opposite direction from what is expected. For W_B, all the instances except for one (/ma.nɨl + caŋ.a.c*i/ ‘garlic + slices of radish or cucumber dried and seasoned with soy’ → [ma.nɨl.caŋ.a.c*i] ~ [ma.nɨl.c*aŋ.a.c*i] ‘pickled garlics’) are local (a laryngeally marked segment appears in the second syllable of W_B), so the locality effect cannot be checked in W_B.

Table XIV

Compound tensification rate as a function of position of W_A consonant type.

Examples are given in Table XV.

Table XV

Examples of compound tensification as a function of locality of W_A consonant type.

2.6 Summary

As a summary of the phonological effects mentioned so far, I constructed a mixed effects logistic regression model using the glmer function from the lme4 package (Bates et al. Reference Bates, Maechler, Bolker and Walker2014) in R (R Development Core Team 2011). Tensification was the dependent variable (non-tensified as the reference category) and the various factors mentioned above were the independent variables: familiarity ratings (numerical), W_A/W_B stem length (numerical), W_A final-segment type (obstruent, sonorant, vowel) and W_A/W_B consonant type (aspirated, tense, lax/sonorant). Random intercepts were set for item and subject. For W_A final-segment type and W_A/W_B consonant type, backward difference coding was utilised: W_A final-segment types (vowel > sonorant > obstruent), W_A/W_B consonant type (tense > aspirated > lax/sonorant). An ANOVA() for likelihood-ratio test was carried out for model comparison, to rule out the more complicated models. Table XVI shows the results. As the familiarity ratings increases, items tend to be tensified; as the stem length increases, the tensification rate declines; vowel-final W_A is tensified less than a sonorant coda, which is in turn tensified less than an obstruent coda; tense consonants in W_A/W_B are not significantly different from aspirated consonants in W_A/W_B, while aspirated consonants in W_A/W_B are less tensified than lax/sonorant consonants. More complicated models, with an interaction of each of the predictors (e.g. a model with an interaction of W_A and W_B consonant types), either failed to converge or were not significantly different from this simpler model.

Table XVI

Results of a mixed effects logistic regression model (lexical data). Significant values are indicated by * (p<0·05), ** (p<0·01) and *** (p<0·001).

I also tried to find a gradient constraint weighting, using Hayes' (Reference Hayes2009) Maxent Grammar Tool. The constraints employed in the simulation are given in (3).

1. (3)

   
   Footnote ¹⁷

The weights obtained are shown in Table XVII.

Table XVII

Weights obtained using Hayes' Maxent grammar tool.

Two constraints with opposing effects (RealiseMorpheme and Faith(W_B)) are weighted highest, with the slightly higher value for the former (5·32 vs. 4·68) probably mirroring the overall distribution of sai-sios in Yanbian: tensified (24,844; 56%) vs. non-tensified (19,501; 44%). *Tense/Vowel-W_A is weighted higher than *Tense/Son-W_A (0·79 vs. 0·50), reflecting the higher tensification rate in the latter. OCP-W_B is weighted higher than OCP-W_A (0·65 vs. 0·26), which is similar to Japanese rendaku, in that rendaku can be blocked if the second element already contains a voiced segment. *Lax/Monosyllabic W_B is weighted higher than *Lax/Monosyllabic W_A (0·64 vs. 0·35). In my observation, monosyllabic W_B nouns rarely constitute an independent phonological phrase in Yanbian, while monosyllabic W_A nouns may do so more frequently, explaining the less frequent occurrence of sai-sios. Finally, Post-obsTens/Comp plays no role in this model. Unlike *Tense/Vowel-W_A and *Tense/Son-W_A, which block the application of sai-sios tensification, Post-obsTens/Comp by definition does not work against RealiseMorpheme: even if Post-obsTens/Comp does not promote more tensification, RealiseMorpheme forces sai-sios, unless other factors against tensification are involved. Based on this result, we can assume that the relatively higher tensification rate after an obstruent coda is not because it is promoted by post-obstruent tensification, but because the relevant environment does not introduce any violation. Figure 11 shows the correlation between observed and predicted distributions. Two outlier combinations were excluded, due to the relatively small number of datapoints (36 and 35 respectively).Footnote ¹⁸ A linear regression analysis with observed distribution as the dependent variable and predicted distribution as the independent variable shows a significant result (est.=1·03, SE=0·05, t=19·32, p<0·0001).

Figure 11

Correlation between observed and predicted distributions of compound tensification, calculated using Hayes' Maxent Grammar Tool.

In sum, compound tensification in Yanbian, which is motivated by a junctural floating [tense] feature, shows considerable variation between speakers, and its application is significantly affected by various phonological factors: familiarity ratings, W_A/W_B stem length, W_A final-segment type and W_A/W_B consonant type. The higher the familiarity ratings, the higher the tensification rate. The longer the stem length, the lower the tensification rate. The probability of tensification is much greater after an obstruent coda than after a sonorant coda, which in turn shows a higher tensification rate than after a vowel. The tensification rate is also relatively lower when W_A or W_B contains either aspirated or tense consonants. As discussed in Zuraw (Reference Zuraw2011), all of these factors contribute probabilistically to the distribution of the occurrence of tensification. The overall distribution can be predicted by several weighted constraints with a substantial degree of accuracy. On the other hand, speaker age, gender and locality are not factors which have a clear effect on sai-sios distribution.

3.1 Wug test 1: existing word + nonce word

In wug test 1, W_A was an existing monosyllabic or disyllabic word with a sonorant or obstruent coda, while W_B was a nonce word with the structure CV.CV, CV.CʰV, CV.C*V or CV.RV (where C=a lax consonant). The onset of W_B was /p/, /t/, /k/, /c/ or /s/. The four W_A words and 40 W_B words, listed in (4), were combined, resulting in 160 test words.

1. (4)

   

Eleven Yanbian native speakers participated in the experiment. They were instructed to report whether the onset consonant of W_B is pronounced as tense, lax or either. The test items were sorted randomly and presented as a list, with spaces for marking the presence or absence of tensification, which was distributed to each speaker. Table XVIII shows the results. On the whole, tensification was reported more frequently when W_B contained medial lax or sonorant consonants than when it contained medial aspirated or tense consonants. The tensification rate was also lower for W_A sonorant codas than for obstruent codas. For obstruent codas, the tensification rate was lower when the W_A syllable count changed from 1 to 2, whereas it was more or less the same for sonorant codas, as seen in Fig. 12.

Figure 12

Compound tensification rate as a function of W_A coda type, W_A syllable count and W_B consonant type in wug test 1.

Table XVIII

Results of wug test 1.

A mixed effects logistic regression model was constructed, with tensification as the dependent variable (non-tensified as the reference category) and W_A coda type (obstruent, sonorant), W_A syllable count (1 or 2) and W_B consonant type (aspirated, tense, lax, sonorant) as the independent variables. W_A coda type and W_A syllable count were sum-coded, while backward difference coding was utilised for W_B consonant type: tense > aspirated > lax > sonorant. Random intercepts were set for item and subject. Table XIX shows the results.

An obstruent coda and a lower syllable count for W_A increase the tensification rate. Their interaction is significant. Tukey's multiple comparison test, using the glht function from the multcomp package (Hothorn et al. Reference Hothorn, Bretz, Westfall, Heiberger and Schuetzenmeister2013), showed that the obstruent coda–2 syllable combination results in significantly less tensification than the obstruent coda–1 syllable combination (est.=−1·04, SE=0·1800, z=−5·80, p=0·0001), while the sonorant coda–2 syllable combination is not significantly different from the sonorant coda–1 syllable combination (est.=−0·24, SE=0·15, z=−1·62, p=0·364). For W_B consonant type, tense consonants are not significantly different from aspirated consonants, which are in turn less tensified than lax consonants. Lax and sonorant consonants are not significantly different.

Table XIX

Results of a mixed effects logistic regression model (wug test 1).

3.2 Wug test 2: nonce word + existing word

A second wug test was conducted. This time W_A was a disyllabic nonce word with the structure RV.CVR, RV.CʰVR, RV.C*VR or CV.RVR, CʰV.RVR, C*V.RVR. W_B was an existing monosyllabic or disyllabic word. The twelve W_A words and six W_B words, listed in (5), were combined, resulting in 72 test words.

1. (5)

   

Ten Yanbian native speakers were instructed to report whether the onset consonant of W_B is pronounced as tense, lax or either, using the same procedure as wug test 1 (seven speakers were also participants in wug test 1). Table XX shows the results. When W_A contained a tense consonant, the tensification rate was very low, regardless of the position (24% for tense in the first syllable; 18% for tense in the second syllable) and the W_B syllable count (27% for monosyllabic W_B; 16% for disyllabic W_B). On the other hand, when W_A contained a lax consonant, the tensification rate was relatively high, regardless of the position (56% for lax in the first syllable; 56% for lax in the second syllable) and the W_B syllable count (54% for monosyllabic W_B; 58% for disyllabic W_B). Aspirated consonants showed a different pattern from tense and lax consonants: when an aspirated consonant appeared in the second syllable of W_A, the tensification rate was quite low (22%), but when it appeared in the first syllable of W_A, the tensification rate was around the same as lax (51%); there was no clear difference for the W_B syllable count (37% for monosyllabic W_B; 37% for disyllabic W_B).

Table XX

Results of wug test 2.

Although it is unclear if there is any locality effect in existing words, the effect seems to be found in wug compounds, in that they tend to resist tensification more strongly when W_A contains an aspirated consonant in the second syllable (i.e. when the aspirated consonant is adjacent to the site of tensification) than in the first syllable, as seen in Fig. 13.

Figure 13

Compound tensification rate as a function of W_A consonant position, W_A consonant type and W_B syllable count in wug test 2.

A mixed effects logistic regression model was constructed, with tensification as the dependent variable (non-tensified as the reference category) and the laryngeal features contained in W_A (aspirated, tense, lax), the position of the laryngeal features in W_A (first or second syllable) and W_B syllable count (1 or 2) as the independent variables. For W_A consonant type, backward difference coding was utilised (tense > aspirated > lax), while W_A consonant position and W_B syllable count were sum-coded. A random intercept was set for item and subject. Table XXI shows the results. Sai-sios tensification is less frequent if W_A contains tense consonants rather than aspirated consonants, and most frequent if it contains lax consonants. Tensification is observed more when the laryngeal features in W_A are in the first syllable. The interaction between W_A consonant and W_A C position is also significant, reflecting the strong locality effect observed in aspirated consonants. Tukey's multiple comparison test, using the glht function, showed that lax–first and tense–first are not significantly different from lax–second and tense–second respectively (est.=0·01, SE=0·27, z=0·04, p=1·000 for lax; est.=−0·58, SE=0·36, z=−1·63, p=0·575 for tense), whereas aspirated–first is significantly more tensified than aspirated–second (est.=−1·70, SE=0·32, z=−5·35, p=0·0001). On the other hand, W_B syllable count was not a significant factor.

Table XXI

Results of a mixed effects logistic regression model (wug test 2).

In sum, the two wug tests showed the OCP effect as well as the effect of coda type, mirroring tendencies observed in existing compound nouns: the probability of tensification is higher after an obstruent coda than after a sonorant coda, and the tensification rate is relatively lower when W_A or W_B contains either aspirated or tense consonants. On the other hand, a locality effect emerged in the experiment which was not clearly confirmed in the existing lexicon. The effect of stem length (the longer the stem, the lower the tensification rate) was partially observed in wug test 1 in W_A with an obstruent coda, but was not clearly observed in the other cases. A possible reason for this is that stem length or the number of syllables in a compound noun correlates with familiarity: more familiar compound words tend to be shorter, as seen in Fig. 14 (mean familiarity ratings are 3·86, 3·75, 3·43 and 2·87 for compound nouns of two, three, four and five syllables respectively). A linear regression analysis with familiarity as the dependent variable and size of compound nouns as the independent variable shows a significant result (est.=−0·25, SE=0·06, t=−4·38, p<0·0001). This may suggest that the syllable-count effect is in fact due to lexical familiarity. If so, the fact that the syllable-count effect did not clearly appear in the results of wug tests except for one case is understandable, since the novel compound words used in wug tests have by definition frequency zero. However, two mixed effects logistic regression models, one of which includes both W_A/W_B syllable length and familiarity as factors and the other which includes W_A/W_B syllable length but excludes familiarity, are significantly different according to an ANOVA() for a likelihood-ratio test (χ²=81·92, df=1, p<0·0001); furthermore, the model in which both W_A/W_B syllable length and familiarity are included is significantly different from one in which familiarity is included but W_A/W_B syllable length is excluded (χ²=128·57, df=2, p<0·0001). Given this, we can conclude that syllable count and familiarity are separate factors affecting sai-sios distribution, even though these two overlap at least partially.

Figure 14

Correlation between syllable count of compound nouns and averaged familiarity ratings.

The fact that the syllable-count effect was not clearly observed in the wug-test results may be because the association between shorter/longer words and more/less compound tensification is not straightforwardly explained as a ‘universal’ preference. Hence there is no reason to expect it to be generalised to the novel words. We will return to this question in §4.2 below.

4.1 Acoustic features

What is the cause of the co-occurrence restriction between two aspirated/tense consonants in Yanbian? Gallagher (Reference Gallagher2011) mentions that when the co-occurrence of multiple laryngeally marked segments in the stem or word is restricted in some languages, the restricted segments always share an articulatory or acoustic property, and she concludes that the feature shared by aspirated consonants and ejectives in Quechua is [long VOT]. The same observation was made by MacEachern (Reference MacEachern1999). Can our Yanbian data be explained in the same way? I conducted a preliminary acoustic analysis of the Yanbian laryngeal contrast.Footnote ¹⁹ Six words or phrases of the same type (Ca/C*a/Cʰa, with a high or low tone in the initial syllable) were recorded in isolation at a sampling rate of 44·1 kHz and 16 bits, resulting in a total of 54 items for each speaker (6 words×3 laryngeal categories×3 places of articulation).Footnote ²⁰ The recording was conducted in a quiet room, using a Sony linear PCM recorder PCM-D50 and a Sony microphone ECM-MS957, and was analysed in Praat (Boersma & Weenink Reference Boersma and Weenink2012). VOT was measured from the point of stop burst release to the onset of voicing in the following vowel. A list of the items recorded can be found in Appendix A.

Table XXII shows the mean VOT values in word-initial position based on the phonation types, taken from 61 Yanbian native speakers born between 1935 and 1992 (25 male, 36 female), among whom were twelve speakers who were subjects for our sai-sios survey. Yanbian aspirated consonants have a relatively long VOT (average 98 ms), while tense consonants have a much shorter VOT (average 17 ms). Unlike Quechua, where both aspirated consonants and ejectives have a similarly long VOT, Yanbian (and Korean in general) has a long VOT only for aspirated consonants. Tukey's multiple comparison test on a linear mixed model, using the lmer function from the lme4 package and the glht function from the multcomp package, with item and subject as random factors showed that the VOTs of the three laryngeal categories are all significantly different from one another (lax vs. aspirated: est.=−71·60, SE=2·39, z=−29·97, p<0·0001; tense vs. aspirated: est.=−81·32, SE=2·39, z=−33·98, p<0·0001; tense vs. lax: est.=−9·72, SE=2·39, z=−4·07, p=0·0001).

Table XXII

Mean VOT values (in ms) for phonation types in word-initial position in Yanbian. Standard deviation is given in parentheses.

The same is true for word-medial position. 54 trisyllabic words/phrases containing Ca/C*a/Cʰa in the penultimate syllable (6 words×3 laryngeal categories×3 places of articulation) were recorded in isolation and analysed for 14 Yanbian speakers (7 males, 7 females), using the same procedure mentioned above. The recorded items can be found in Appendix B.

Lax consonants are voiced intervocalically for some speakers but are voiceless for others. The VOT value for lax is based on data from speakers who pronounced lax as voiceless. As seen in Table XXIII, the VOTs of both lax and tense are shorter than aspirated, and as a whole the VOT values in word-medial position are shorter than those in word-initial position. Tukey's multiple comparison test on a linear mixed model with item and subject as random factors shows that the VOTs of lax and tense are significantly shorter than that of aspirated (est.=−35·39, SE=2·43, z=−14·58, p<0·0001 for lax; est.=−35·13, SE=2·30, z=−15·28, p<0·0001 for tense), while the VOTs of tense and lax are not significantly different (est.=0·26, SE=2·43, z=0·11, p=0·994). In addition, the closure duration of tense is significantly longer than that of aspirated and lax (est.=23·99, SE=4·55, z=5·27, p<0·0001 for aspirated; est.=115·73, SE=4·56, z=25·40, p<0·0001 for lax), and the closure duration of aspirated is significantly longer than that of lax (est.=−91·74, SE=4·56, z=−20·13, p<0·0001). Thus aspirated and tense consonants in word-medial position share similar acoustic correlates: they have comparably long silent periods, which is due to their long closure duration. The mean total of closure duration plus VOT is 195 ms for aspirated and 184 ms for tense, while it is 68 ms for lax. Figure 15 shows sample waveforms and spectrograms.

Figure 15

Sample waveforms and spectrograms from speaker KMH (female, born in 1992) for (a) /s*a.pʰa.ta/ ‘It is a sofa’ (CD (closure duration) 127 ms; VOT 59 ms) and (b) /a.p*a.ta/ ‘It is my father’ (CD 184 ms; VOT 9 ms).

Table XXIII

Mean VOT values and closure durations (in ms) for phonation types in wordmedial position in Yanbian. Standard deviation is given in parentheses.

This suggests that, in Yanbian, the feature shared between aspirated and tense consonants is [long non-modal voicing]. This interpretation does not contradict the findings of MacEachern (Reference MacEachern1999) and Gallagher (Reference Gallagher2011), but can cast light on the acoustic factor involved in the co-occurrence: having two relatively less sonorant sequences characterised by [long non-modal voicing] tends to be avoided. In our analysis, [non-modal voicing] includes frication, thus /s*/ is also the target of co-occurrence restrictions. As pointed out by Cho et al. (Reference Cho, Jun and Ladefoged2002), there is a significant duration difference between /s/ and /s*/ in word-medial position in both Seoul and Cheju Korean: /s*/ has a much greater duration than /s/. This supports the argument that /s*/ differs from /s/ in ‘long’ [non-modal voicing].

Furthermore, the duration of a vowel followed by a syllable with an aspirated or tense onset tends to be shorter than that followed by a syllable with a lax onset in Yanbian. Table XXIV and Fig. 16 show the mean vowel durations of V₁ in CV₁.CV₂ words in each environment, based on recordings of the words taken from four native speakers (three words each with twelve structures, giving 36 words for each speaker). A list of the recorded items can be found in Appendix C.

Figure 16

Mean vowel duration (ms) of V₁ in CV₁. CV₂ words as a function of initial and peninitial onset type.

Table XXIV

Mean vowel durations (ms) of initial syllable. O1=onset of the initial syllable; O2=onset of the peninitial syllable.

Tukey's multiple comparison test on a linear mixed model with item and subject as random factors showed that the vowel durations in a first syllable followed by a syllable with a lax or sonorant onset are significantly greater than those in one followed by a syllable with an aspirated or tense onset (lax vs. aspirated: est.=73·83, SE=7·28, z=10·15, p<0·001; sonorant vs. aspirated: est.=89·65, SE=7·28, z=12·32, p<0·001; tense vs. lax: est.=−68·63, SE=7·32, z=−9·38, p<0·001; tense vs. sonorant: est.=−84·45, SE=7·32, z=−11·55, p<0·001), while the vowel durations in a syllable followed by a tense onset vs. aspirated onset and those in a syllable followed by a sonorant onset vs. lax onset are not significantly different (tense vs. aspirated: est.=5·20, SE=7·35, z=0·71, p=0·894; sonorant vs. lax: est.=15·82, SE=7·24, z=2·19, p=0·127). These results indicate that the shorter vowel duration makes the two onsets closer, which may be an additional factor regulating compound tensification.

4.2 Simplex nouns

The next question with regard to the co-occurrence restriction on two laryngeally marked segments observed in compound words is whether or not the same pattern appears in simplex words as well. If there is such an effect in simplex nouns, this suggests that the OCP effect in compound words is generalised from simplex words. In this section, I examine the co-occurrence patterns of onset consonants in disyllabic and trisyllabic simplex nouns taken from the two Korean dictionaries mentioned in §1, as well as some words collected from Yanbian native speakers.Footnote ²¹ The total of simplex nouns collected this way is 1991 for disyllables and 884 for trisyllables. For the phonotactics of Korean monosyllabic stems, see Ito (Reference Ito2007).

Table XXV shows the distribution of consonant types in the onsets of initial and peninitial syllables in disyllabic and trisyllabic simplex nouns. Interestingly, the co-occurrence of two aspirated consonants is strongly underrepresented, while that of two tense consonants is overrepresented in the lexicon.Footnote ²² The co-occurrence of aspirated and tense onsets is also underrepresented, but this does not reach statistical significance, and the co-occurrence of lax (initial)-tense (peninitial) is also underrepresented.

Table XXV

Co-occurrence of onset types in initial and peninitial syllables in disyllabic and trisyllabic simplex nouns. The observed/expected value is given in parentheses. Significant correlations (p<0·05) are indicated by *·(χ²=45·72, df=9, p=<0·0001).

Examples are given in (6).

1. (6)

   

Tables XXVI and XXVII show the onset distribution of peninitial and final syllables, and initial and final syllables in trisyllabic simplex nouns. We do not show χ² statistics here, since aspirated/tense onsets rarely appear in the final syllable (respectively 23 and 24 occurrences) and there are many cells with a value below five. Based on the observed/expected values, some combinations seem to be underrepresented (aspirated (peninitial)–aspirated (final), tense (initial)–aspirated (final), etc.), while others seem to be overrepresented (aspirated (initial)–aspirated (final), aspirated (initial)–tense (final), etc.). However, it is difficult to draw any conclusions about the laryngeal OCP effect on the basis of this limited data.

Table XXVI

Co-occurrence of onset types in peninitial and final syllables in trisyllabic simplex nouns. The observed/expected value is given in parentheses.

Table XXVII

Co-occurrence of onset type in initial and final syllables in trisyllabic simplex nouns. The observed/expected value is given in parentheses.

Examples are given in (7) and (8).

1. (7)

   

2. (8)

   

How can we explain the apparent contradiction between the OCP effect observed in compound tensification (recall Tables X and XII) and the overrepresentation of the co-occurrence of two tense onsets in initial and peninitial syllables of simplex words (Table XXV)? According to Gallagher (Reference Gallagher2010, Reference Gallagher2012), long-distance restrictions on laryngeal features fall into one of three patterns: assimilation, dissimilation and mixed. In ‘assimilatory’ and ‘dissimilatory’ languages roots with heterorganic and homorganic pairs of consonants show the same co-occurrence patterns, while in ‘mixed’ languages heterorganic and homorganic pairs pattern differently. The three patterns are shown in (9), where TT and TK denote a pair of homorganic and heterorganic consonants respectively, and C' is a laryngeally marked consonant.

1. (9)

   

In our database of simplex words (initial and peninitial onsets), marked–marked structures (CʰV(C).CʰV(C), CʰV(C).C*V(C), C*V(C).CʰV(C) and C*V(C).C*V(C)) are not biased to homorganic consonant pairs, as seen in Table XXVIII.

Table XXVIII

Initial and peninitial onset type in simplex words and the ratio of homorganic pairs.

Homorganic consonant pairs represent 34% of marked–marked structures, more or less the same as for marked–unmarked and unmarked–marked structures (CʰV(C).CV(C), CV(C).CʰV(C) and CV(C).C*V(C), C*V(C).CV(C)), and for unmarked–unmarked structures (CV(C).CV(C)). In addition, more detailed examination of words with a marked–marked structure (Table XXIX) shows that a homorganic effect is observed when both onsets are aspirated, but heterorganic pairs are more frequently observed in other cases. Given this, the Yanbian data cannot straightforwardly be classified as a ‘mixed’ pattern.

Table XXIX

Initial and peninitial onsets in simplex words and the ratio of homorganic pairs: marked–marked structure.

As mentioned above, I consider that the overrepresentation of two tense onsets in initial and peninitial syllables in simplex words is due to a long-distance regressive assimilation process that anticipates the creaky voice of the following syllable: the glottal constriction in the second syllable is assimilated to the first syllable. This is due to harmony between the vowel of the initial CV and the vowel of the following C*, which underlies the consonantal harmony: CV.C* > C.C* > C*.C*. I assume that the change from CV.C*V to C*V.C*V can happen especially easily in Yanbian, given that the VOTs of lax and tense consonants are quite close in Yanbian (27 ms for lax and 17 ms for tense), and that creaky voice (lower H1 – H2, where H1 – H2 is the difference in amplitude between the first and second harmonics) is the major distinctive cue in Yanbian.Footnote ²³ Table XXX shows the mean H1 – H2 values in vowel-onset position of the initial syllable taken from 61 speakers based on the recorded data in Appendix A. H1 – H2 was measured in the first 25 ms by using Fast Fourier Transform spectra. Tukey's multiple comparison test on a linear mixed model with item and subject as random factors showed that the H1 – H2 values of lax and tense are significantly lower than that of aspirated (lax vs. aspirated: est.=−1·19, SE=0·15, z=−7·76, p<0·0001; tense vs. aspirated: est.=−7·30, SE=0·16, z=−46·98, p<0·0001), and that the H1 – H2 value of tense is significantly lower than that of lax (est.=−6·11, SE=0·15, z=−39·46, p<0·0001).

Table XXX

Mean H1 – H2 values (in dB) in Yanbian. Standard deviation is given in parentheses.

In fact, there are many words which display this assimilation diachronically, e.g. MK /kos.kal/ > /ko.k*al/ > [k*o.k*al] ‘Buddhist priest's hood’ and MK /tuk.tə.pi/ > /tut.kə.pi/ (metathesis) > /tu.k*ə.pi/ > [t*u.k*ə.pi] ‘toad’, but no instances of progressive assimilation (C*V.CV > C*V.C*V) are observed. Recall that tense consonants in Contemporary Korean originate from consonant clusters in Middle Korean. The fact that the co-occurrence of lax initial and tense peninitial consonants is underrepresented in simplex words also supports this hypothesis.Footnote ²⁴

The fact that W_A C*V.C*V is dispreferred in compound words may also be because the compound tensification itself is a progressive process: the floating [tense] feature is between the two morphemes in the input and docks onto the first consonant of W_B in the output, since any coda in W_A is unreleased, and hence cannot be tensified, due to a general syllable-structure constraint of Korean phonology. Thus the floating [tense] feature is realised on the second morpheme, and its association can be characterised as progressive. The OCP constraint on W_B blocks tensification, since the tensification here is not the one triggered by the harmony to a following tense consonant in W_B. Thus, Yanbian Korean exhibits two co-occurrence patterns in long-distance restrictions on laryngeal features (assimilation and dissimilation), based not on the homorganic/heterorganic patterns of the onset pairs, but on the presence or absence of the environment for regressive assimilation.Footnote ²⁵

In any case, the overrepresentation of the co-occurrence of two tense consonants in simplex nouns and the fact that the underrepresentation of the co-occurrence of aspirated and tense onsets does not reach statistical significance suggest that OCP[tense] and OCP[aspirated-tense] in compound nouns are emergent patterns without clear lexical precedents. Furthermore, in §3.2 I found that a locality effect which has not yet managed to shape the lexicon in a measurable way emerged in the wug test, whereas the syllable-count effect, which is observed in the lexicon, was not clearly apparent in the wug test, except in one case. These results show that Yanbian speakers have knowledge of typological universals that goes beyond what can be inferred from the data in the lexicon. As mentioned above, sai-sios originates historically from a genitive marker in Middle Korean, and hence is assumed here to have appeared regardless of the phonological properties of W_B. However, the attachment of the genitive marker /-s/ in Middle Korean, which would originally have been morphologically and phonologically transparent, was obscured, due to historical sound changes. These changes have eliminated the traces of sai-sios in Contemporary Korean, except for the tensification of the first consonant of W_B, making the compound tensification process opaque. Thus a process that was originally morphologically and phonologically transparent, but became opaque due to sound change, has been reshaped and reinterpreted, and phonological factors have arisen in conditioning sai-sios. On the one hand, speakers have restructured the generalisation on the basis of the available synchronic data, and the application of sai-sios (RealiseMorpheme) may interact with lower-weighted constraints on tense consonants in the lexicon, such as faithfulness of [−tense] in longer and/or less familiar words. On the other hand, ‘universal’ preferences also emerge for our Yanbian speakers, such as OCP[tense] and OCP[aspirated-tense] with a locality effect, which are evidently not present in the linguistic data of their native language. The syllable-count effect, which overlaps at least partially with the familiarity effect, cannot be straightforwardly explained as a universal preference, and hence is not actively incorporated in novel words. This paper thus provides further evidence for the existence of analytic biases (or ‘overlearning’) as well as ‘underlearning’ (Wilson Reference Wilson2003, Moreton Reference Moreton2008, Zhang & Lai Reference Zhang and Lai2008, Becker et al. Reference Becker, Ketrez and Nevins2011), not only based on wug tests, but also through the analysis of historical restructuring of the lexicon.

4.3 Alternatives

Finally in this section, I investigate whether the OCP effect observed in Yanbian compounds (as well as simplex words) can be motivated by additional factors.

First, the restriction on the co-occurrence of aspirated and tense in Yanbian may be due to articulatory factors. Following Kim (Reference Kim1965) and Halle & Stevens (Reference Halle and Stevens1971), tense can be interpreted as [constricted glottis], and aspirated as [spread glottis]. The feature [constricted glottis] is at the opposite pole from [spread glottis] on the VOT parameter, in that glottal opening is minimal vs. maximal, whereas the state of glottis for lax and sonorant onsets, as well as for vowels, is assumed to be intermediate. Given this, a significantly large amount of articulatory effort is presumably required not only for two extreme glottal states separated by a vowel but also for two opposing glottal states (see Gallagher Reference Gallagher2013 and Walter Reference Walter2007 for relevant discussion).

Another possible explanation is that the feature shared between aspirated and tense is higher F0 in the following vowel, which is observed in Seoul and Kyengsang Korean (e.g. Kenstowicz & Park Reference Kenstowicz and Park2006, Silva Reference Silva2006, Lee & Jongman Reference Lee and Jongman2012) and which could thus give rise to an OCP effect on high tones or raised F0 as the factor grouping these two laryngeal classes. We measured F0 values in both vowel-onset position (the second pulse of the vowel following the target stop) and vowel-medial position (the medial position of the entire vowel duration) of the first syllable, based on the data in Appendix A from 14 speakers (7 males, 7 females), taken from citation forms. As seen in Table XXXI, the F0 difference between tense/aspirated and lax is quite small in Yanbian: around 1–20 Hz for vowel-onset position and 4–10 Hz for vowel-medial position. Two linear mixed models were run to test the F0 patterns, using the lmer function in R with item and subject as random factors: for one case, the dependent variable is vowel-onset F0 and for the other case it is vowel-medial F0; in both cases, independent variables are gender (male, female), tone (initial-syllable high, low) and laryngeal class (aspirated, tense, lax). Tukey's multiple comparison test on a linear mixed model with vowel-onset F0 as the dependent variable showed that gender and tone are significant factors (gender: est.=−105·70, SE=17·07, z=−6·19, p<0·0001; tone: est.=−34·41, SE=1·09, z=−31·69, p<0·0001). For laryngeal class, aspirated is significantly higher than lax and tense, and tense is significantly higher than lax (lax vs. aspirated: est.=−13·03, SE=1·32, z=−9·88, p<0·0001; tense vs. aspirated: est.=−7·00, SE=1·33, z=−5·28, p<0·0001; tense vs. lax: est.=6·02, SE=1·33, z=4·53, p<0·0001). In vowel-medial position, gender and tone are again significant factors (gender: est.=−98·92, SE=14·17, z=−6·98, p<0·0001; tone: est.=−47·36, SE=1·28, z=−37·11, p<0·0001). The F0 values of aspirated are significantly higher than lax, while the F0 values of aspirated vs. tense and tense vs. lax are not significantly different (lax vs. aspirated: est.=−6·71, SE=1·55, z=−4·33, p=0·0001; tense vs. aspirated: est.=−3·06, SE=1·56, z=−1·97, p=0·1208; tense vs. lax: est.=3·65, SE=1·56, z=2·34, p=0·0509). As a whole, the effects of gender and tone are much stronger in magnitude than laryngeal class, as shown in the estimates. This is understandable, given the fact that Yanbian has a distinctive pitch accent (i.e. HL and LH are contrastive), which suggests that the relatively small differences of F0 between aspirated, tense and lax are unlikely to override the underlying tones to affect the decision on tensification. Thus an analysis involving the OCP applying to H tones or raised F0 probably does not yield the proper generalisation.

Table XXXI

Mean F0 values (in Hz) of each laryngeal category in Yanbian. Standard deviation is given in parentheses.