Why languages show variation: A case study from Japanese
INTRODUCTION: It is a truism in natural language that the same word can be pronounced in multiple ways; i.e., there is phonetic and phonological variation. While there is now a substantial body of research on this topic (e.g. Coetzee 2013; Coetzee & Pater 2011), a fundamental question is rarely addressed: why does such variation exist in the first place? We propose that variation follows from language as an effective system of message transfer. From this proposal, we derive a theory which explains (i) why language users deploy variation, and (ii) what kind of variation will be observed in what environments. As a case study, we apply the theory to geminate devoicing in Japanese loanwords and demonstrate how our approach provides increased coverage of the facts over previous analyses.
THE THEORY: If we regard language as an effective system of message transfer (Hall et al. 2016), it follows that messages are understood from the signal in the context produced. In terms of probability theory, we can formalize this idea as (1):
(1) Let �(�������|������, �������) be the probability of listeners retrieving the intended message given its signal (=phonetic form) and the context. Speakers keep �(�������|������, �������) high and constant.
Applying Bayes Theorem to �(�������|������, �������), we get:
(2) �(�������|������, �������) = �(�������|�������) ∗ �(������|�������) where �(�������|�������) is the probability of the message given the context, and
�(������|�������) is how clearly each phonetic signal differentiates the intended message from its lexical competitors, referred to as signal specificity (Hall et al. 2016).
Given the postulate in (1) that �(�������|������, �������) be kept high and constant, then from (2) it follows that �(�������|�������) and �(������|�������) should systematically co-vary. This already explains why language users deploy variation; if
�(�������|������, �������) is kept constant, then �(������|�������) needs to vary as a function of �(�������|�������); i.e. speakers use different phonetic forms in order to vary signal specificity according to message predictability. This theory makes one more specific prediction: the higher the predictability of the message, the lower �(������|�������). THE JAPANESE DATA: We illustrate the key prediction with geminate devoicing in Japanese loanwords, which has been analyzed in various theoretical frameworks (Kawahara 2015). Japanese does not have voiced geminates in the native vocabulary, but in loanwords word- final voiced stops are borrowed as voiced geminates (e.g. /doɡɡu/ ‘dog’: Ito & Mester 1995). Voiced geminates have high signal specificity (=�(������|�������) ). Because native words do not contain voiced geminates and there is a small number of voiced geminates in the overall Japanese lexicon, having a voiced geminate makes a particular word “stand out” from other words. Therefore, devoicing a geminate decreases signal specificity.
Japanese has an active OCP(voice) constraint, which prevents two voiced obstruents in a morpheme in native words. Loanwords can violate this constraint, possibly because they need to retain a high degree of signal specificity to balance low predictability, according to (2). Voiced geminates optionally devoice when they co-occur with another voiced obstruent in the same word ([doɡɡu]~[dokku]: Kawahara 2006). Devoicing geminates in this context decreases signal specificity but to a lesser degree than devoicing singletons, as demonstrated by Kawahara (to appear). Further, since devoicing geminates in this context exerts a smaller reduction in signal specificity than devoicing singletons, this process is sensitive to subtler
variation in predictability. This was illustrated by Kawahara’s (2011) study in which the lexical frequency of the word that contains a voiced geminate and the number of moras of that word both impact geminate devoicing, as shown in Figure 1.
The two effects in Figure 1 are almost always set aside in previous analyses; one exception is Coetzee and Kawahara (2013), who add an extra mechanism to Harmonic Grammar to account for lexical frequency. No previous work has accounted for the word length effect. Frequency is an estimate of predictability out of context. Hence higher frequency words are predicted to show lower signal specificity. Increased geminate devoicing as a function of increased frequency functions to vary signal specificity according to predictability. Geminate devoicing in long words follows from the requirement that signal specificity remains constant at a given level of predictability. Longer words have fewer lexical competitors than shorter words and, hence, higher signal specificity. Geminate devoicing lowers signal specificity, compensating for the effect of word length to keep word signal specificity proportional to message predictability. The current theory therefore provides coverage of the two aspects of Japanese geminate devoicing that other theories fail to address.
CONCLUSION: According to (1) and (2), phonological variation is a response to the requirement that speakers keep �(�������|������, �������) high and constant. The prediction is borne out in patterns of Japanese geminate devoicing, but some questions remain. Why is geminate devoicing, as opposed to some other phonological process, the response to varying message predictability? We speculate that other phonological changes may exert an impact on signal specificity that is disproportionate to the changes in predictability exerted by word frequency. This may also explain why geminates do not devoice outside of OCP contexts. In the OCP context, the impact of geminate devoicing on signal specificity is smaller and may be proportional to changes in predictability resulting from word frequency differences or variation in word length. Confirming this requires further theoretical development around the quantification of predictability and signal specificity. Overall, although some crucial questions remain, the current theory offers a framework for discovering why and where variation occurs.
SELECTED REFERENCES: [1] Coetzee, A. & J. Pater. (2011). The place of variation in phonological theory. In Handbook of Phonological Theory. 2nd Ed, 401-434. [2] Hall, K, E. Hume, F. Jeager, & A. Wedel (2016) The message shapes phonology. Ms. [3] Kawahara, S. (2006) A faithfulness ranking projected from a perceptibility scale. Lg 82:536–574. [4] Kawahara, S. 2011. Aspects of Japanese loanword devoicing. JEAL 20:169–194.
