Testing Japanese loanword devoicing: Addressing task effects

In the loanword phonology of Japanese, voiced obstruent geminates ([bb, dd, gg]) have been claimed to devoice when they co-occur with another voiced obstruent within the same morpheme (e.g. /beddo/ → [betto] ‘bed’). This devoicing pattern has contributed much to address a number of theoretical issues in the recent phonological literature. However, the relevant data have been primarily based on intuition-based data provided by Nishimura (2003) and Kawahara (2006). Kawahara (2011a, 2011b) addressed this issue by conducting rating studies using naive native speakers of Japanese. The results generally supported the intuition-based data by Nishimura (2003) and Kawahara (2006). However, the rating studies also revealed several aspects of the devoicing pattern that go beyond the intuition-based data as well. The current study further investigates the devoicing pattern by varying several task variables. In particular, this paper builds on Kawahara (2011a, 2011b) by adding (i) nonce word stimuli, (ii) a binary yes/no experiment, and (iii) auditory stimuli. The results show that (i) nonce words and real words behave similarly, but nonce words nevertheless show less variability across different grammatical conditions than real words; (ii) the binary yes/no experiment shows results similar to those of the scale-based experiment; and (iii) while auditory stimuli yield results comparable with those of orthographic stimuli, they also show an exaggerated effect of a phonetic implementation pattern. Overall, this paper uses Japanese as a case study, and finds some task effects in phonological judgment experiments. It is hoped that this paper stimulates further experimental research on phonological judgments of other phenomena in Japanese as well as in other languages. 1. Introduction 1.1. The phenomenon The experiments reported in this paper were first reported in a longer manuscript circulated in 2010 (Kawahara, 2010). For their insightful comments on this paper and Kawahara (2010), I am grateful to Osamu Fujimura, Kazu Kurisu, Julien Musolino, Jeremy Perkins, Jason Shaw, Mariko Sugahara, Kristen Syrett, Kyoko Yamaguchi, four anonymous reviewers, and the audience at the colloquium talks at the University of Pennsylvania and Johns Hopkins University, and the participants of Japan Phonology Forum 2011. Sophia Kao and Melanie Pangilinan helped me with the stimulus preparation for Experiment II. Finally, I would like to thank Hope McManus for her edits at the final stage of the manuscript preparation. This research is partly supported by Research Council grant to the author from Rutgers University. All remaining errors are mine. This paper is about devoicing of obstruents in the loanword phonology of Japanese. It has been known that voiced obstruent geminates ([bb, dd, gg]) in Japanese loanwords can be devoiced (Itô and Mester 1995, 1999; Quakenbusch 1989; Vance 1987), but exactly when such devoicing occurs remained unclear. For example, Itô and Mester (1999) argued that some items can undergo devoicing while other items cannot, and considered the first type of words as “assimilated foreign items” and the second, non-devoicing type as “unassimilated foreign items”. Instead of relying on a (more or less) arbitrary etymological distinction, Nishimura (2003) proposes a phonological characterization of this distinction, claiming that voiced obstruent geminates optionally devoice when they co-occur with another voiced obstruent within the same stem, as exemplified by the data in (1). He further claims that this devoicing is due to a restriction against having two voiced obstruents within the same stem. In Japanese phonology, this restriction has long been known as Lyman’s Law (Kawahara 2012b; Lyman 1894; Vance 2007), and has been formalized as OCP[voice] (Obligatory Contour Principle: Leben 1973; henceforth simply the OCP) (Itô and Mester 1986, 1998, 2003). In other words, devoicing is possible in (1) whereas it is impossible in non-OCP-violating voiced geminates, as shown in (2). Moreover, in an interesting twist, Nishimura (2003) argues that devoicing is also impossible in OCP-violating singletons, as in (3). (1) Voiced obstruent geminates optionally devoice if they co-occur with another voiced obstruent; i.e. when they violate OCP[voice]. beddo → betto ‘bed’ baggu → bakku ‘bag’ biggu → bikku ‘big’ (2) Voiced obstruent geminates do not devoice if they do not violate OCP[voice]. sunobbu → sunobbu *sunoppu ‘snob’ heddo → heddo *hetto ‘head’ reggu → reggu *rekku ‘leg’ (3) Voiced singletons do not devoice even when they violate OCP[voice]. dabu → dabu *dapu ‘Dove’ doguma → doguma *dokuma ‘dogma’ dagu → dagu *daku ‘Doug’ The patterns in (1)-(3) have attracted much attention in the recent phonological literature. It is beyond the scope of this paper to settle these debates; however, to briefly summarize, the devoicing pattern triggered three major theoretical debates: (i) how to explain the difference between singletons (=the data in (3)) and geminates (=the data in (1)) (Kawahara 2006, 2008; Rice 2006; Steriade 2004); (ii) how to capture the cumulative markedness requirement of devoicing in (1) (Farris-Trimble 2008; Nishimura 2003; Pater 2009, forthcoming; Tesar 2007); and (iii) how the spontaneous emergence of loanword devoicing in (1) bears on the theory of lexical stratification—a theory of how loanword phonology is related to native phonology (Crawford 2009; Itô and Mester 2003, 2008; Tateishi 2002). See Kawahara (2011a) and Kawahara (2012a) for recent summaries (the former in English and the latter in Japanese). In short, the Japanese loanword devoicing pattern has contributed much to several theoretical debates in recent years. However, Kawahara (2011b) raises one issue: the Japanese loanword devoicing data are primarily based on the intuitions of two linguists, namely, Nishimura (2003) and Kawahara (2006); i.e. the grammaticality judgments in (1)-(3) primarily come from the authors themselves. 2 Many studies have raised concerns about research exclusively relying on authors’ own introspections (e.g. Da̧browska 2010; Gibson and Fedorenko 2010; Griner 2001; Labov 1996; Myers 2009; Ohala 1986; Schütze 1996). To address this problem, Kawahara (2011b) conducted a rating experiment with 38 native Japanese speakers who did not know about the devoicing pattern. The experiment indeed showed that Japanese speakers generally judge devoicing of OCP-violating geminates as more natural than devoicing of non-OCP-violating geminates or devoicing of OCP-violating singletons. In this regard, Kawahara (2011b) succeeded in supporting the empirical basis of the claims made about the patterns in (1)(3). Kawahara (2011a) reports a follow-up experiment using a larger set of stimuli with 49 naive native speakers, which again supported the idea that devoicing of OCP-violating geminates is the most natural environment for native speakers of Japanese. 1.2. The current study There are some remaining questions, however. First, both Kawahara (2011a) and Kawahara (2011b) used only real words. In the case of Japanese loanword devoicing, it is of some interest to investigate whether the results obtained for real words generalize to nonce words. An often-used test on phonological productivity is a wug-test (Berko 1958), in which participants are asked to inflect nonce words. Some previous wug-tests have failed to replicate phonological patterns that apply to real words, in which case it is often concluded that the alleged phonological patterns are not productive; i.e., they are lexicalized (Griner 2001; Ohala 1974; Sanders 2003). (See also Shademan (2007) for some related discussion.) If the phonological pattern under discussion is not productive with nonce words, the pattern should probably not be used for phonological argumentation. In fact, there is an example from Japanese phonology whose productivity has been questioned by way of experiments using nonce words. Several phonological changes occurring in Japanese verbal paradigms (Davis and Tsujimura 1991; Tsujimura 1996) were not replicated in nonce word experimentation (Batchelder 1999; Griner 2001; Vance 1987). In short, there is no guarantee that we can generalize the patterns of real words to nonce words, and it is vital to test the productivity of the phenomenon under question using nonce words. Experiment III in Kawahara (2012) addressed this question, although in that paper, the comparison between real words and nonce words was not the main focus. The section 2 of this paper therefore reports that experiment in more detail to address this issue of whether the previous results can be replicated with nonce words. The first aim of this paper is to therefore re-examine Experiment III of Kawahara (2012) to mainly address the question of whether the results obtained in Kawahara (2011a) and Kawahara (2011b) can be replicated with nonce words. This report also allows us to compare the results of that experiment with two other experiments reported in the paper. See Nishimura (2003), Kawahara and Sano (2012) and Sano and Kawahara (2012) for some evidence based on corpus data. The second aim of this paper is to test the gradiency of judgment patterns found in the previous experiments. Kawahara (2011a, 2011b) found that Japanese speakers distinguish the naturalness of two processes that were both judged to be “ungrammatical” by Nishimura (2003) and Kawahara (2006), with the devoicing of non-OCP violating geminates (=the examples in (2)) rated as more natural than the devoicing of OCPviolating singletons (=(3)). One may wonder whether this gradient effect was due to a task effect; the reason being that Kawahara (2011a, 2011b) uses a gradient scale. Testing this issue is in part motivated by the debate concerning the gradient nature of phonological judgments. It is known that grammatical judgments show distinctions beyond a simple, binary “grammatical” vs. “ungrammatical” dichotomy, especially in experimental settings (see e.g., Albright 2009; Coetzee 2008; Coleman and Pierrehumbert 1997; Daland et al. 2011; Dankovičová et al. 1998; Goldrick 2011; Greenberg and Jenkins 1964; Hayes 2000; Hayes and Wilson 2008; Pertz and Bever 1975; Pierrehumbert 2001; Shademan 2007 for phonological/phonotactic judgments; Chomsky 1965; Myers 2009; Schütze 1996; Sorace and Keller 2005 for syntactic judgments). However, one may contend the idea that we obtain gradient results in experimental settings because these experiments use scales. Therefore, the second aim of this paper is to test whether the gradient results that Kawahara (2011a, 2011b) obtained can be replicated using a binary yes/no task. Some previous studies (Bader and Mäussler 2010; Coleman and Pierrehumbert, 1997; Dankovičová et al., 1998; Frisch et al., 2004) raised similar issues and found gradient results using a binary yes/no format. The current study thus builds on them and aims to address the gradient nature of phonological judgments in the case of Japanese loanword devoicing. Finally, the third aim of this paper is that Kawahara (2011a, 2011b) used visual, orthographic stimuli, although the instructions in these studies encouraged the participants to read the stimuli in their heads and use an auditory impression to make judgments. While many judgment experiments in linguistics are run with orthography, it is worth running the same experiment with auditory stimuli for a few reasons. First, one explanation for why voiced geminates, but not voiced singletons, can devoice is because a phonological voicing contrast is auditorily less perceptible in geminates than in singletons (Kawahara 2006, 2008). An auditory judgment experiment would help to address this specific hypothesis. Second, it would be interesting to investigate whether the results of Kawahara (2011a, 2011b) can be replicated with auditory stimuli, because phonology is concerned with sounds. Replicating the Japanese devoicing pattern with auditory stimuli is therefore the third aim of this paper. To summarize, there are three issues that this paper aims to address: (i) thejudgment patterns on devoicing as revealed by nonce words, (ii) the effect of using abinary yes/no format, and (iii) the effect of using auditory stimuli. This paper reexaminesKawahara (2012b) and reports two additional experiments in order to address these threeissues. More generally, by varying experimental variables, the current project aims to further examine the empirical basis of the theoretical debates reviewed in section 1.1,beyond Kawahara (2011a, 2011b).Before reporting the actual experiments, a few remarks are in order. First, theexperiments reported in this paper are judgment experiments for a phonological process,i.e., devoicing. The task is for native speakers to judge the naturalness or possibility of aphonological pattern, or in other words, a pairing between one form and another form (i.e. in this case, a phonological form and its optional variant). This task therefore differsfrom phonotactic wellformedess judgment tasks in which speakers judge the wellformedness of surface forms only (e.g., Bailey and Hahn 2001; Coetzee 2008;Coleman and Pierrehumbert 1997; Daland et al. 2011; Dankovičová et al. 1998;Greenberg and Jenkins 1964; Shademan 2007). Second, this paper offers a case study inJapanese of such a phonological judgment study. Although its scope is thus limited, it ishoped that this paper will stimulate further studies on different phonological phenomenain different languages (including Japanese). 2. Kawahara (2012b) Experiment III and beyondThis experiment, briefly reported as Experiment III in Kawahara (2012b), is anorthography-based rating experiment. This section reexamines this experiment in detail,since some details and analyses of this experiment were omitted from Kawahara 2012b(Kawahara 2012b was written after the current paper), 3 and also since the other two experiments in the current paper crucially build on this experiment and I will make manycross-experimental comparisons in what follows. In particular, this section reexaminesthe experiment from the perspectives that are discussed in the introduction, mainly withthe focus of comparing real words and nonce words.This section thus addresses three issues: (i) to replicate Kawahara (2011a, 2011b);and, more importantly, (ii) to test whether the results obtained with real words in theprevious studies generalize to nonce words; and finally, (iii) to compare the patterns ofreal words and nonce words. 2.1. Method2.1.1. Stimuli. The stimuli consisted of four grammatical conditions: (i) OCP-violatinggeminates, (ii) non-OCP-violating geminates, (iii) OCP-violating singletons, and (iv)non-OCP-violating singletons, as summarized in (4), each with a representative example.In this design, two factors—OCP and GEM—were fully crossed. This paper usesCAPITAL LETTERS to represent variable names. (4)The four grammatical conditionsa. OCP-violating geminates (e.g. [baggu])b. non-OCP-violating geminates (e.g. [eggu])c. OCP-violating singletons (e.g. [dagu])d. non-OCP-violating singletons (e.g. [magu]). The experiment had 9 items per each condition. All the stimulus items weredisyllabic, and all the target consonants were word-internal (since all lexical geminates in In Kawahara (2012b), together with two other experiments on Rendaku, this experiment is also reported(in much less detail) to show the activity of the OCP—or Lyman’s Law—in both loanwords and nonce words. That paper does not extensively discuss the effect of geminacy, its interaction with OCP, or on the comparison between real words and nonce words. Neither does that paper report the issue of gradiency at all; i.e. the analyses presented in Figures 2 and 3 below are new to this paper. Figure 1 is reproduced withpermission from Elsevier. Japanese appear word-internally: Kawahara (forthcoming)). The stimulus set wasconstructed in the following way: first, real disyllabic words containing OCP-violating geminates were chosen; this case has the least number of existing items in the Japaneselexicon. This selection process resulted in 9 items. Among those 9 items, 6 itemscontained [dd] followed by epenthetic [o], and the remaining 3 items contained [gg]followed by epenthetic [u]. No stimuli with [bb] were found; in fact, no disyllabic wordswith OCP-violating [bb] exist which is not unexpected given that [bb] is very rare inJapanese loanwords (Katayama 1998). Then the items for the other three conditions wereselected, consisting of 6 items for [d(d)] and 3 items for [g(g)], as listed in Table 1.Across all conditions, the number of items for each place of articulation was controlledfor. Short vowels were used before geminates and singleton [g]. Long vowels anddiphthongs had to be used before singleton [d], because disyllabic loanwords with aninitial short vowel almost always have a geminate [dd], and not a singleton [d]. Thispattern is due to a productive gemination process in loanword adaptation (e.g. [baddo] ‘bad’; see e.g. Katayama1998; Kubozono et al. 2008).4All of the stimuli have a pitchaccent on the initial syllable, which is phonetically realized as a HL falling F0 contour. Table 1: The list of the stimuli that are real words.OCP-GemGemOCP-SingSingbaddo ‘bad’heddo ‘head’ bado ‘badminton’muudo ‘mood’beddo ‘bed’reddo ‘red’ gaido ‘guide’waido ‘wide’daddo ‘dad’uddo ‘wood’ zoid common name haido ‘hide’deddo ‘dead’kiddo ‘kid’ boodo ‘board’roodo ‘road’guddo ‘good’maddo ‘mad’ gaado ‘guard’riido ‘lead’goddo ‘god’roddo ‘rod’ baado ‘bird’huudo ‘food’baggu ‘bag’eggu ‘egg’dagu ‘Doug’hagu ‘hug’biggu ‘big’reggu ‘leg’ bagu ‘bug’magu ‘mag’doggu ‘dog’taggu ‘tag’ jogu ‘jog’ragu ‘rag’ The nonce word stimuli are listed in Table 2. These stimuli had the same phonologicalstructures as the real word stimuli, except that all nonce word stimuli had short initialvowels, including those nonce words that contain a singleton [d], which can also have ashort vowel. Table 2: The list of the stimuli that are nonce words.OCP-GemGemOCP-SingSing buddokeddobudohudoboddokoddodadoradododdoruddododorudo geddoyuddodedoridogaddotaddogadoyudogiddokuddogudowadobogguuggudeguhegu [bado] is a truncated form of [badominton]. gagguoggudogunegu goggunaggugegumugu 2.1.2. Task. In this experiment Japanese speakers were asked to rate the naturalness ofdevoicing in the four grammatical conditions. The instructions stated that thequestionnaire was about the naturalness of devoicing in Japanese loanwords. Using thesame format as Kawahara (2011a, 2011b), for each question, the participants werepresented with one stimulus item and asked to judge the naturalness of the form thatundergoes devoicing of word-internal consonants (e.g. given [baddo], how natural wouldyou find it to pronounce it as [batto]?). The instructions and stimuli were presented inJapanese orthography. The katakana orthography was used for the stimuli (for both[baddo] and [batto] in the example above), for both real words and nonce words, sincekatakana is conventionally used for loanwords and nonce words in standard Japaneseorthography (Labrune 2012). Although the test was based on orthography, theparticipants were asked to read each stimulus in their heads, and make judgments basedon their auditory impression rather than on the orthography.Following Kawahara (2011a, 2011b), in this experiment, the speakers judged thenaturalness of devoicing using a 5-point scale, as follows: A. “very natural”, B.“somewhat natural”, C. “neither natural nor unnatural”, D. “somewhat unnatural”, and E.“very unnatural”. Since the software that ran the experiment (see Section 2.1.3) could notpresent the scale numerically, the responses were later converted to a numerical scale.The main session was blocked into two parts. The first block presented all the realword stimuli, followed by a break sign. The second block presented all the nonce wordstimuli. The entire experiment was structured in this way because it was assumed thatmaking judgments about real words would be easier than making judgments about noncewords for the participants, allowing the participants to first gain familiarity with the taskbefore giving judgments for nonce words. 2.1.3. Procedure. Sakai, an online system which runs questionnaires, was used to runthe current online experiment. An advantage of this internet-based methodology is thefact that it is easy to get a large number of participants. This advantage is particularlyimportant when the researcher does not reside in an area where there are many localspeakers of the target language. A potential disadvantage is that the researcher cannotcontrol the environment in which the participants take the experiment, although Sprouse (2011) shows that linguistic judgment data gathered with this sort of method arecomparable with the data gathered in the laboratory (see Reips (2002) and Sprouse(2011) for further, general discussion about online experimentation in psychology andlinguistics). The experimental website first presented a consent form and the instructions ofthe experiment. Then the main session started, with one trial presented per page. Theorder of the stimuli within each block was randomized. Kawahara (2010) reports an experiment that addresses the question of how this organization may have affected the results. 2.1.4. Participants. Thirty-three native speakers of Japanese, who were mainly studentsat a Japanese university, participated in the experiment. One speaker reported that they are familiar with the devoicing pattern, and hence his/her data were excluded from thefollowing analysis. 2.1.5. Statistics. The responses were first converted to numerical values as follows: A.“very natural”=5; B. “somewhat natural”=4; C. “neither natural nor unnatural”=3; D.“somewhat unnatural”=2; E. “very unnatural”=1. For statistical analyses, a general linearmixed model was run (Baayen et al. 2008; Baayen 2008) using R (R Development CoreTeam, 1993–2013) with the lme4 package (Bates et al. 2011). The fixed factors wereOCP and GEM. The p-values were calculated by the Markov chain Monte Carlo methodusing the languageR package (Baayen 2009). 2.2. ResultsFigure 1 illustrates average rating scores. In real words, the average naturalness ratingsshowed the following order: OCP-violating geminates (4.23) > non-OCP-violatinggeminates (3.29) > OCP-violating singletons (2.69) > non-OCP-violating singletons(2.21). Simply put, devoicing of OCP-violating geminates was rated as more natural thanthe devoicing of non-OCP-violating geminates which, in turn, was rated as more naturalthan the devoicing of OCP-violating singletons, which was rated as more natural thandevoicing of non-OCP-violating singletons, replicating the results of previous studies(Kawahara, 2011a, 2011b). Statistically, for real words, all factors are significant: OCP (t= 5.29, p < .001), GEM (t = 11.81, p < .001), and the interaction between OCP and GEM(t = 2.68, p < .01). The significance of the main effects shows that OCP and GEM eachaffect naturalness ratings on devoicing, and the significant interaction term indicates thatthe effect of OCP is bigger on the geminate pair (4.23-3.29=0.94) than on the singletonpair (2.69-2.21=0.48). To make the interpretation of the statistical analyses simpler, this model left out the effect of lexical usagefrequencies on naturalness ratings. See Coetzee and Kawahara (2013), Kawahara (2011a) and Kawahara and Sano (2012) for discussion and also modeling of lexical frequency effects in the Japanese loanword devoicing pattern. Also, to avoid interpreting complex interaction terms, the difference between real words and nonce words was not coded in this model. The targeted comparison between real words and noncewords is provided in the discussion section (Section 2.3). Figure 1: The average naturalness ratings in the orthography-based rating experiment.Error bars represent 95% confidence intervals. Reprinted with permission from Elsevier. For nonce words, the order of the naturalness ratings is the same as the real wordcondition: OCP-violating geminates (3.64) > non-OCP-violating geminates (3.41) >OCP-violating singletons (3.06) > non-OCP-violating singletons (2.81). The statisticalanalysis shows that both OCP (t = 2.56, p < .05) and GEM (t = 6.44, p < .001) aresignificant, but, unlike the result of real words, their interaction is not (t = 0.06, n.s.). Fornonce words, the effect of OCP on naturalness ratings is comparable between thesingleton condition (3.64-3.41=0.23) and the geminate condition (3.06-2.81=0.25). 2.3. Discussion2.3.1. Real words vs. nonce words. First, we observe the same order of the four grammatical conditions across real words and nonce words. This order also matches withthe results of two previous studies using real words (Kawahara 2011a, 2011b). In thissense, the current experiment has shown that the results of the previous studies that usereal words generalize to nonce words. Most importantly, even in nonce words, OCP-violating geminates received the highest naturalness ratings, supporting the original observation by Nishimura (2003). The current experiment thus contributes furtherempirical support for theoretical claims made about the Japanese loanword devoicingpattern (see Section 1.1).At the same time, we observe a difference between real words and nonce words:in nonce words, there is less variability in naturalness ratings across the four grammatical conditions than in real words. In other words, devoicing in nonce words showed lessvariation in naturalness ratings across the four conditions than devoicing in real wordsOCP+Gem Gem OCP+Sing SingReal Words Averagenaturalnessratings 12345 OCP+Gem Gem OCP+Sing SingNonce Words Averagenaturalnessratings 12345 did. The condition rated as having the most natural devoicing pattern is OCP-violatinggeminates; devoicing in OCP-violating geminates is judged to be less natural in nonce words than in real words. The least natural devoicing pattern is non-OCP-violatinggeminates; this condition is judged to be more natural in nonce words than in real words.To statistically assess this difference between real words and nonce words, foreach speaker, the standard deviations across all tokens were calculated separately for realwords and nonce words. These standard deviations were then compared between the twoconditions using a non-parametric within-subject Wilcoxon test. This analysis shows thatthe average standard deviations are 1.30 for the real words and 1.03 for the nonce words,and that the difference is significant (p < .001).This reduction of variability across the four grammatical conditions in noncewords could be responsible for the absence of a significant interaction between OCP andGEM in nonce words; there may not be a space left for OCP-violating geminates to havenaturalness ratings that are high enough to yield a significant interaction between OCPand GEM.A question arises as to where the difference between real words and nonce wordscomes from. Presumably the participants have encountered real instances of devoicing inreal words, which would make them more confident about what would happen to eachtarget word. On the other hand, the participants have not seen nonce words before, andtherefore they may feel less committed about making extreme judgments in general; i.e.they are reluctant to use endpoints of judgment scales. Despite this difference betweenreal words and nonce words, as discussed, we observe the same ordering between thefour grammatical conditions in real words and nonce words. 2.3.2. Gradiency. Second, the current study found gradient grammatical distinctionsamong the four grammatical conditions, just like the two previous studies (Kawahara,2011a, 2011b). It does not seem possible to divide the judgment patterns simply into the“grammatical” category and the “ungrammatical” category. In this sense, the currentresults agree with the previous studies in finding distinctions that go beyond whatNishimura (2003) first proposed.One question that arises is whether this four-way distinction is due to a non-homogeneous speech community. That is, one could argue that the response from eachspeaker is always binary which follows a “grammatical” vs. “ungrammatical” dichotomy,but averaging over the responses from different speakers results in a gradient pattern.This hypothesis predicts that, in a sample of a homogenous speech community, distributions of responses are at two extremes, because people should consistently rateeach devoicing pattern either as completely natural (i.e. grammatical; =5 in rating) orcompletely unnatural (i.e. ungrammatical; =1 in rating). In this view, the differences Kawahara (2011a, 2011b) speculates about why Japanese speakers find the devoicing of non-OCPviolating geminates more natural than that of OCP-violating singletons. Beyond the speculation presented there, yet another possibility is that a constraint against voiced geminates is a phonetically natural one (Ohala 1983), whereas OCP[voice] in Japanese is not (Kawahara 2008; Ohala 1981). In fact, there isevidence that children acquiring Japanese show a stage in which they apparently do not show the effect of OCP[voice] (Fukuda and Fukuda 1994), implying that this constraint may have to be learned rather than being innate. Given the characteristics of OCP[voice] in Japanese, the speakers may have found the phonetically natural devoicing (=geminate devoicing) more grammatically natural. This possibility wasbrought to my attention by Armin Mester (p.c. August 2011). between the four grammatical conditions arise from the difference in the number ofspeakers who assign grammatical status (=5 in rating) to each condition. To examine this prediction, Figures 2 and 3 provide histograms that show the distributions of averagescores for each speaker in each grammatical condition. We observe that, contra thehypothesis, there are many speakers who show intermediate average scores in eachgrammatical condition. Figure 2: A histogram of naturalness ratings (number of speakers) for real words.OCP+Gem Average ratingsNumberofspeakers 1 2 3 4 5051015Gem Average ratingsNumberofspeakers 1 2 3 4 5051015