3.1 Against linear order and sonority relationship

Due to the inconsistency as to which consonant is subject to deletion or preservation on the surface, it seems at first glance hard to systematically explain the mechanism of CCS in the isolation forms of Korean nouns. One might argue for a linear order approach or a sonority relationship approach. However, neither of them works for CCS in Korean.

First, the linear order between two input consonants does not provide a consistent account since it is not always C₁ or C₂ that is preserved or deleted. In some clusters like /lk/ (e.g. /talk/ [tak] ‘chicken’), C₁ is deleted, while C₂ is preserved (9a), suggesting that Anchor-IO (7) dominates Contiguity-IO (8). In other clusters like /lp/ (e.g. /jʌtʌlp/ [jʌ.tʌl] ‘eight’), the conflicting ranking is observed (9b), indicating that Anchor-IO is dominated by Contiguity-IO (definitions are based on Kager 1999).

The discrepancy between /lk/ and /lp/ contradicts Whitman (1985) and Young-mee Yu Cho (1988) that the clusters /lk/, /lp/, and /lp^h/ pattern together, either deleting /l/ in Seoul Korean or preserving it in Kyungsang Korean.

Second, the sonority relationship between C₁ and C₂ also fails to account for CCS in Korean. It is true, though, that sonority is an important factor that accounts for certain shapes of output forms in phonology.

There are languages in which underlying consonant clusters fully surface in a way that they are syllabified through vowel epenthesis, as shown in the examples from Côté (2000) for Cairene Arabic and Chaha (5b). Crucially, the position of the epenthetic vowel may depend on the sonority of the consonants so that low sonority onsets and high sonority codas are created, as in the Vediceto dialect of Italian.

Sonority has also been investigated as a motivation for word-final cluster simplification via deletion. Given that falling sonority is universally preferred from a peak, Côté (2000) shows how CCS occurs when final clusters are of increasing sonority in Québec French. In obstruent+liquid clusters, it is invariably the second consonant that gets deleted, as in /livr/ [liv], *[lir] ‘deliver-pres’ and /sufl/ [sʊf], *[sʊl] ‘blow-pres’ (Côté 2000: 234). This is because the general faithfulness constraint Max-C is dominated by a markedness constraint that militates against rising sonority and the positional faithfulness constraint Max-C/V_ (‘no consonant deletion immediately after a vowel’). In other word, what motivates CCS is sonority, but how it is embodied is determined by the linear order of the consonants, in the way that the consonant that does not immediately follow a vowel undergoes deletion. Preference for the consonant in proximity to a vowel is in agreement with Wilson’s (2001) argument for first consonant deletion in intervocalic position (i.e. VC₁C₂V VC₂V) in Diola-Fogny and West Greenlandic.

However, sonority relationship between the two members of a consonant cluster does not provide a unified explanation for Korean nouns. Sometimes less sonorous consonants delete (e.g. /jʌtʌlp/ [jʌ.tʌl] ‘eight’), which would be predicted by the preference for high sonority codas (Vennemann 1988), while other times more sonorous consonants delete (e.g. /talk/ [tak] ‘chicken’).

In sum, neither the order between two input stem-final consonants nor their sonority relationship is uniformly decisive in the different CCS patterns in the isolation forms of nouns in Korean. Arguing against the two theoretically possible accounts, I provide a comprehensive analysis of isolation forms in Section 7.2.

4.1 Previous research on variation in inflected forms

There have been attempts to account for variation in inflected forms of Korean nouns. However, it is noteworthy that there has also been an asymmetry in the amount of attention that conservative forms and innovative forms have received in the literature. While much previous research focused on the emergence of innovative forms, conservative forms were taken for granted since they are fully faithful to input forms. This eluded a complete understanding of variation with important questions, such as whether different clusters prefer conservative forms or innovative forms, and if so, why. In Section 7.1, this paper fills this gap.

In previous studies, some researchers have explained inflectional variation as phonological opacity, arguing that there is no phonological motivation to delete one of the two consonants in a particular context. Jin-hyung Kim (2005) notes that complex codas are forbidden in Korean and that this leads to CCS both in isolation forms (e.g. /kaps/ [kap] ‘price’) and inflected forms with a consonant-initial suffix attached (e.g. /kaps-to/ [kap.to] ‘price too’), as in (1). This alternation is phonologically predictable by the language-specific restriction on syllabification. She points out that CCS also occurs in a context where it would not be predicted, such as in inflected forms with a vowel-initial suffix attached (e.g. /kaps-i/ [ka.pi] ‘price-nom’). Even though the standard form (e.g. /kaps-i/ [kap.si] ‘price-nom’) does not violate the syllable well-formedness restriction, the unexpected output form with CCS does occur. She accounts for such opacity by claiming that the younger generation has restructured underlying forms from the ones with a consonant cluster /CC/ (i.e. complex) to the ones with a single consonant /C/ (i.e. simplex). For example, /kaps/ ‘price’ has undergone leveling in a set of uninflected and inflected forms, as in {kap, ka.pi, ka.pɨl, ka.pɨn…}. Furthermore, she reports that younger speakers never preserve both elements in consonant clusters, thus distinguishing their speech from older speakers’ production. However, Jin-hyung Kim’s (2005) argument for total leveling toward innovative forms is extreme in some sense, since both conservative and innovative forms do occur in the younger generation’s speech. This is supported by the results of a production experiment of the present study.

Others have argued for a relation between innovative forms and isolation forms. Kenstowicz (1996) accounts for variation in inflected forms by relating output forms at the two different morphological levels: isolation and inflection. First, at the isolation level, the markedness constraint *Complex outranks the faithfulness constraint Parse-C (i.e. *Complex >> Parse-C). For /talk/ ‘chicken’, for example, *Complex penalizes the candidate with a complex coda *[talk], and thus the simplified candidate [tak] is chosen as the optimal output, at the expense of parsing every element in input to output. He claims that this ranking also works in evaluating candidates at the level of inflection, especially for the conservative forms (e.g. /talk-i/ [tal.ki] ‘chicken-nom’), as it satisfies *Complex by syllabifying C₂ /k/ as the onset to /i/. Additionally, there are innovative forms, as in [ta.ki]. What plays a crucial role to select innovative forms as an optimal output form is the Base-Identity constraint, which requires a similarity between the Base (i.e. isolation form) and derived forms, by the ranking of *Complex >> Base-Identity >> Parse-C. In other words, innovative forms satisfy not only *Complex but also Base-Identity because the consonant that survives in the innovative form is simplex but not complex via deletion, and it is the very consonant that appears in the isolation form, as in [tak]. The Base-Identity-based analysis resolves the opacity issue raised above since what seems to be phonologically opaque is in fact predicted by other morphologically related forms. The present study follows Kenstowicz’s (1996) Base-Identity analysis of innovative forms in Korean nominal inflection. However, it still remains mysterious why one over the other consonant is deleted or preserved even in isolation forms, and under what mechanisms the quality related to CCS in both isolation and innovative inflection is determined. This is discussed in Sections 7.2 and 7.3.

5.1 Participants

Participants were recruited using social media. Forty-four native speakers of Korean who were at the age of 19 or in their 20’s at the time of participation participated in the production experiment (mean age: 22.48; range: 19-28). There were 22 male speakers and 22 female speakers. Twenty-three participants were native speakers of Seoul Korean who were from Seoul and Kyunggi area, and the other 21 were native speakers of Kyungnam Korean from Busan and South Kyungsang area. Ten of them participated in the in-person format, whereas the other 34 participated in the online format. Participants were paid compensation that is commensurate with $5 for their time.

5.2 Stimuli

Eleven native Korean monosyllabic nouns and one disyllable numeral that occur with one of the six consonant clusters /ps/, /ks/, /lp/, /lk/, /ls/, and /lm/ were selected for target noun stems. For the sake of convenience, the clusters were grouped into three types, based on the [±sonorant] feature of the elements, whether they are obstruents (Obs.) or sonorants (Son.): “Obs.+Obs.” for /ps/, /ks/, “Son.+Obs.” for /lp/, /lk/, /ls/, and “Son.+Son.” for /lm/. Although the cluster “Son.+Son.” has only one cluster that falls into this category (i.e. /lm/), which could lack representativeness, there is uniqueness of the cluster /lm/ that has something to do with morphology: it is in fact not a monomorphemic cluster but a combination of part of the verb stem /l/ from al- ‘to know’ or sal- ‘to live’ and the noun-forming morpheme -m. This will be elaborated in Section 7.2.4.

Thirty-six inflected forms of target items were prepared (= 12 nouns stems × 3 suffixes), as presented in Table 5: one noun stem with /ps/ (kaps ‘price’), three with /ks/ (saks ‘wage’, moks ‘share’, nʌks ‘soul’), one with /lp/ (jʌtʌlp ‘eight’), four with /lk/ (talk ‘chicken’, salk ‘leopard cat’, cʰilk ‘arrowroot’, hɨlk ‘soil’), one with /ls/ (kols ‘way’), and two with /lm/ (alm ‘knowledge’, salm ‘life’), each of which was followed by the three vowel-initial inflectional suffixes -i (nom), -ɨl (acc), and -ɛ (dat/loc). The inflected forms of target noun stems were embedded in the medial position of present tense declarative sentences.

There were also fillers, which included five stem-final consonant clusters /lpʰ/, /ltʰ/, /lh/, /nc/, and (/nh/) in verb stems, as in /ɨlpʰ-/ ‘to recite’, /hultʰ-/ ‘to skim through’, /alh-/ ‘to suffer’, /anc-/ ‘to sit’, and (/k’ɨnh-/ ‘to quit’) in three verbal inflections.⁴ Fillers also ended in a consonant cluster, as target noun stems did. This could have explicitly led to participants’ overall attention to consonant clusters, as a reviewer suggested. However, I chose fillers with a consonant cluster because if fillers with a simplex consonant had been used, it would have provided participants with a more stark contrast between simplex (fillers) and complex (target items) consonants.

Fillers were embedded in declarative sentences in a random order as were the target noun stems. In this way, 15 sentences for fillers were prepared (= 5 verb stems × 3 inflections). Combining all target items and fillers, 39 sentences were prepared for the in-person format of the experiment: (9 target noun stems excluding those with /lp/ and /lm/ × 3 suffixes) + (4 fillers excluding those with /nh/ × 3 verbal inflections). For the online format, 51 sentences were prepared: (12 target noun stems including those with /lp/ and /lm/ × 3 suffixes) + (5 fillers including those with /nh/ × 3 verbal inflections). (See the supplementary file.)

5.3 Procedure

The 51 sentences that contain target items and fillers were written in Korean orthography and prepared in a random order prior to the experiment. Experimental sentences in written forms were presented to participants.⁵ For the in-person format of the experiment, the 39 sentences (excluding the target items with /lp/ and /lm/ and the fillers with /nh/) were printed on a sheet of paper and shown to participants. For the online format, the 51 sentences (including the target items with /lp/ and /lm/ and the fillers with /nh/) were presented on the participants’ computer screen using the screen-sharing function that is built in the Zoom web conferencing platform. For both formats of the experiment, participants were asked to read aloud the sentences with three repetitions in a natural tone and at a normal speed as they spoke in daily conversations, either in the Seoul dialect or in the Kyungnam dialect depending on the areas they are originally from.

For the in-person format, recordings took place by using a Zoom H2n Handy Recorder in a quiet room on a university campus. The recorded files were automatically saved as a Waveform Audio File Format (.wav). The online format was conducted by having both the experimenter’s and the participants’ videos turned off throughout the recordings in order to obtain only audios, excluding videos that were unnecessary to the experiment. Participants’ audio was recorded by using their computer’s microphone and Zoom’s built-in recording function. Since the recorded files were automatically saved by Zoom as an Audio-only MPEG-4 file (.m4a), they were converted by the experimenter to a Waveform Audio File Format (.wav) after all recording sessions were complete, in order to make the files readable in Praat.

This resulted in reading and recording 6,372 sentences that consist of 1,170 recorded tokens from the in-person format (= 39 sentences × 10 participants × 3 repetitions) and 5,202 recorded tokens from the online format (= 51 sentences × 34 participants × 3 repetitions). Out of the 1,170 recorded tokens from the in-person format, 810 tokens were examined (= 9 target noun stems × 3 suffixes × 10 participants × 3 repetitions), excluding the 360 tokens of fillers. Among the 5,202 recorded tokens from the online format, 3,672 tokens were examined (= 12 target noun stems × 3 suffixes × 34 participants × 3 repetitions), excluding 1,530 fillers. In sum, a total of 4,482 tokens (= 810 tokens from the in-person format + 3,672 tokens from the online format) were analyzed.

5.4 Acoustic analysis

Recordings were analyzed acoustically using Praat (Boersma & Weenink 2021, version: 6.1.54) as well as impressionistically in order to overcome inconsistency, if any, of the sound files that were recorded via different means (i.e. with a Zoom H2n Handy Recorder and over Zoom). In what follows, I provide criteria of the acoustic analysis of two output forms (i.e. conservative or innovative) with some samples. Below spectrograms, there are three levels: the uppermost level is for the sentence that was read by a participant, the one in the middle is for the input form of the inflected noun, and the bottom shows segmentation of the inflected form.

First, for /ps/, /ks/, and /ls/, the output forms were either with an [s] or not. If there was a frication noise, this indicates the presence of [s] in the output form (left in Figure 1). Otherwise, it suggests the absence of [s] (right in Figure 1).

Figure 1

/kaps-i/ [kap.si] (conservative) vs. [ka.pi] (innovative) ‘price-nom’.

Second, the output forms of /lp/ were either with a [p] or not. Thus, the criterion was whether there is a stop closure and/or a release burst for [p] (left in Figure 2) or not (right in Figure 2).

Figure 2

/jʌtʌlp-i/ [jʌ.tʌl.pi] (conservative) vs. [jʌ.tʌ.li] (innovative) ‘eight-nom’.

Lastly, the output forms of /lk/ were [lk] or [k], and those of /lm/ were [lm] or [m]. Therefore, the criterion was whether there is an [l] or not. Miyeon Ahn (2017) shows the trajectories of the first formant (F₁) and the second formant (F₂) of post-vocalic /l/ in Korean under the assumption that the formant frequencies of the lateral approximant are affected by preceding vowels due to a coarticulatory effect; for example, if /l/ is preceded by the high vowel /i/ (e.g. /il/ in /cʰilk-i/ ‘arrowroot-nom’), F₂ increases slightly (as the tongue moves slightly forward from /i/ to /l/), whereas there is not much change in F₁. Following the criterion, if such a change in spectral properties was observed in the present spectrograms, this indicates the presence of [l], as in [cʰil.ki] (left in Figure 3). Otherwise, it suggests the absence of [l], as in [cʰi.ki] (right in Figure 3).

Figure 3

/cʰilk-i/ [cʰil.ki] (conservative) vs. [cʰi.ki] (innovative) ‘arrowroot-nom’.

Based on the above analysis, the number of conservative forms (with stem-final consonant clusters intact) and the number of innovative forms (with simplified stem-final consonant clusters) were counted, and the percentages were calculated. Results of the production experiment are presented in Section 6.

6.1 Results of raw data

In the raw data, there were three output consonant types: conservative, innovative, and other forms. Conservative forms are those with both input consonants preserved (e.g. /kaps-i/ [kap.si] ‘price-nom’), whereas innovative forms are those with one of the input consonants being deleted and the other being preserved, which is produced even with different laryngeal features (e.g. /talk-i/ [ta.k’i] (tense) and [ta.kʰi] (aspirated) as well as [ta.ki] (lenis)) since this is not considered to be important. Other forms include output forms with an unexpected consonant that does not occur in the corresponding isolation forms. For /jʌtʌlp/ ‘eight’, for example, [l] survives in the isolation form [jʌtʌl], but there were instances in the recorded tokens where the other consonant [p] appeared in the suffixed forms, as in [jʌ.tʌ.pi] ‘eight-nom’. In addition, there were cases where the non-etymological consonant [s] is inserted, as in /talk-i/ [tak.si] ‘chicken-nom’, which were also included in the “other” category. (See Ji Yea Kim 2019 for [s]-insertion in Korean.) The “other” category was excluded from both the raw data analysis and the statistical analysis since it is rare and out of the scope of this study. (See Section 6.1.1 for more details.)

6.1.1 Clusters

Percentages of conservative, innovative and other forms across the six stem-final consonant clusters were calculated by dividing the raw count of each of the three output types by the total number of output forms produced by the participants (% rounded to the nearest tenth). They are shown in Table 6 and Figure 4.

Table 6

Percentages (%) of output forms by clusters. (Raw counts are provided in parentheses, and dominant output consonant types are highlighted.)

	Consonant clusters
Output types	/ps/	/ks/	/lp/	/lk/	/ls/	/lm/
Conservative	99.0 (392)	75.0 (891)	10.8 (33)	10.2 (162)	12.6 (50)	97.2 (595)
Innovative	1.0 (4)	25.0 (297)	84.6 (259)	86.8 (1375)	82.3 (326)	2.8 (17)
Other	0.0 (0)	0.0 (0)	4.6 (14)	3.0 (47)	5.0 (20)	0.0 (0)

Figure 4

Percentages (%) of output forms by clusters.

Interestingly, a stark dichotomy was observed by consonant clusters. Conservative forms were dominant for /ps/, /ks/, and /lm/, while innovative forms were favored for /lp/, /lk/, and /ls/. Note that, as a reviewer suggested, it is highly likely that for some speakers noun stems with final /lp/, /lk/, and /ls/ have been lexicalized as the ones with a simplex consonant: /lp/ and /ls/ to /l/ (e.g. /jʌtʌlp/ > /jʌtʌl/ ‘eight’, /kols/ > /kol/ ‘way’) and /lk/ to /k/ (e.g. /talk/ > /tak/ ‘chicken’). This is supported by /tols/ > /tol/ ‘first birthday’. (See 7.1.) One might claim that such restructuring in underlying representations (URs) would not predict conservative forms, since the restructured URs are no longer with a consonant cluster but with a simplex final consonant. However, conservative forms did occur to a less extent (10.2–12.6%), though, for these clusters. This might be due to a spelling effect, since written forms were used in the current experiment.

It seems reasonable at this point to roughly group the clusters into two, based on the results and the [±sonorant] feature of the consonantal elements of the clusters: conservative-dominant (“Obs.+Obs.” and “Son.+Son.”) and innovative dominant (“Son.+Obs.”). Since the percentages of output forms that fall into the “other” category are negligibly low (4.6% for /lp/, 3.0% for /lk/, and 5.0% for /ls/) and they are out of the scope of this study, “other” was excluded, and only “conservative” and “innovative” forms were analyzed and will be discussed henceforth.

6.1.2 Clusters and suffixes

Percentages of conservative and innovative forms by the six stem-final consonant clusters and the three vowel-initial inflectional suffixes were calculated, as shown in Table 7 and Figure 5.

Table 7

Percentages (%) of output forms by clusters and suffixes. (Raw counts are provided in parentheses, and dominant output consonant types are highlighted.)

	Suffixes
	nom -i						acc -ɨl						dat/loc -ɛ
	Consonant clusters						Consonant clusters						Consonant clusters
	/ps/	/ks/	/lp/	/lk/	/ls/	/lm/	/ps/	/ks/	/lp/	/lk/	/ls/	/lm/	/ps/	/ks/	/lp/	/lk/	/ls/	/lm/
Conservative	97.7 (129)	74.7 (296)	11.2 (11)	13.5 (69)	12.9 (16)	97.5 (199)	100 (132)	77.5 (307)	8.4 (8)	8.2 (42)	14.1 (18)	97.5 (199)	99.2 (131)	72.7 (288)	14.1 (14)	9.9 (51)	12.9 (16)	96.6 (197)
Innovative	2.3 (3)	25.3 (100)	88.8 (87)	86.5 (442)	87.1 (108)	2.5 (5)	0.0 (0)	22.5 (89)	91.6 (87)	91.8 (469)	85.9 (110)	2.5 (5)	0.8 (1)	27.3 (108)	85.9 (85)	90.1 (464)	87.1 (108)	3.4 (7)

Figure 5

Percentages (%) of output forms by clusters and suffixes.

The general patterns by clusters and suffixes are identical to those by clusters only. That is, conservative forms were favored for /ps/, /ks/, and /lm/, while innovative forms were dominant for /lp/, /lk/, and /ls/. These patterns were observed regardless of whether the suffix was nom -i, acc -ɨl, or dat/loc -ɛ.

6.1.3 Clusters and dialects

Percentages of conservative and innovative forms by the six stem-final consonant clusters and the two regional dialects are presented in Table 8 and Figure 6.

Table 8

Percentages (%) of output forms by clusters and dialects. (Raw counts are provided in parentheses, and dominant output consonant types are highlighted.)

	Dialects
	Seoul						Kyungnam
	Consonant clusters						Consonant clusters
	/ps/	/ks/	/lp/	/lk/	/ls/	/lm/	/ps/	/ks/	/lp/	/lk/	/ls/	/lm/
Conservative	99.0 (205)	77.0 (478)	14.3 (22)	11.8 (96)	15.3 (31)	95.7 (310)	98.9 (187)	72.8 (413)	8.0 (11)	9.2 (66)	11.0 (19)	99.0 (285)
Innovative	1.0 (2)	23.0 (143)	85.7 (132)	88.2 (720)	84.7 (172)	4.3 (14)	1.1 (2)	27.2 (154)	92.0 (127)	90.8 (655)	89.0 (154)	1.0 (3)

Figure 6

Percentages (%) of output forms by clusters and dialects.

Output forms by clusters and dialects were overall identical to those by clusters only, and by clusters and suffixes: conservative forms were dominant for /ps/, /ks/, and /lm/, whereas innovative forms were for /lp/, /lk/, and /ls/. Such a dichotomy was found regardless of participants’ dialects (i.e. Seoul or Kyungnam).

6.1.4 Clusters and gender

Percentages of conservative and innovative forms by the six stem-final consonant clusters and participants’ gender were calculated, as shown in Table 9 and Figure 7.

Table 9

Percentages (%) of output forms by clusters and gender. (Raw counts are provided in parentheses, and dominant output consonant types are highlighted.)

	Gender
	Male						Female
	Consonant clusters						Consonant clusters
	/ps/	/ks/	/lp/	/lk/	/ls/	/lm/	/ps/	/ks/	/lp/	/lk/	/ls/	/lm/
Conservative	98.0 (194)	74.1 (440)	14.8 (22)	11.6 (91)	16.1 (30)	95.4 (292)	100.0 (198)	75.9 (451)	7.7 (11)	9.5 (71)	10.5 (20)	99.0 (303)
Innovative	2.0 (4)	25.9 (154)	85.2 (127)	88.4 (696)	83.9 (156)	4.6 (14)	0.0 (0)	24.1 (143)	92.3 (132)	90.5 (679)	89.5 (170)	1.0 (3)

Figure 7

Percentages (%) of output forms by clusters and gender.

The overall distribution of conservative and innovative forms by clusters and gender was identical to the findings by clusters only, and by clusters with either suffixes or dialects. In other words, conservative forms were dominant for /ps/, /ks/, and /lm/, whereas innovative forms were preferred for /lp/, /lk/, and /ls/. Such patterns were observed regardless of participants’ gender (i.e. male or female).

6.2 Mixed effects analysis

A statistical analysis was conducted in order to examine whether and to what extent stem-final consonant clusters, vowel-initial inflectional suffixes, speakers’ regional dialects, and gender have an impact on the choice between conservative forms and innovative forms in Korean nominal inflection. A mixed effects logistic regression model was fitted using R (R Core Team 2021) and the glmer function in the lmerTest package (Kuznetsova et al. 2017). In the model, the dependent variable was the binary categorical factor “output consonant type” (conservative or innovative). I included the following four independent variables (i.e. fixed effects): “clusters” (6 levels: /ps/, /ks/, /lp/, /lk/, /ls/, /lm/), “suffixes” (3 levels: nom -i, acc -ɨl, dat/loc -ɛ), “dialects” (2 levels: Seoul, Kyungnam), and “gender” (2 levels: male, female). Since all independent variables were categorical, appropriate coding processes were done as follows: two 2-level factors (i.e. “dialects”, “gender”) were sum-coded, whereas the other two tri-or-more-level factors (i.e. “clusters”, “suffixes”) were treatment-coded. The reference level of each independent variable was reordered: /ps/ for “clusters”; nom -i for “suffixes”, Seoul for “dialects”, and male for “gender”. In the previous section, results of the raw data suggested that interactions may not be significant since the general patterns of output forms, either conversative or innovative, were not different when suffixes (6.1.2), dialects (6.1.3), and gender (6.1.4) were calculated with each of the six consonant clusters. However, I included the interactions in the statistical analysis to examine whether there is any significance. Participants and target noun stems (i.e. test items) were included as random effects, which indicate that there was variation depending on participants (variance = 1.067) and target noun stems (variance = 1.128). By running a mixed effects model with these random effects, results show that “clusters” is the only main effect that is statistically significant and that no interactions turned out to be significant. Results of the significant factor (i.e. “clusters”) are shown in (15), leaving out those of the insignificant ones.

Among the six clusters, it is notable that the reference level /ps/ is significantly different from /lp/ (p < .05), /lk/ (p < .01) and /ls/ (p < .05) for the output consonant types. On the other hand, it is not significantly different from /ks/ and /lm/ (both p > .05). This dichotomy agrees with the results of the raw data: /ps/, /ks/, and /lm/ pattern together by preferring conservative forms, while /lp/, /lk/, and /ls/ are identical by favoring innovative forms. As was expected from the raw data, neither of the three predictors other than “clusters” (i.e. “suffixes”, “dialects”, and “gender”) nor any of their interaction with “clusters” impacts output consonant types of the stem-final consonant clusters in Korean nominal inflection.

In sum, these results suggest that different clusters prefer different output consonant types in Korean nominal inflection, either conservative or innovative. When it comes to innovative forms, it is noteworthy that they include only one stem-final consonant via simplification and that the very consonant is identical to what surfaces in the corresponding isolation form. (cf. The consonants that survive in the “other” forms in inflection are not the ones that surface in the isolation forms.) This suggests that there is a relation between consonants in the innovative forms in inflection and those in the isolation forms. These results will be discussed in Section 7.

7.1 Inflected forms: conservative-dominant and innovative-dominant clusters

Results showed in Section 6 that conservative forms were dominant for /ps/, /ks/, and /lm/, and innovative forms were for /lp/, /lk/, and /ls/. Conservative forms are accounted for by the ranking in which the anti-deletion faithfulness constraint Max-C-IO (16) is ranked higher than the universal markedness constraint NoCoda (17): Max-C-IO >> NoCoda (definitions are based on Kager 1999).

Despite the shared pattern, what distinguishes the “Obs.+Obs.” clusters /ps/ and /ks/ from the “Son.+Son.” cluster /lm/ is that the former are rising sonority clusters that are prohibited by Syllable Contact (Syllcon) in Korean (Davis & Seung-Hoon Shin 1999: 286 based on Bat-El 1996: 304).

SyllCon is dominated by NoCoda, which is ranked lower than Max-C-IO, as illustrated in the tableaux for “Obs.+Obs” clusters (19) and “Son.+Son.” cluster (20).

The constraint ranking for inflected forms that work for conservative-dominant output forms at this point is summarized in (21).

However, this ranking fails to predict innovative forms, which are dominant for the “Son.+Obs.” clusters.

Recall that the six stem-final consonant clusters can be grouped based on the [±sonorant] feature of the consonantal elements. The clusters in which consonantal elements share the same [±sonorant] feature (i.e. “Obs.+Obs.” for /ps/, /ks/, and “Son.+Son.” for /lm/) were conservative-dominant, which indicates that both of the input consonants were more likely to be preserved in inflected forms. In contrast, those of which elements do not agree in the [±sonorant] feature (i.e. “Son.+Obs.” for /lp/, /lk/, /ls/) were innovative-dominant with only one of the input consonants preserved in inflection. As can be expected from the generalization, the agreement in the [±sonorant] feature between the consonantal elements of a cluster is the core that distinguishes “Son.+Obs.” clusters from the other two cluster types. This is stated in Agree(son), which is a kind of Identical Cluster Constraints that require a sequence of consonants be identical in a particular featural dimension (Pulleyblank 1997).

Agree(son) contrasts with SyllCon, which is a special kind of Obligatory Contour Principle (Leben 1973) that is specific to sonority. Agree(son) should outrank Max-C-IO in order for innovative forms to win over conservative forms for the “Son.+Obs.” clusters.⁶

The final ranking for the two dominant inflected forms (conservative for /ps/, /ks/, /lm/, and innovative for /lp/, /lk/, /ls/) is in (25).⁷

It will be helpful at this point to further discuss the peculiarity of the “Son.+Obs.” clusters /lp/, /lk/, and /ls/. If they preferred conservative forms, they would have /l/ as the coda, followed by an obstruent as an onset. This is not problematic at all from a synchronic perspective, since it does not violate SyllCon (24a). However, it seems that the /l/+Obs. sequences have recently undergone simplification. It was reported that speakers who were in their 80’s at the time of a survey conducted in early 2015 were highly likely to pronounce conservative forms for /lp/ and /lk/: /jʌtʌlp-i/ [jʌ.tʌl.pi] ‘eight-nom’, /hɨlk-i/ [hɨl.ki] ‘soil-nom’, and /talk-i/ [tal.ki] ‘chicken-nom’ (Seok-kyu Lim 2016). This suggests a diachronic change from conservative forms produced by speakers in their 80’s (e.g. Seok-Kyu Lim 2016) to innovative forms for /lp/ and /lk/ produced by speakers in their 20’s (e.g. the present study). Such a change is even more obvious for /ls/, of which evidence is found in orthography. For the noun meaning ‘first birthday’, what had been written as <tols> was revised to <tol> by the Standard Language Rules that were enacted in Korea in 1988. This shows that there was a stark discrepancy between the written form <tols> and the spoken forms not only in isolation /tols/ [tol] but also in inflection /tols-i/ [to.li] (i.e. innovative) ‘first birthday-nom’. These pieces of evidence for the change in progress as well as the synchronic analysis provided in this section support the preference for innovative forms for the “Son.+Obs.” clusters, as opposed to the two other cluster types (i.e. “Obs.+Obs.” and “Son.+Son.”) that favor conservative forms in inflected forms.

7.2 Isolation forms

Variation occurs in inflected forms of nouns in Korean: while conservative forms are fully faithful to the input stem-final clusters, innovative forms are simplified outputs as isolation forms are. Due to the close relation of innovative forms with isolation forms, it is important to address isolation forms first, in order to provide a unified account of innovative forms in the end. This section thus discusses isolation forms of nouns in Korean. It is crucial to note that there is no variation in isolation forms, as opposed to inflected forms. The output consonants that survive obligatory CCS in isolation forms are based on the standard pronunciations from the Great Dictionary of Standard Korean (National Institute of the Korean Language 2022) (footnote 3).

As discussed in Section 3.1, neither the linear order nor the sonority relationship consistently accounts for the different CCS patterns in isolation forms. In this section, I propose a comprehensive analysis that shows the interaction among phonology (Sections 7.2.1 and 7.2.2), phonetics and segmental frequency (Section 7.2.3), and morphology (Section 7.2.4) that determines which consonant is deleted or preserved in the isolation forms of nouns.

7.2.3 Phonetic saliency and segmental frequency: /l/-deletion from /lk/

Unlike the cluster /lp/ for which the sonority preference in coda works, the input cluster /lk/ shows the opposite pattern. That is, the more sonorous consonant /l/ is always deleted, while the less sonorous consonant /k/ is preserved. For example, the noun talk ‘chicken’ is produced in isolation not as *[tal] but as [tak]. Results of previous studies suggest that there is a phonetic reason for the preservation of /k/. Following Jongho Jun’s (2004) and Taehong Cho & Sahyang Kim’s (2009) velar preference (Section 4.1.1), I propose that the phonetic saliency of the velar obstruent /k/ also plays a role in CCS in the isolation forms of nouns with /lk/.

The hypothesis of velar preference in Korean is also supported by the results of previous studies on segmental frequencies. They show the predominant status of /k/ in frequency. According to Jiyoung Shin’s (2010) analysis of phonemes in 47,401 entries in Yonsei Korean Dictionary, /k/ is the most frequent consonant (13.3%) among 19 consonants in Korean, as shown in Figure 8.

Figure 8

Frequency of 19 consonants in Korean (Jiyoung Shin 2010: 98).

In addition, Jongho Jun (2010) provides the frequency of 14 obstruents that appear in Korean nouns, based on corpus counts and dictionary counts. It is clear that /k/ is by far the most frequent obstruent, as highlighted in Tables 10, 11, and 12.

Table 10

Corpus counts of 14 output obstruents in Korean nouns. (Jongho Jun 2010: 149 based on Albright 2008: 171 citing Heung-Gyu Kim & Beom-Mo Kang 2000).

Labial		Coronal		Velar
p	1,360	t	1	k	5,994
pʰ	64	tʰ	113	kʰ	18
p’	0	t’	0	k’	6
		c	17
		cʰ	160
		c’	0
		s	375
		s’	0

Table 11

Corpus counts of 14 output obstruents in Korean nouns. (Jongho Jun 2010: 149 based on Beom-Mo Kang & Heung-Gyu Kim 2004).

Labial		Coronal		Velar
p	2,515	t	2	k	7,537
pʰ	84	tʰ	116	kʰ	24
p’	0	t’	0	k’	9
		c	18
		cʰ	164
		c’	0
		s	308
		s’	0

Table 12

Dictionary counts of 14 output obstruents in Korean nouns. (Jongho Jun 2010: 149 based on Jaewon You 1985).

Labial		Coronal		Velar
p	935	t	4	k	1,234
pʰ	96	tʰ	184	kʰ	17
p’	0	t’	0	k’	10
		c	38
		cʰ	171
		c’	0
		s	473
		s’	0

On the basis of the prominence of /k/ observed in both phonetic and frequency-based analyses, I propose Max-k-IO, a more detailed version of anti-deletion constraint, which penalizes deletion of the specific segment /k/.

1. (36)

1. Max-k-IO: Input /k/ must have the correspondent [k] in output. (‘No /k/-deletion’)

The optimal output form of /talk/ ‘chicken’ is [tak], and it is chosen as the winner in the ranking in which Max-k-IO dominates *Coda(obs). Between Max-k-IO and Max-C-IO, there is no crucial ranking.

1. (37)

1. Tableau for /lk/
2. 

In sum, the updated ranking for CCS in the isolation forms of nouns in Korean is in (38).

1. (38)

1. *CC, *FricCoda >> Max-C-IO, Max-k-IO >> *Coda(obs) >> *Coda(liq) (to be revised)

7.3 Non-dominant innovative forms in inflection: in relation to isolation forms

In order to account for non-dominant innovative forms in Korean nominal inflection, isolation forms should be referred to, since the very consonant that survives in an isolation form also appears in corresponding innovative forms. For example, the isolation form of the noun /kaps/ ‘price’ is [kap], and the inflected form /kaps-i/ ‘price-nom’ can be either [kap.si] (conservative) or [ka.pi] (innovative) but not *[ka.si], which is ungrammatical because there is no isolation form like *[kas]. This is the gist of Kenstowicz’s (1996) Base-Identity effect. Otherwise, CCS in innovative forms might seem opaque, as Jin-hyung Kim (2005) points out (Section 4.1). Heejeong Ko (2006) argues for the Base-Output Correspondence Theory, following Kager (1999). Similarly, by adopting Kenstowicz’s (1996) Base-Identity constraint, which is one kind of segmental output-output correspondence constraint, I show in this paper that isolation forms are the Base to which innovative forms in inflection make reference.

Base-Identity requires the consonants in innovative inflected forms be segmentally identical to those in isolation forms (i.e. Base). In other words, it states that inflected forms should be equal to the structure “isolation form + inflectional suffix”. Therefore, a violation mark should be given if any consonant in an inflected form does not have a segmental correspondent in the isolation form. Base-Identity is vacuously satisfied when selecting the optimal output of nouns in isolation because there is no Base for isolation forms to refer to.

While Base-Identity is not in effect for isolation forms, it is important in the selection of the optimal output of innovative forms in inflection.

Candidates (47a) [kap.si] and (47c) [ka.si] are ruled out because they violate Base-Identity. These two candidates lose since [s] does not make reference to the final consonant in the isolation form [kap]. Since candidate (47b) [ka.pi] satisfies Base-Identity by having [p] in the inflected form as well as in the isolation form [kap] at the expense of deleting /s/, it is the optimal output for innovative inflection.

The effect of the constraint Base-Identity is in line with Steriade’s (2000) principle of Lexical Conservatism, which states that novel forms must find a precedent in a listed form. The existing isolation form [kap] ‘price’ is the precedent in our case, and for this reason, [ka.pi] is selected for ‘price-nom’. On the other hand, *[ka.si] and *[kap.si] are ruled out because there is no listed isolation form *[kas] or *[kaps] in Korean. (cf. For the constraint ranking that selects conservative forms as an optimal output form, see Section 7.1.)

The constraint rankings established for dominant inflected forms (either conservative or innovative), isolation forms, and non-dominant innovative inflected forms are summarized in (48).

The final constraint rankings are consistent with the rest of Korean phonology. First, (48a) holds for Korean nasalization and lateralization. (See footnote 6.) Second, (48b) also works, particularly in that *CC and *FricCoda are undominated in Korean, and high sonority codas are universally preferred (Vennemann 1988) (i.e. *Coda(obs) >> *Coda(nas) >> *Coda(liq)). Lastly, the ranking in which Base-Identity dominates Max-C-IO accounts for non-dominant innovative forms (e.g. /kaps-i/ [ka.pi]≻[kap.si]) (48c). This is consistent with other parts of Korean, such as verbal inflection, since Base-Identity is vacuously satisfied in that case due to the lack of isolation forms for verb stems. (See Section 2.)

7.4 Notes on non-significant factors

Experimental results show that, in contrast to “clusters”, the other three predictors “suffixes”, “dialects”, and “gender” were not statistically significant. First, some studies have suggested a relationship between vowel-initial inflectional suffixes and surface forms of stem-final consonants. Jongho Jun & Jeehyun Lee (2007) and Jongho Jun (2010) show that suffixes of different vowel qualities -i, -ɨl, and -ɛ have an impact on variation in stem-final coronal obstruents /s, t, tʰ, c, cʰ/ in Korean. They ascribe this to a frequency effect: [cʰ] occurs frequently before -i and -ɨl, while [tʰ] before -ɛ in Korean, which results in the preference for a particular output consonant before different vowel-initial suffixes. However, as for variation in stem-final consonant clusters, we observed that vowel-initial inflectional suffixes had no significant effect on conservative forms and innovative forms in Korean nominal inflection. As a reviewer suggested, I ascribe this to the nature of consonant clusters, as opposed to the nature of simplex coronal obstruents. The frequency-based analysis works for variation in coronal obstruents since the speakers’ job is to choose the quality of the output consonants among [s, t, tʰ, c, cʰ] that occurs frequently before a particular vowel. In other words, their decision is based on the extension of dominant patterns in the language to this particular phenomenon. On the other hand, speakers undergo a different kind of decision process for variation in consonant clusters: it is either conservative with two consonants or innovative with one consonant. More importantly, the quality of output consonants in innovative forms is not something that speakers can choose: it is predetermined in isolation forms and extended to innovative forms due to the Base-Identity effect. Thus, it makes no sense that some output consonants prefer some suffixes since the former in our case are preset for consonant clusters.

Second, another factor that was of interest to previous researchers is speakers’ regional dialects. There might have been a dialectal difference between Seoul Korean and Kyungsang Korean at least until early to mid 1990’s (e.g. Whitman 1985; Young-mee Yu Cho 1988; Mira Oh 1994). However, this study shows that there is no such difference among younger speakers of Korean. Even though there used to be dialectal variation in the past, it is likely that it has been diminished over time. Although discussing a different topic, Jongho Jun & Jeehyun Lee (2007) mention that their participants who were from the North Kyungsang area and in their 20’s had been exposed to standard Seoul Korean via the mass media, consequently showing patterns that were similar to Seoul speakers. This can also be a plausible account for the lack of dialectal variation for consonant clusters in nominal inflection.

Lastly, there was no significant difference between male speakers and female speakers in the variants surfacing stem-final consonants, either conservative or innovative. What was expected from Labov’s (1990) Principle I is that male speakers would use more innovative forms in which one of the two consonants was dropped, whereas female speakers would favor conservative forms in which both consonants were preserved. However, results of the production experiment did not turn out to be so in this study. This suggests that, for the younger generation in Korea, male and female speakers do not show a significant difference in variation in nominal inflection.