Categorical and gradient homophony avoidance: Evidence from Japanese

2.1 Effects within morphological paradigms

The literature contains numerous descriptions of patterns in which a phonological rule exceptionally applies, or fails to apply, just in those cases where it would yield homophony between members of the same morphological paradigm. Perhaps the clearest examples are those of Crosswhite (1999), who analyzes two Slavic languages along these lines. First, Crosswhite describes a reduction pattern in Trigrad Bulgarian, where unstressed mid back vowels generally become [a], as illustrated in (1). In a handful of cases, this rule fails to apply: unstressed /o/ is permitted to surface in some specific morphological environments. Crosswhite notes that these exceptional cases involve a final /-o/ that might otherwise be confused with an /-a/ suffix: the masculine animate ending /-o/ is unchanged because the related accusative suffix is /-a/, and the neuter singular ending /-o/ is similarly unaffected because the neuter plural suffix is /-a/. These competing suffixes, illustrated in (2), crucially fail to block reduction when the relevant forms can be distinguished by other means – for example, by a stress shift, as shown in (3). (In all of the following examples, the segments of interest are underlined.)

Crosswhite observes a similar pattern in Russian, where unstressed /a/ generally reduces to [i] after palatalized consonants and [ə] after non-palatalized consonants. Exceptionally, /ă/ reduces to [ə] after palatalized consonants in the third plural forms of certain verbs; in these cases, reduction to [i] would result in identical singular and plural third-person forms.

A crucial component of Crosswhite’s analysis is the observation that these blocking effects occur only for homophones within the same morphological paradigm. In Russian, for example, the same vowel reduction pattern applies freely even when it creates homophones involving morphologically unrelated words. Thus, as illustrated in (5), the plural forms of the adjectives meaning ‘frequent’ and ‘clean’ both reduce to the same surface form, [ʧ^jistatá]. This homophony is allowed because the two forms have different stems; crucially, the singular and plural forms of each adjective are distinct from each other.

Other examples of homophony-sensitive patterns in the literature, summarized in Table 1, have the same property: in every case, when some rule is blocked from applying to a specific form, the potential homophone that blocks the rule belongs to the same morphological paradigm as the form to which the rule fails to apply. ^[2] The crucial role played by morphology is captured by formal analyses that refer specifically to inflectional morphology, such as the ParadigmContrast constraint in Optimality Theory (Kenstowicz 2002; Itô and Mester 2004). Crosswhite (1999) suggests that this is a sensible restriction: to determine whether a particular form has a potential homophone anywhere in the lexicon requires a large number of comparisons; by contrast, related words in a morphological paradigm present a much more restricted search space. Another argument could be made for the primacy of morphological paradigms on the grounds that it is particularly important for morphologically distinct forms to be phonologically distinct as well (Kurisu 2001; Kenstowicz 2013).

Table 1:

Examples of regular phonological processes blocked or altered by potential intra-paradigmatic homophony. Arrows indicate blocking among potential homophones.

Language	Rule affected by homophony	Competing forms	Examples
Trigrad Bulgarian (Crosswhite 1999)	/ŏ/ → [a] blocked	neut. pl.
		neut sg.
		masc. anim. acc.
		masc. anim. nom.
Russian (Crosswhite 1999)		3 pl.
	/ă/ → [i]/Cj __	3 sg.
	/ă/ → [ə] instead
Dakelh (Gessner and Hansson 2004)		trans.
	/sɬ/ → [s]	intrans.
	/sɬ/ → [ɬʌ] instead
Chi-Mwiːni (Kisseberth and Abasheikh 1974)		perf.
	/lVlVl/ → [lVl]	appl. perf.
	blocked
Damascene Arabic (Kenstowicz 2002)	/CCVCVCV/ → [CCV́CCV]
	blocked + stress shift	1/2 sg. subj., masc. sg. obj.
		3 sg. fem. subj., masc. sg. obj.
Iraqi Arabic (Blevins 2003)		base
	/CiCiVCi/ → [CiCi]	class II
	blocked

It is worth noting at this point that whether these patterns are genuine cases of synchronic homophony avoidance remains a point of debate in the literature. For Dakelh, Gessner and Hansson (2004) explicitly argue that the relevant pattern involves diachronic syncope blocking, not synchronic epenthesis; Mondon (2009) argues for a diachronic account of all patterns of this type. Urbanczyk (2005), on the other hand, argues that at least some anti-homophony patterns cannot be explained diachronically. For our purposes, the important point is that all of the plausible examples of synchronic homophony avoidance involve words with the same stem. Thus far, then, the published literature appears to find a consistent pattern: there is no such thing as a phonological rule that is blocked specifically when it would create a pair of homophonous surface forms, except possibly in the special case where the two forms are morphologically related. Phonologists working in Dispersion Theory, where the grammar is given the power to regulate contrasts among surface forms, have reached the similar conclusion that the grammar is simply not sensitive to the actual lexical items of a language. Padgett (2003) makes this argument explicitly:

The idea of neutralization avoidance, if understood in the wrong way, can make strange predictions. For example, consider the fact that Standard English has the words beat [bit], boot [but], and peat [pit], but no poot [put].... [I]f there were a process backing [i] to [u], would we expect that it might affect [pit] but not [bit], since only the latter would entail a neutralization (with [but])?

These questions arise when we take the domain of explanation to be the set of actual lexical items in a language. But this is in fact not the practice in generative phonology. Instead, theories model the set of possible words of a language.... (Padgett 2003: 78–79)

2.2 Japanese nasal contraction

Ichimura (2006) is, to our knowledge, the single exception in the literature to the above generalization: Ichimura claims that a particular reduction pattern in Japanese fails to apply when it would create homophonous pairs of words, even when those words have different stems. In this section, we review the pattern as described by Ichimura and consider whether it is a genuine counterexample to the generalization previously established in the literature. The relevant pattern is nasal contraction. ^[3] This is an optional rule in casual speech that takes the general form /rVn/ → [nn]; it applies across a variety of stem-suffix boundaries and involves deletion of a vowel. For our purposes, the most interesting sub-case is the behavior of the negative verbal suffix /-(a)nai/.

The form of /-(a)nai/ is phonologically determined: it appears as /-anai/ if the preceding stem ends in a consonant and as /-nai/ if the stem ends in a vowel. Nasal contraction is indifferent to the morphological affiliation of the deleted vowel; it may be the /a/ of the suffix, as in (6a), or the final vowel of the stem, as in (6b).

The examples in (6) illustrate the two classes of Japanese verbs that may participate in this pattern: those whose stems end in r and those whose stems end in re. ^[4] Herein lies the potential for homophony. If two verbs are identical except for the presence or absence of a final stem vowel, that vowel will be the only clue to the identity of the verb in the negative form. If nasal contraction applies, it will delete that crucial vowel, rendering the two surface forms homophonous.

Ichimura’s observation is that nasal contraction never actually produces ambiguous forms. Whenever there is a pair of verb stems that might be rendered homophonous, nasal contraction systematically fails to apply to one of them:

Furthermore, Ichimura observes that it is always contraction of the re-verb that is blocked, never contraction of the r-verb. He argues that this cannot be a general restriction against nasal contraction in re-verbs, since contraction does apply to forms such as /kure-nai/ in (6b). Ichimura concludes that nasal contraction is sensitive to homophony between words with different stems: it cannot apply to a re-verb if the result would be homophonous with a corresponding r-verb.

To test whether this is a productive generalization, Ichimura conducted an experiment in which native speakers of Japanese produced r- and re-verbs. He found that nasal contraction systematically applied to both real and nonce r-verbs, but that it failed to apply to re-verbs (whether real or nonce) with a potentially homophonous r-verb counterpart. Ichimura concludes that the pattern is indeed a productive one, and that it is conditioned by homophony between unrelated lexical items.

2.3 An alternative account of nasal contraction

If Ichimura is right, then current approaches to phonology are too restrictive: it is possible (although apparently not common) for phonological patterns to be sensitive to relationships between morphologically unrelated words, and therefore any descriptively adequate theory of phonology must be able to make such comparisons. It turns out, though, that there are reasons to doubt that nasal contraction in Japanese is actually homophony-avoiding in this way.

The crucial part of this pattern is the behavior of re-verbs. Ichimura notes that re-verbs are less likely in general to undergo nasal contraction than r-verbs; contraction is often impossible for a given re-verb even if it has no r-verb counterpart. In other words, absence of a potential homophone is a necessary but not a sufficient condition for re-verbs to contract. Ichimura compiled a list of real r- and re-verbs and found that nasal contraction was possible for the overwhelming majority of the r-verbs (155 out of 157) but less than two-thirds of the re-verbs (23 out of 37). Thus, the picture presented in Ichimura (2006) is one in which nasal contraction is (almost) fully productive for r-verbs but only moderately productive for re-verbs. ^[5]

By contrast, our own observations suggest that the productivity of nasal contraction for re-verbs is extremely limited. The second author compiled a comprehensive list of re-verbs from one of the most comprehensive Japanese dictionaries available (Iwanami Shoten 2008, Kôjien, 6th ed.), and elicited informal judgments from herself and two other native speakers of Japanese as to whether they could contract. Of the 149 re-verbs, only 31 were deemed eligible for contraction by at least one judge; these verbs are listed in Table 2. At 20.9%, this is a much lower rate of contraction for real re-verbs than Ichimura found; moreover, in 19 cases, only the most permissive judge (A) believed that contraction was possible. It is also interesting to note that several of these re-verbs do have potentially homophonous r-verb counterparts, contrary to the description of the pattern presented above.

This informal survey suggests that, in practice, very few re-verbs can actually undergo nasal contraction. It is possible that at least some of the inter-judge variation reflects social evaluations of this pattern rather than actual practice. Judge A, for example, explicitly stated that nasal contraction is something males are more likely to do (an intuition shared by the second author); because Judge A is male himself, it is possible that his judgments were influenced by his evaluation of nasal contraction as a marker of male speech, perhaps similar to the “covert prestige” of certain sociolinguistic variables documented by Trudgill (1983: Ch. 10). Whatever the ultimate explanation for the differences between judges, the overall picture is clear: contraction of re-verbs is limited.

These results, combined with the native-speaker intuitions of the second author, lead us to suspect that nasal contraction is simply not productive for re-verbs: a few re-verbs may exceptionally undergo contraction, but the vast majority cannot. If our hypothesis is correct, then it is not meaningful to describe the restrictions on re-verbs as homophony-avoiding. Diachronically, anti-homophony pressures may have helped guide the selection of which re-verbs (such as kureru) are exceptionally able to contract; but synchronically, aside from these exceptions, Japanese speakers simply fail to apply contraction to re-verbs in general, and thus there is no need to compare actual lexical items before deciding whether or not to contract a given re-verb.

To determine whether this account of nasal contraction is correct, we need to know whether Japanese speakers are willing to apply nasal contraction to nonce re-verbs with no potential homophone. Unfortunately, Ichimura’s experiment did not include stimuli of this type, and therefore it is impossible to learn from his results whether nasal contraction is in fact productive for re-verbs. To remedy this problem, we conducted a similar experiment designed to explore the range of verb types for which nasal contraction is productive.

3.1 Subjects

Study participants were 10 native speakers of Japanese, tested in pairs. All participants were Japanese college students participating in an exchange program at the University of Utah; further demographic information is provided in Table 3. Subject pair 1 was run to pilot the experimental procedure; their data is omitted here because subject 1A had participated in a previous pilot version of the experiment. Numbering of the other subject pairs is not continuous because some scheduled pairs cancelled their sessions before data could be collected. All subjects gave written consent to participate in the experiment.

All of the subjects used standard Japanese during the experiment; however, four of the subjects (3B, 4B, 8A, and 8B) had a noticeable accent – particularly subjects 3B, 4B, and 8B, who had strong western accents. It is possible that these subjects were less likely to use nasal contraction because of their native dialect, in which negation is realized with a different suffix (-hen) to which nasal contraction does not apply. Every subject in the study produced nasal contraction at least some of the time, despite the fact that the pattern was not explicitly mentioned at any point during the experiment; thus, the pattern is present to at least some degree in the speech of every subject. Moreover, we have no reason to believe that dialect differences would affect the contraction of r- and re-verbs differently. Therefore, we believe that our data is capable of telling us something about the behavior of nasal contraction in Japanese in general, so long as it is interpreted with an appropriate degree of caution.

As noted above, we also suspect that nasal contraction is more frequent in the speech of men than in that of women, particularly in all-male groups. Since most of our subjects were female, we may have observed a lower rate of contraction than we would have seen with more male subjects. Again, though, we have no reason to believe that gender differences affect r- and re-verbs differently.

3.2 Stimuli

Stimuli consisted of 36 real and nonce verbs. As summarized in Table 4, half of the stimuli were r-class verbs and half were re-class verbs. Within each class of verbs, we included both real and nonce verbs with and without potential homophones in the other class; thus, these stimuli cover all possible combinations of verb class, real/nonce status, and existence of a potentially homophonous verb.

Table 4 also summarizes the predictions for the behavior of each class of stimulus items made by three different hypotheses about the nature of nasal contraction. Ichimura’s anti-homophony account (“Anti-homophony”) predicts that the only verbs that will not contract are re-verbs with a potentially homophonous r-verb counterpart, regardless of whether the re-verb is real or nonce. By contrast, if nasal contraction applies productively only to r-verbs (“r-verbs only”), then none of the nonce re-verbs are predicted to contract. Some of the real re-verbs might behave like kureru and exceptionally contract (indicated by a question mark in the table), but they should be less likely to do so than r-verbs. Finally, the inclusion of nonce r-verbs allows us to test for the unlikely possibility that nasal contraction is simply not productive at all (“Nonproductive”), in which case nasal contraction should apply only to real r-verbs (and possibly a few real re-verbs). Finally, the real r-verbs allow us to assess whether the experimental procedure successfully induced subjects to use this optional pattern in the first place.

The crucial stimuli are the nonce re-verbs with no potential homophone; these are the stimuli that most clearly distinguish between Ichimura’s anti-homophony account (which predicts that they will contract) and the alternative account in which contraction is not productive for re-verbs (which predicts that they will not). It is precisely this class of stimuli that was omitted in Ichimura’s original study; thus, these are the items that will allow us to test more accurately whether anti-homophony is a crucial factor in the productivity of nasal contraction.

Table 5 gives the actual stimuli used in the experiment. Real verbs are listed with their meanings; nonce verbs are listed with the meaning that was assigned to them during the experiment (in parentheses).

3.3 Procedure

All experimental procedures were conducted at the University of Utah under the approval of the IRB. Subjects were tested in pairs; each pair consisted of two subjects who were already friends and agreed to come in to the lab together. The purpose of this procedure was to encourage more casual speech, since nasal contraction is an optional process that is less likely to apply in formal settings.

During the experimental session, subjects were seated across a desk from one another with a microphone placed between them and a computer screen to one side. Subjects were told that they would be playing a game in which they would be given a series of topics to discuss; for each topic, they would have 2 minutes for their discussion, during which they would attempt to use a particular word as many times as possible. Games similar to this one are sometimes played on Japanese game shows; thus, the task was not an entirely novel one. To demonstrate for the subjects what was being asked of them, the experimenter played a sample game that had been recorded during pilot testing. The target word for each discussion was one of the stimulus verbs from Table 5.

At the start of each trial, the computer would play a warning sound and display the text of the next discussion topic and the new target word; subjects also heard a recorded female voice reading the text. When the target word was a nonce verb, a definition was provided. Stimuli were presented in random order; the entire session lasted about an hour and a half. See the Appendix for the experimental instructions and sample discussion topics.

3.4 Results

The second author listened to the recordings and marked each occurrence of each target verb. Tokens were eliminated from analysis for any of the following reasons:

–

• The target verb was mispronounced or was too difficult to hear (12 tokens, 0.81%). In example (9), the subject mistakenly produced taorenaitte, a form of taoreru (a real re-verb), rather than the target nonce verb taoru.

–

• The subject was repeating the prompt, or appeared to be practicing the target verb rather than using it in context (122 tokens, 9.5%). In example (10), all negative productions of the target verb okuru were eliminated for this reason.

–

• The production sounded forced or extremely self-conscious, as though the subject were saying the verb out of context in order to meet his or her quota (65 tokens, 4.4%). In example (11), the subjects have been discussing whether they stay up late in Utah. When subject 2B inquires about subject 2A’s previous habits in Japan, subject 2A responds with a present-tense form of the target nonce verb haguru, rather than the past-tense form that would have been natural in context.

In all, 199 tokens (13% of the total) were eliminated from further analysis.

Each remaining token was coded by the second author as either contracted or not contracted. These tokens were then extracted from the longer recordings with up to one second of context on either side (less if another token of the target word appeared in that window). The extracted tokens were played in a random order for three other raters, all native speakers of Japanese who were naïve to the purposes of the experiment. As summarized in Table 6, inter-rater agreement was reasonably good, although not perfect; in particular, Rater 3 was far more conservative than the other raters in identifying tokens as contracted. In the following analyses, we adopt the criterion that a particular token was contracted if it was rated as contracted by at least two of the four raters; 128 tokens met this criterion. Under a more stringent requirement that only tokens rated as contracted by at least three raters are counted (108 in all), all of the significance results reported below remain unchanged at α = 0.05, except where otherwise noted.

All statistical analyses were conducted in R (R Core Team 2014). Overall, subjects were quite successful at using the target verbs during the allotted 2-minute discussion window. Subject pairs produced an average of 8.9 tokens of each verb between them on a given trial; the actual number of tokens ranged from 2 to 24. Unsurprisingly, subjects produced more tokens of real verbs (an average of 10.0 per verb per pair) than of nonce verbs (an average of 7.6), a statistically significant difference (W = 3367, p = 0.0011 by a two-sample Wilcoxon test). Despite this difference, it is clear that the procedure successfully encouraged subjects to produce a large number of nonce verbs as well as real ones.

The procedure was also moderately successful at encouraging subjects to use nasal contraction. Of real r-verbs, which are predicted by all analyses to contract, subjects contracted 27.0% (110) of 407 tokens; of nonce r-verbs, subjects contracted only 6.0% (16) of 267 tokens. It appears, then, that nasal contraction does indeed apply productively to new verbs, although perhaps not as robustly as it does to existing verbs; it is unclear to us whether the lower rate of contraction for nonce words demonstrates that the pattern is not fully productive, or is due to the fact that subjects were uncomfortable with the novel words and therefore spoke less casually.

By contrast, subjects did not produce nasal contraction for any re-verbs at all, whether real (389 tokens) or nonce (220 tokens). This result is consistent with the hypothesis that nasal contraction does not apply productively to re-verbs at all, but instead applies exceptionally to just a handful of verbs such as kureru. Given the low rate of contraction among r-verbs, we cannot rule out the possibility that re-verbs do contract at a low but non-zero rate. Our results do, however, support the claim that re-verbs are at least less likely to contract than r-verbs; among nonce verbs (summarized in Table 7), re-verbs are significantly less likely to contract than r-verbs by Fisher’s exact test (p = 1.02 × 10⁻⁴). The same holds true if we consider only nonce verbs with no potential homophone (those predicted by Ichimura to contract): the difference between r-verbs and re-verbs remains statistically significant (p = 0.015). Thus, our results provide no evidence that nasal contraction applies productively to novel re-verbs.

To further understand the factors that affected whether or not subjects were willing to apply nasal contraction, we fitted a mixed-effects logistic regression model using the lme4 package (Bates et al. 2014) to the data for r-verbs only. (It was not possible to include both r- and re-verbs because the Class factor results in quasi-complete separation of the data, since no re-verbs were contracted.) We began with a model that predicted contraction of each token from a fixed effect of Frequency and a random effect of Subject. Log frequencies from the NINJAL corpus (National Institute for Japanese Language and Linguistics 2014) were used; nonce verbs were assigned a log frequency of zero, and the smallest raw frequency of any real verb was 2. ANOVAs revealed that adding a fixed effect of Homophone (whether or not there exists a potentially homophonous re-verb) improved the model significantly (χ²(1) = 18.4, p = 1.8 × 10⁻⁵), but adding a fixed effect of Nonce (whether or not the target was a nonce verb) did not, regardless of whether Nonce was added to the base model (χ²(1) = 2.96, p = 0.085) or to a model with Homophone (χ²(1) = 2.53, p = 0.11). However, adding an interaction between Nonce and Homophone resulted in a marginally significant improvement over a model with just the two main effects (χ²(1) = 4.15, p = 0.042). ^[6] This final model is summarized in Table 8.

As expected, more frequent verbs were more likely to contract. Moreover, the model provides further support for the pattern we observed informally above, that subjects were less likely to contract nonce verbs than real verbs. Most interestingly for our purposes, the model reveals an additional effect of homophony: subjects were less likely to contract an r-verb if there exists a corresponding re-verb that is a potential homophone. Finally, the positive estimate for the interaction between Nonce and Homophone shows that this effect of homophony is attenuated for nonce verbs.

3.5 Discussion

Our results do not support the absolute blocking pattern described by Ichimura. Where Ichimura argues that re-verbs fail to contract specifically when contraction would result in an ambiguous surface form, we find no evidence that contraction applies productively to re-verbs at all. Although a handful of re-verbs (such as kureru) do contract, we did not observe any contraction of novel re-verbs, regardless of the existence of a potential homophone; thus, at the very least, our results provide no evidence that re-verbs with potential homophones are different from other re-verbs. In fact, the pattern we observe looks like an ordinary root-affix asymmetry: nasal contraction is much more likely to delete a vowel that is part of an affix (as in wakar-anai) than one that is part of a root (as in wakare-nai). Thus, the pattern can be easily understood as an example of positional faithfulness (Beckman 1998).

We conclude that productive nasal contraction in Japanese is restricted to r-verbs; it is not absolutely blocked by potentially homophonous words with other stems. On the other hand, our results do support Ichimura’s more general suggestion that nasal contraction is sensitive in some way to homophony. Although we failed to find evidence for the absolute anti-homophony effect he describes, we did find evidence for stochastic avoidance of homophony among r-verbs. Speakers in our experiment were significantly less likely to contract r-verbs with potentially homophonous re-verb counterparts, although they did not categorically refrain from doing so.

This probabilistic homophony avoidance is interesting for two reasons. First, the homophony that is apparently being avoided is entirely hypothetical, given the fact that the potential homophones in these cases are all re-verbs and therefore never actually contract. When speakers contract okur-anai to okunnai, for example, there is a real sense in which the surface form okunnai is not actually ambiguous: the corresponding re-verb okure-nai never contracts to okunnai at all. Thus, the fact that speakers avoid contraction even when the potential homophony would never actually be realized is surprising.

The second interesting aspect of our results is the fact that this anti-homophony effect was apparent only among real verbs; indeed, the size of the estimate for the interaction in Table 8 is more than enough to eliminate the effect of Homophone among nonce verbs entirely. At least two possible explanations for this difference suggest themselves. One is that speakers are simply not sensitive to potential homophony when they are producing nonce verbs. The other is the fact that the potential re-class homophones were much more contextually salient in this experiment for real verbs than for nonce verbs. For each real target r-verb, the corresponding re-verb was also a target during a different trial of the experiment; for example, both huru and hureru were used as target words. By contrast, the potential re-verb homophones of the nonce r-verbs were not used as targets during the experiment: subjects saw haguru but not hagureru. It is possible, then, that the apparent difference between real and nonce verbs is actually an effect of the contextual salience of the potential homophone.

We find some evidence for this second explanation when we look at presentation order effects. If subjects are reluctant to contract huru because of the contextual salience of hureru (as opposed to the mere fact that hureru exists), then subjects should be less likely to contract huru if they have already seen hureru in the experiment than if they have not. It turns out that this is exactly what happened: subjects contracted 29.5% of real r-verbs (38 of 129) whose re-verb counterpart they had not already seen during the experiment, but only 11.1% of real r-verbs (15 of 135) whose re-verb counterpart they had seen in a previous trial. A mixed-effects model of these r-verbs with a fixed effect of Frequency and a random effect of Subject leads to the same conclusion: there is a statistically significant effect of Order (p = 0.0025), such that verbs were more likely to contract if their re-verb counterpart had not already been seen.

This second interpretation also suggests an explanation for the first puzzle, that these re-verbs are ‘potential homophones’ only in the most hypothetical of senses. What we see here as an effect of homophony might be an example of the more general tendency for speakers to produce less reduction for words with many or salient competitors (Munson and Solomon 2004; Baese and Goldrick 2009; Goldrick et al. 2013). Indeed, homophony avoidance itself might be simply a specific case of a more general phenomenon of hypercorrect speech for ‘hard’ words.