Systematic Irregularity in Japanese Rendaku: How the grammar mediates patterned lexical exceptions

Abstract

Exceptions to Japanese rendaku voicing that are independent of Lyman's Law have usually been considered to be random and unsystematic. This article proposes that such exceptions are largely systematic and can be explained through lexical specification and Positional Markedness. Two main types of systematicity are examined: the clustering of blocking cases around particular lexical items, and a prosodic size effect, where "long" compounds, with at least one constituent exceeding two moras, will disable blocking under most conditions. Lexical clustering is explained through lexical specification of features under Combinatorial Underspecification while the prosodic size effect is seen as an expression of Positional Markedness. It is argued that only in long compounds is the morpheme boundary at the edge of a ProsodicWord, a prosodically strong position that more freely permits the marked [sonorant, +voice] featural combination of rendaku voicing to occur.

1. Introduction

Rendaku voicing in Japanese has been subject to much recent investigation (McCawley 1968; Vance 1980; Otsu 1980; Itô and Mester 1986, 1998; Mester and Itô 1989), but previous analyses have focussed mainly on exceptions that are due to Lyman's Law (Lyman 1894; Ogura 1910). This article examines exceptions to rendaku voicing that are independent of Lyman's Law and which much of the relevant literature either ignores or treats as random and unsystematic (McCawley 1968; Otsu 1980; Vance 1987). I argue that there is in fact a systematicity to these exceptions which can largely be accounted for by the effects of the grammar on possible output forms, if input representations are sufficiently specified.

Exceptions to regular phonological processes are known to often occur in systematic ways; for example, the irregular forms of the English past tense show phonological patterning (Bybee and Slobin 1982; Stemberger and MacWhinney 1988; Myers 1999). Within rule-based theories, exceptions to regular processes have been explained through exception features (Aronoff 1976; Kiparsky 1982). This article proposes that the systematicity of exceptions to rendaku voicing can be accounted for in a principled way through featural specification.¹

The patterning of exceptions in phonology is also addressed by such recent work as Zuraw (2000) and Albright and Hayes (2002), both of whom argue that systematic irregularity can be encoded in the speakers' mental representation of the generalizations of their language. For example, Zuraw (2000:24) observes that exceptions to regular lexical patterns do not occur randomly but are more likely to occur under certain phonological conditions. She proposes that knowledge of lexical regularities occurs in a speaker's grammar in the form of low- and variably-ranked constraints.

Albright and Hayes (2002) propose a learning model through which a speaker might learn the kinds of detailed generalizations that constitute a knowledge of the patterning of exceptions. They give the example of English past tense often changing [I] to [Λ] when the final consonant is a velar nasal, arguing that English speakers do learn such generalizations. They refer to experimental evidence which demonstrates that speakers can make predictions about the form of novel words based on this kind of detailed knowledge.

If speakers do in fact have knowledge of the patterning that irregular forms take, then our task is to explain how, in current models of grammar, such knowledge is represented. In this context, the goals of this article are to show that (a) exceptions to rendaku voicing are strongly systematic and (b) that this systematicity can be encoded in a speaker's knowledge through (i) an appropriate choice of underlying representations and (ii) a grammar that is sensitive to the kinds of phonological environments in which these exceptions are likely or are not likely to occur.

1.1. Rendaku voicing and its irregularity

The process of rendaku voicing (McCawley 1968; Vance 1980; Itô and Mester 1986, 1998; Mester and Itô 1989) voices the initial obstruent of the second member of a Japanese compound word and applies mainly to compounds of Yamato (native Japanese) origin. Although rendaku voicing is notorious for its lexical variation and unpredictability (McCawley 1968:87, note 18; Otsu 1980:1; Vance 1987), recent accounts of rendaku voicing have tended to abstract away from its irregularity in order to deal with phonologically regular processes such as Lyman's Law, which blocks rendaku in a 99.9% predictable fashion (see, e.g., Itô and Mester 1986, 1998). Rendaku occurs in about 75% of Yamato noun-noun compounds that have the right phonological conditions to trigger it, namely, the presence of an initial voiceless obstruent (that is not followed by a voiced obstruent) in the simplex form of the second member.

The irregularity of rendaku is illustrated by pairs such as the following, in which rendaku mysteriously fails to occur in one compound but not the other, in spite of apparent phonological, morphological, and semantic similarity between the two compounds:

Rendaku voicing occurs in kuse 'habit' in (1a) but not in (1b). Both compounds in (1a, b) are unaccented. Both kuti and asi are bimoraic, end in a high front vowel, and contain voiceless obstruents. There is thus no obvious phonological, morphological, or semantic difference between the two compounds that could account for the lack of voicing in (1b). Moreover, the meanings of the two compounds are parallel, where both compounds have a first member that refers to a body part that is the focus of the particular habit that the compound refers to. Such a semantic near-minimal pair rules out the possibility that morpho-semantic factors could determine whether voicing occurs or not in noun-noun compounds such as these.

Further examples of irregularity in rendaku voicing are the compounds in (2) with second member hara 'field', which has a pitch accent on the initial mora as a simplex noun. The noun hara fails to voice in compounds with the initially accented bimoraic noun sino 'bamboo' in (2a) but voices in compounds formed with the initially accented bimoraic nouns tono 'nobility' and matu 'pine' in (2b, c). All three compounds in (2) have a pitch accent on the second mora.

In example (3), the noun kawa 'skin, hide', which has a pitch accent on the second mora as a simplex noun, resists voicing when it combines with the unaccented bimoraic noun himo 'cord' but voices when it combines with the unaccented bimoraic noun momi 'unhulled rice'. Both compounds in (3) are unaccented.

Seen from the perspective of an individual noun, rendaku voicing indeed appears completely unpredictable. But on a global scale, as shall be elaborated on in section 2, rendaku voicing will be seen to exhibit a strong degree of patterning.

1.2. Theoretical framework and assumptions

The analysis presented here adopts the constraint-based framework of Optimality Theory (Prince and Smolensky 1993), which derives output forms through a ranked set of violable constraints. Within this framework, I adopt the following assumptions:

Rendaku voicing occurs because of a junctural morpheme whose underlying form is a floating [+voice] feature (Itô and Mester 1986:57, 1998:26).

"Combinatorial Underspecification" (Archangeli and Pulleyblank 1994). Under this proposal, more than one possible input form may occur for one particular combination of features in the output. For a binary feature F, any of three possible input values may occur: (a) +F, (b) -F, or (c) 0F, where both 0F and one of +F or -F will surface with the same value. This hypothesis reflects the principle of "Richness of the Base" in Optimality Theory.

"Lexical prespecification" (Kiparsky 1993; Inkelas 1995; Inkelas, Orgun, and Zoll 1996). This proposal argues that exceptions to phonological processes should be accounted for through prespecification of a feature that blocks the process, rather than by partitioning the grammar into separate cophonologies (Inkelas and Orgun 1995; Itô and Mester 1995) where cophonology A accounts for the regular process and cophonology B for exceptions to the process.

2. Systematicity of exceptions to rendaku in noun-noun compounds

We now examine evidence that rendaku voicing is not only irregular, but that its irregularity exhibits clear systematicity. Section 2.1 sketches out the nature of rendaku voicing. Section 2.2 discusses the corpus of data that was used to identify patterns of rendaku voicing and blocking. In sections 2.3 to 2.5, we identify three ways in which rendaku blocking is not completely random: (a) the existence of nouns that are immune to rendaku, (b) the effects of prosodic size on rendaku blocking, and (c) lexical clustering of blocking cases.

2.1. Compounds that are eligible for rendaku voicing

Excluded from the process of rendaku voicing are (a) dvandva or headless compounds which predictably block rendaku voicing (see Sakurai 1966:41; Okumura 1980; Otsu 1980; Nishikawa 1987; Vance 1987) and (b) compounds whose second member contains a non-initial voiced obstruent, which predictably blocks voicing through Lyman's Law (Itô and Mester 1986, 1998). The present analysis considers only compounds that are eligible for rendaku voicing, that is, non-dvandva compounds that have an initial voiceless obstruent and no voiced obstruent in the second conjunct in its surface simplex form.

2.2. Corpus of data

The relevant data, a corpus of 1004 headed Yamato noun-noun compounds, was gathered from a list of about 1200 Yamato monomorphemic nouns in Martin's (1987) inventory of Yamato nouns, by excluding all nouns that are transparently polymorphemic. Using this database, a list of rendaku-eligible noun-noun compounds was developed from entries in the NHK²Japanese Accent and Pronunciation Dictionary (1999) and the Shogakukan Progressive Japanese-English Dictionary (1993). A few further compounds were added from entries in Breen (2003). This corpus reveals two significant patterns in the blocking of rendaku voicing, elaborated in sections 2.3 to 2.5. One is a clustering of rendaku blocking around groups of nouns that either frequently resist rendaku or show complete immunity to rendaku voicing. The second is a prosodic size threshold which, when exceeded by a compound, disables blocking of rendaku by rendaku resisters. We begin by examining, in section 2.3, nouns that are immune to rendaku.

2.3. Immunity to rendaku

As pointed out by Martin (1987) and Vance (1987), a small number of Yamato nouns, such as the following, never undergo rendaku voicing:

For example, hime 'princess' only occurs in compounds as hime, not *bime (e.g., uta hime (song-princess) 'songstress', oto hime 'youngest princess', ori hime 'woman textile worker').³

Another rendaku-immune noun in (4) is kasu 'dregs', which, like other rendaku immune nouns, fails to voice in compounds even when the compound is long (i.e., has at least one member that exceeds two moras). In such long compounds, rendaku voicing is never blocked, except in the case of rendaku-immune nouns (see section 2.4). In the compound abura kasu 'oil dregs', for example, voicing fails to occur, even though its first conjunct abura 'oil' exceeds two moras. Its lack of voicing cannot be due to any property of the noun abura, because all other rendaku-eligible compounds of which it is the first member will undergo voicing:

Other rendaku-immune nouns also fail to voice in long compounds:

These facts distinguish nouns like kasu 'dregs' from "rendaku resisters", discussed below, which, even if they resist voicing in short compounds, always voice in long compounds (see section 2.4.2).

2.4. Rendaku resisters and prosodic size effect

A second class of nouns also resists rendaku voicing, but unlike rendaku-immune nouns, these nouns (a) do undergo voicing in a minority of "short-short" compounds (see section 2.4.1) and (b) always voice in long compounds. Excluding the cases of rendaku-immune nouns (above), exceptions to rendaku voicing occur solely in compounds in which neither member exceeds two moras. Following Kubozono (1995, 1996), I adopt the label "short-short" to refer to this type of compound. We first examine the behaviour of some rendaku resisters in short-short compounds.

2.4.1. Blocking of rendaku by resisters in short-short compounds

In (7) and (8) are examples of short-short compounds with the second member kusa 'grass'. We find 16 such compounds that do not voice (7), and only three that do (8). We can thus label kusa a "rendaku resister" because it resists voicing in most but not all short-short compounds in which it occurs.

(8) Short-short compounds with kusa that voice:

The noun kuse 'habit' occurs in fewer compounds than kusa 'grass' but also robustly resists rendaku, failing to voice in six out of eight short-short compounds.

(9) Short-short compounds with kuse that do not voice:

(10) Short-short compounds with kuse that voice:

The noun saki 'tip'⁴ looks almost like a rendaku-immune noun, blocking voicing in all known noun-noun compounds except for the reduplicated form saki zaki 'distant future':

(11) All non-reduplicated forms with second conjunct saki block voicing:

The nouns hara 'field' (12)–(13), ki 'tree' (14)–(15), and te 'hand' (16)–(17) block rendaku in more than half of the short-short compounds in which they occur.

(12) Short-short compounds with hara 'field' that block voicing:

(13) Short-short compounds with hara 'field' that voice:

(14) Short-short compounds with ki 'tree, wood' that block voicing:

(15) Short-short compounds with ki 'tree, wood' that voice:

(16) Compounds with te 'hand' that block voicing:

(17) Compounds with te 'hand' that voice:

2.4.2. Failure of blocking by resisters in long compounds

We have just observed that rendaku-resisting nouns robustly resist rendaku in short-short compounds; however, they differ from rendaku-immune nouns in two ways: (a) they do not block voicing in 100% of cases; (b) they never block rendaku when they occur in long compounds. Not all rendaku resisters actually occur in long compounds but those that do never fail to voice. Note that the important long compounds for us to examine here are those with a long first member and a short second member that resists rendaku in short-short compounds. Compounds with a long second member always undergo voicing unless they are rendaku-immune. The following data show long compounds whose second conjunct is one of therendaku-resisting nouns examined in section 2.4.1. In these data, all first conjuncts exceed two moras:

To add to the data above, we can also examine long compounds whose second conjunct is a noun that blocks voicing in at least some short-short compounds. The examples below illustrate the voicing behaviour of the following nouns in short-short compounds: tori 'bird' (19), siru 'soup' (20), kumo 'cloud' (21), kawa 'hide' (22), and tama 'ball, jewel' (23). A few compounds with Sino-Japanese first members are also included for comparison (marked "SJ"⁵). Henceforth the notation "N2" is used to denote the second conjunct of a noun-noun compound.

(19) Compounds with N2 tori 'bird':

(20) Compounds with N2 siru 'soup':

(21) Compounds with N2 kumo 'cloud':

(22) Compounds with N2 kawa 'skin, hide, leather':

(23) Compounds with tama 'ball, jewel':

Out of the 52 short-short compounds listed above, blocking occurs in 13 cases, or about 25% of the time. We now examine compounds in which the same nouns occur as N2, but this time, with long first members.

(24) Long compounds with N2 tori 'bird':

(25) Long compounds with N2 tama 'ball, jewel':

(26) Long compounds with N2 siru 'broth':

(27) Long compounds with N2 kumo 'cloud':

(28) Long compounds with N2 kawa 'hide':

(29) Long compound with N2 tori 'bird':

miyako           'shrine'           miyako dori           'oystercatcher'

Out of the 31 long compounds listed in (24)–(29), blocking never occurs. This is in contrast to the measurable rate of blocking (25%) among the 52 short-short compounds with the same second conjuncts in the previous set of data (19)–(23). These results suggest that when we factor out compounds with rendaku-immune nouns, blocking is disabled in long compounds.

2.5. Clustering of blocking around rendaku resisters

We saw from the data in (7) through (13) that certain nouns such as kusa 'grass', hara 'field', kuse 'habit', te 'hand', ki 'tree', and saki 'tip' block voicing in compounds more frequently than not. On a global scale, cases of rendaku blocking do in fact cluster around particular lexical items rather than being randomly dispersed.

Factoring out compounds with rendaku-immune nouns,most cases of rendaku blocking in short-short noun-noun compounds occur among compounds headed by a small number of rendaku resisters: nouns that voice in only a small minority of compounds. From the 781 short-short compounds in our corpus, subtracting 49 that are rendaku-immune leaves 732. Of these, 167 block rendaku. Eighty-three, or 50%, of these cases occur among just eight nouns, as shown in Table 1.

Table 1.

Blocking cases among eight rendaku-resisting nouns

In addition to these particular nouns, there are other Yamato nouns that occur in fewer compounds, and which also resist voicing. For example, the nouns given in Table 2 never voice in known compounds. (The number of compounds in which the noun occurs as a second conjunct is given following each noun.)

Table 2.

Further nouns that resist rendaku voicing

Because each of these nouns (Table 2) occurs in only a few compounds, all of them short-short, it is difficult to tell if it is behaving as a rendaku resister or as a rendaku-immune noun. The actual number of nouns we can call rendaku resisters is therefore difficult to determine, since many compounds have second conjuncts that only occur in one or two compounds. The nouns in Table 2 could be either rendaku-immune or rendaku-resistant. I suggest that there is a class of nouns of indeterminate number that resist voicing of which the eight nouns listed in Table 1 are those that occur in enough compounds to determine whether they are rendaku-immune or rendaku-resistant. Whether or not these nouns have any other common characteristics has yet to be determined.

The eight rendaku-resisting nouns listed in Table 1 each occur in at least eight noun-noun compounds. If we look for other rendaku-eligible nouns that occur as a second conjunct in at least as many short-short compounds, we find 17 others, listed below. Thirteen of these, shown in Table 3, voice in every compound in which they occur. The remaining four, shown in Table 4, voice in 36 out of 48 compounds.

The set of nouns in Table 1 (83 cases of blocking in 119 compounds) added to the set in Table 3 (zero cases of blocking in 140 compounds) and Table 4 (12 cases of blocking in 48 compounds) form a set of 25 rendaku-eligible nouns that account for a total of 307 compounds. Of these, 119, or 39%, occur with the eight rendaku-resisting nouns in Table 1. These eight nouns account for 83 out

Table 3.

Nouns that voice in every compound

Table 4.

Nouns that voice in most compounds in which they occur

of 95 cases of blocking among the 307 compounds in Tables 1, 3, and 4. The resistant nouns in Table 1 have no obvious phonological or semantic feature that distinguishes them from the non-resistant nouns in Tables 3 or 4. For example, both kuse 'habit', a resister, and kami 'paper', which always voices, have final pitch accent. Both kusa, which resists rendaku, and kuti 'mouth', which robustly voices, have two voiceless obstruents. Examples of compounds formed with kami 'paper', one of the nouns listed in Table 3 that does not block voicing, are given in (30).

A summary of the lexical clustering of rendaku blocking among short-short noun-noun compounds is given in Table 5. Exceptions to rendaku voicing are not randomly dispersed, but cluster around particular nouns. In section 3, we examine how this kind of lexically based systematicity of exceptions can be accounted for through the appropriate choice of input representation for nouns that resist or are immune to rendaku voicing.

Table 5.

Summary of blocking patterns among Yamato noun-noun compounds

3. Analysis of the Systematicity of Rendaku Blocking

3.1. Exceptionless blocking by rendaku-immune nouns

The data in section 2 suggest that there are different classes of nouns that behave differently with respect to rendaku voicing: (a) rendaku-immune nouns, which never voice under any circumstances; (b) rendaku resisters which robustly resist voicing, but only in short-short compounds; and (c) the elsewhere case of all other nouns, which undergo rendaku voicing in most compounds. The three types can be distinguished lexically as follows. Rendaku-immune nouns have a [ -voice] feature that is linked to the root node of the initial obstruent (31a); rendaku resisters have a floating [ -voice] feature (31b), and non-resisters have an initial obstruent that is underspecified for voicing (31c). A capital "K" represents a velar stop minus its voicing feature.

(31) a.

Adopting the model of Combinatorial Underspecification (Archangeli and Pulley-blank 1994; see section 1.2), an obstruent can have three possible values for voicing in the input: [-voice], [+voice], and [0voice] (underspecified). The [0voice] feature will normally surface as [-voice] through the effects of a grounded markedness constraint for voicing in obstruents, of the type proposed by Archangeli and Pulleyblank (1994):

(32) IF [ -SON] THEN [ -VOICE]

Given the three types of underlying representations in (31), we can provide the following analysis. The linked [ -voice] feature of rendaku-immune nouns (31a) persists in all compounds, whereas the floating [ -voice] feature of rendaku resisters (31b) persists only in short-short compounds. In non-resisting nouns, which have no relevant underlying voicing feature (31c), the [+voice] feature of the junctural morpheme will link to the initial obstruent of the second conjunct in the output. I propose that the exceptionless surfacing of the [ -voice] feature in immune nouns is due to a constraint that specifically targets a linked feature as opposed to a floating one. The relevant constraint is as follows:

(33) DEP-PATH-FEATURE: For every path of association P' between a feature F' and a root node R' in the output, there is a corresponding path of association P between feature F and root node R in the input, where P corresponds to P', F corresponds to F' and R corresponds to R.' ⁶

Under the present proposal, the initial obstruent of a rendaku-immune noun has a path of association to a [ -voice] feature in the input; the initial obstruent of any other noun eligible for rendaku has no path of association to any voicing feature in the input. Thus, any output candidate derived from a non-immune noun will equally violateDEP-PATH-FEATURE whether it has voicing or not, since neither [+voice] nor [ -voice] is linked in the input (see (43)–(44)). However, DEP-PATH-FEATURE will rule out a candidate with initial voicing for a rendaku-immune noun, since only a candidate with initial [ -voice] will satisfy this constraint. Since rendaku-immune nouns block voicing without exception, I take DEP-PATH-FEATURE to be undominated.

We can account for the surfacing of the [+voice] feature of the junctural morpheme in non-blocking cases through a MAX faithfulness constraint for the feature [+voice], MAX[+VOICE] in (34a), where MAX[+VOICE] outranks MAX [-VOICE].

(34)

1. a. MAX[+VOICE]: For every [+voice] feature in the input, there is a corresponding [+voice] feature in the output.

2. b. MAX[-VOICE]: For every [ -voice] feature in the input, there is a corresponding [ -voice] feature in the output.

(35) MAX[+VOICE] > > MAX[-VOICE]

The tableau in (36) shows how blocking occurs to satisfy the constraint DEP-PATH-FEATURE in compounds such as abura kasu 'oil dregs'.

As shown, the rendaku-immune noun blocks voicing because the velar consonant has a specified, linked [- voice] feature in the input. Candidate (a) satisfies DEP-PATH-FEATURE since a linked [ -voice] feature occurs in the input. Candidate (b) violates DEP-PATH-FEATURE since no linked [+voice] feature occurs in the input. The [+voice] feature of the junctural morpheme is not linked to any root node in the input. This kind of account of lexical exceptions to rendaku follows the proposal of Inkelas, Orgun, and Zoll (1996), who account for exceptions to regular phonological processes through prespecification of features. I shall delay until section 3.2.2 an illustration of how the constraint MAX[+VOICE] becomes relevant in choosing the optimal candidate.

3.2. Blocking by rendaku resisters

Because the nouns that we are calling rendaku resisters block voicing only in short-short compounds, such blocking must be due to a phonological difference between long and short compounds in addition to any lexical feature that distinguishes rendaku resisters from non-resisters. Accordingly, an analysis of these cases must take account of the differing prosodic structures of long and short compounds.

3.2.1. Prosodic differences between long and short compounds and Positional Markedness

In short-short compounds, neither conjunct exceeds a bimoraic Foot. The primacy of the bimoraic Foot in Japanese phonology is well known (see, for example, Poser 1990). Itô (1991) and Hewitt (1994) show that truncations to a bimoraic Foot occur in several situations in the language. One is in foreign borrowings, where words are often truncated to one bimoraic Foot per morpheme when imported into the language. For example 'word processor' becomes waa-puro, 'remote control' becomes rimo-kon, and 'pocket monster' becomes poke-mon. In addition, personal names are often truncated to fit a bimoraic template before the hypocoristic suffix -tyan, or after the hypocoristic prefix o-. For example, Keiko can become Kei-tyan or O-kei and Hiroko can become Hiro-tyan, or O-hiro. Itô, Kitagawa, and Mester's (1992) examination of "Zyuuzya-go", the Japanese secret language used by jazz musicians, gives further support for the bimoraic Foot as the canonical prosodic unit in Japanese. They argue convincingly that the bimoraic Foot is the template into which an original word is mapped to derive the corresponding word in the secret language.

An important part of the analysis is Kubozono's (1999) proposal that in Japanese, every Foot is entirely contained within the same morpheme. In his analysis of compound accent, two adjacent monomoraic morphemes will not form a single bimoraic Foot. Instead, each will be parsed as a separate degenerate Foot. Following analyses of Japanese prosody in Poser (1984, 1990), Tateishi (1989), Itô (1991), Mester (1990), Itô and Mester (1992), Itô, Kitagawa, and Mester (1992) and Kubozono (1995), let us further adopt the position that a Prosodic Word must maximally consist of two Feet (37a). Where binary Feet cannot be achieved, degenerate, monomoraic Feet are formed, as in (37b–c).

(37) Licit Prosodic Words for short-short compounds:

Short-short compounds will then be dominated by a single Prosodic Word, as in (37). Compounds in which both members are monomoraic will consist of two degenerate Feet dominated by a Prosodic Word.

Long compounds, on the other hand, are prosodically too large to be dominated by a single Prosodic Word, given our hypothesis that a Prosodic Word can contain at most two bimoraic Feet. Long compounds of the form 3μ + 2 μ and 2μ + 3 μ are shown in (38). The crucial assumption is that all prosodic material is dominated by a Prosodic Word. In (38), a single Prosodic Word could not dominate both conjuncts because it would consist of three Feet: two bimoraic Feet plus a degenerate monomoraic Foot.

(38) Licit Prosodic Words for long compounds:

Our hypothesis that long compounds differ prosodically from short compounds is supported by independent evidence from the pitch-accent behaviour of Japanese compounds. Kubozono (1995, 1996) and Alderete (1999) show that the pitch accent pattern of a long compound is completely predictable. On the other hand, the pitch accent pattern of a short-short compound cannot be predicted from the input forms of its constituents. According to Haruo Kubozono (personal communication), a random 70% of 2μ + 2μ compounds are unaccented; most of the remaining 30% have antepenultimate accent, and the rest have unpredictable accent. As for 2μ + 1μ compounds, a random 50% are unaccented; the other 50% are initially accented.

The following example illustrates the unpredictability of pitch accent in short-short compounds. When the accented noun te 'hand' combines in a compound with each of the following final-accented bimoraic nouns, the accent patterns of the resulting compounds will vary in spite of the fact that the inputs have identical accent patterns in all cases (pitch-accent indicated by acute accent):

The predictability of accent in long noun-noun compounds is accounted for by Alderete (1999:85ff), following analyses by Poser (1990) and Kubozono (1995). Predictable default accent in a long compound obeys a prosodic requirement of three or more moras on the noun that receives accent. Default accent does not occur in short-short compounds, which have no member that is three moras or longer. Alderete proposes that default accent in a Japanese long compound occurs on a constituent that forms a prosodic head, which he analyses as a Prosodic Word. In his analysis, a prosodic head must dominate at least a bimoraic Foot plus further prosodic material. An implication of his analysis is that short-short compounds never receive default accent because neither constituent qualifies as a prosodic head, being less than a Prosodic Word. This kind of analysis of Japanese compound accentuation is consistent with the present proposal that long compounds consist of a separate Prosodic Word for each constituent but short-short compounds do not.

Adopting such a proposal enables us to explain why rendaku blocking (other than by rendaku-immune nouns) occurs only in short compounds. The relevant generalization is that rendaku voicing is freely permitted to occur in a syllable that is at the left edge of a Prosodic Word. In short compounds, as shown in (37), the initial obstruent of the second conjunct is not at the left edge of a Prosodic Word; in long compounds, shown in (38), it is. In other words, the [+voice] feature of the junctural morpheme is permitted more freely in what we could call a "prominent" position. Beckman (1998:vii), in her examination of phonological effects that she explains through "Positional Faithfulness", identifies "psycholinguistically prominent positions such as roots, root-initial syllables, stressed syllables, and syllable onsets". Her claim is that these kinds of positions differ from non-prominent positions with respect to faithfulness to an underlying form.

The generalization that rendaku surfaces more readily at the edge of a Prosodic Word could be captured either as Positional Faithfulness (Alderete 1995; Jun 1995; Lombardi 1995a; Steriade 1995; Beckman 1995; Casali 1996; Beckman 1998) or as Positional Markedness (Zoll 1998; Kager 2001; Smith 2002). In the former view, "Head-Dependence" (Alderete 1995) requires that the voicing feature that occurs on an obstruent at the left edge of the second conjunct have some correspondent in the input, when that edge is a prosodically strong position (i.e., the edge of a Prosodic Word in a long compound). In the latter view, the marked [+voice] feature is permitted more freely in the prosodically strong position that occurs at the left edge of the second conjunct of a long compound.

Although either of the two approaches is possible for the case at hand, I will adopt Positional Markedness here for the following reasons. Given our hypothesis that the morpheme boundary in a long compound is in a prosodically stronger position than in a short-short compound, a Positional Faithfulness account would need to view long compounds as being more faithful to the input than short compounds. Yet the fact that blocking in short compounds clusters around particular lexical items suggests that it is some lexically specified feature such as [ -voice] that effects blocking. Under Positional Faithfulness, we ought to expect faithfulness to such a feature to occur more readily in long compounds than in short compounds. What appears to be happening, however, is the opposite, since it is short compounds rather than long ones that are showing lexical idiosyncrasy.

It is also to be noted that in the prosodic structure of Japanese compounds that we have proposed, the left edge of the whole compound is also at the left edge of a Prosodic Word; however, rendaku voicing never occurs in that position. If we consider the junctural morpheme as an affix, we can account for its linear surface position through De Lacy's (1999) Correspondence Theory of Morpheme Order, where the direction of attachment of an affix to a morpheme is determined by an empty position in the input form of a morpheme. Under De Lacy's proposal, N2 of a Japanese compoundwould have an empty position at the left edge of its input form. The junctural morpheme will prefer to surface at that position because of constraints that refer to relations between positions in the input and output, as proposed by McCarthy and Prince (1995).

I will pursue here an account of the effects of prosodic size on rendaku voicing that is based on Positional Markedness, using a type of constraint proposed by Zoll (1998). She proposes a constraint "COINCIDE", which effectively allows marked features to be licensed when they coincide with a strong position. Zoll formulates this kind of constraint through Local Conjunction (see below) of the markedness constraint in question with what she calls a "positional" constraint which requires the coincidence of category x with category y. She argues that local conjunction is necessary for formulating a Positional Markedness constraint.

The principle of "Local Conjunction" of two constraints that form a natural class has been proposed in order to account for a number of cross-linguistic phenomena, including derived environment effects (Smolensky 1995; Lubowicz 1998; Itô and Mester 1998; Alderete 1999). When two constraints are conjoined, they act together as a single constraint, which is violated if and on if both conjuncts are violated in the same domain D. The term "local" refers to the fact that there is some local domainDin which violations of the two sub-constraints are calculated.⁷

In the case we are examining, the markedness constraint IF [-SON] THEN [-VOICE] in (32) is conjoined with the alignment constraint in (40) that in effect identifies the strong position as the left edge of a ProsodicWord.

(40) ALIGN (σ, LEFT, PWD, LEFT): The left edge of every syllable in the output is aligned with the left edge of a Prosodic Word.

The conjoined constraint is violated when both conjuncts are violated in the same syllable. This effectively discourages a [+voice] feature in an obstruent from occurring in a syllable other than the leftmost syllable of a Prosodic Word.

For simplicity, I shall refer to the conjoined constraint IF [-SON] THEN [-VOICE] & ALIGN (σ, LEFT, PWD, LEFT) as COINCIDE([-SON, +VOI],LEFTMOST IN PWD) using Zoll's terminology. This constraint will be violated when a voiced obstruent occurs at a position other than the left edge of a Prosodic Word—most relevantly, here, at the left edge of N2 in a short-short compound but not at the left edge of N2 in a long compound. In section 3.2.2 we examine how this positional markedness constraint derives rendaku blocking in short-short compounds.

3.2.2. Applying Positional Markedness to rendaku blocking

Blocking of rendaku, as we saw in section 2, does not occur in all short compounds. In section 3.1, we distinguished rendaku resisters by positing a lexically prespecified [ -voice] feature in nouns resistant to voicing. Since blocking among resisters occurs only in short-short compounds, it must be both a lexical feature and the violation of positional markedness that blocks voicing. That is, the proposed positional markedness constraint COINCIDE must only be allowed to block voicing when there is a lexical [ -voice] feature in the input. When there is no such feature in the input, no blocking occurs; that is, no [ -voice] feature surfaces in the output. This requirement can be expressed through the following faithfulness constraint:

(41) DEP-FEATURE: For every feature F' in the output there is a corresponding feature F in the input where F' corresponds to F.

DEP-FEATURE must be ranked above the positional markedness constraint COINCIDE, lest voicing be blocked in all short-short compounds. Because the constraint COINCIDE blocks voicing, it must be ranked above the constraint MAX[+VOICE] that was proposed in section 3.1. Including the undominated constraint DEP-PATH-FEATURE that was also proposed in section 3.1, we arrive at the following constraint ranking:

(42) DEP-PATH-FEATURE, DEP-FEATURE > > COINCIDE > > MAX[+VOICE] > > MAX [ -VOICE]

To see how this ranking will determine relevant output forms, consider, for example, the rendaku resister kusa 'grass', which voices in only three out of 19 short-short compounds in the corpus. It has the proposed lexical entry /Kusa/, with the floating feature [ -voice]. Henceforth, capitalized letters "K", "T", "S", and "H" are used to represent input forms of obstruents that have no specified voicing feature. Relevant instances of the features [+voice] and [ -voice] are shown with no association lines when they are floating features and with association lines when they have a path to a root node. In the tableaux that follow, only violations that pertain to the voicing feature on the initial obstruent of N2 are shown.

Both candidates in tableau (43) violate DEP-PATH-FEATURE since there is no path from a voicing feature to the relevant root node in the input. On the other hand, neither candidate violates DEP-FEATURE, which does not refer to paths, since both [+voice] (the junctural morpheme) and [ -voice] (part of the input form of kusa) occur in the input. If the [+voice] feature of the junctural morpheme occurs in the output as in candidate (b), the constraint COINCIDE is violated, since the [+voice] feature is not at the left edge of a Prosodic Word, which here dominates the whole compound, as proposed in (37). Candidate (a), however, in which no voicing occurs, will not violate COINCIDE, and therefore, is optimal. Prosodic Word boundaries are shown henceforth with curly brackets.

If the input has both a floating [+voice] (the junctural morpheme) and a [ -voice] feature (present in some morphemes, as proposed), it does not matter whether more than one instance of each feature occurs in the entire input: an output with no voicing will be optimal for a short compound. For example, a noun with both floating [ -voice] and [+voice] will block voicing as well, since either relevant candidate will satisfy the constraint DEP-FEATURE.

Consider now, in (44), a compound in which N2 is not a rendaku resister, and, as we have proposed here, has no underlying [ -voice] feature. Both candidates violate DEP-PATH-FEATURE, because there is no path from the initial consonant to any voicing feature in the input. Candidate (a), with no voicing on the initial obstruent, will violate DEP-FEATURE, since the [ -voice] feature in that candidate has no correspondent in the input. Candidate (b), which voices the obstruent, satisfies DEP-FEATURE because there is a [+voice] feature (the junctural morpheme) in the input. Because the faithfulness constraint is ranked above the positional markedness constraint, positional markedness cannot block voicing in this case.

Consider now a long compound. Even if N2 were to have the kind of floating [ -voice] feature that we have proposed blocks voicing in short compounds, voicing will not be blocked by the constraint COINCIDE, as it was in (20), since the left edge of N2 is, in our analysis, at the left edge of a Prosodic Word. This is where the proposed constraint MAX[+VOICE] (34) becomes relevant. Candidate (a), in which the [+voice] feature of the junctural morpheme does not surface, will violate this constraint, making the candidate with voicing optimal:

In summary, lexical prespecification of a [ -voice] feature in rendaku-resisting nouns interacts with prosodic length to cause blocking of rendaku voicing specifically in short-short compounds but not in long compounds.

3.3. Exceptions to robust patterns: blocking by non-resisters and voicing by resisters

The previous subsection proposed a set of representations and a constraint ranking that is intended to account for a strong degree of patterning among Yamato nouns: a set of rendaku-resisting nouns robustly blocks voicing in short-short compounds and non-resisting nouns robustly undergo voicing. However, as we observed in section 2.5, there are also exceptions to this general pattern: (a) some nouns that usually voice fail to voice in a few compounds (see, e.g., (19)); and (b), rendaku-resisting nouns do voice in a few compounds (see, e.g., (8)).

A few examples of the first kind of exception are given in (46), with the nouns kumo 'cloud', tama 'ball', and tori 'bird', all of which usually voice in compounds.

Examples of the second kind of exception, voicing by resisters, can be found with the resister te 'hand, helper, means'. In the following four compounds, repeated from (17), the noun te voices:

It is notable that many compounds like those in (46) or (47), which voice in a manner that is contrary to the overall tendencies of resisters and non-resisters, exhibit an idiosyncratic, non-compositional meaning, as we see can above. For these reasons, let us consider the hypothesis that there is a single lexical listing for compounds that depart from the normal patterns of resisters and non-resisters. That is, these compounds are derived from a single "listeme" or lexical entry in the sense of the term proposed by Di Sciullo and Williams (1987).

Because of the OT principle of "Richness of the Base" (Prince and Smolensky 1993), which requires that there be no restriction on possible inputs, whenever an input consists of a single listeme that encompasses several morphemes, there must also be an input that consists of separate listemes for each of the constituent morphemes. For example, if a compoundlike yami-kumo (46), has a single listeme as a input, a second input, that consists of a composition of yami 'darkness' and kumo 'cloud', must also be considered. In order to distinguish between these kinds of possible inputs, Zuraw (2000), in her treatment of systematic exceptions, proposes a constraint USE-LISTED which requires that an output correspond to a one-listeme input form rather than to the separate input forms of the constituent morphemes. For example, in her tableau that determines the surface form of the Tagalog verb bigaj 'distribute' combined with the prefix ma-, she posits a possible input form that is a single listeme: mamigaj. An output candidate derived from this single listeme will satisfy USE-LISTED whereas one that is derived from a concatenation of the prefix with the verb stem will violate the constraint.

There are a number of possible ways of formulating USE-LISTED. One is to consider it a subtype of the *SPEC constraint proposed by Prince and Smolensky (1993), which bans underlying material from the input. Violations of this constraint would be measured gradiently, with one violation mark for each instance of a listeme in the input.⁸

(48) USE-LISTED: The occurrence of a listeme in the input is banned.

The constraint will be violated once by an output candidate that is derived from a single-listeme input and more than once by an output candidate that is derived from more than one listeme. With respect to rendaku voicing, high ranking of this constraint will allow exceptional voicing to occur when a compound is listed as one listeme with a [ -voice] feature that does not occur in the listing of the simplex morpheme N2. This is because this constraint will eliminate candidates that are derived from separate listemes when they compete with candidates that are derived from a single-listeme input. This will be illustrated in tableau (50).

DEP-PATH-FEATURE must be ranked above USE-LISTED, for reasons that will be discussed in section 3.4. However, because we still want single-listeme inputs to permit exceptional blocking by non-resisters,USE-LISTED must be ranked above DEP-FEATURE and COINCIDE, which cause blocking of rendaku. Accordingly, the constraint ranking we shall posit is as follows:

(49) DEP-PATH-FEATURE > > USE-LISTED >> DEP-FEATURE >> COINCIDE >> MAX[+VOICE] >> MAX[-VOICE]

3.3.1. Deriving exceptional blocking in short-short compounds

Given the proposed constraint ranking in (49), consider the compound yami kumo 'recklessly', where the noun kumo 'cloud' is not a rendaku resister. The proposed single-listeme input, shown in (50), contains a floating [ -voice] feature that does not occur in the listing of the simplex noun kumo. The input form of the latter is /Kumo/ with no specification of voicing on the obstruent.

The tableau in (50) illustrates how a single-listeme input can derive exceptional blocking in yami kumo. I follow Zuraw (2000) in distinguishing between candidates in the tableau according to the input form they are derived from.

Candidate (a) differs from candidate (c) in that (a) is derived from a single listeme whereas (c) is derived from three listemes (if we include the junctural morpheme as a listeme). The notation "1L" refers to a candidate that is derived from a single listeme and "C" to a candidate that is derived from the concatenation of separately listed morphemes. In this tableau, all candidates violate DEP-PATH-FEATURE because whatever voicing feature surfaces on the initial obstruent of kumo, it has no linked voicing feature in the input. High ranking of USE-LISTED eliminates candidates (c) and (d), each of whose inputs consists of more than one listeme.

Candidates (a) and (b) are derived from a single listeme that has a [ -voice] feature in the input. Thus the presence of the added [ -voice] feature in the single-listeme input blocks rendaku in the same way as it does in rendaku resisters. The competition among candidates (a) and (b) is analogous to the determination of the winning candidate for a compound with a rendaku resister such as ao-kusa in (43).

3.3.2. The absence of exceptional blocking in long compounds

Recall that exceptional blocking never occurs in long compounds, which always undergo rendaku voicing, the only exceptions being a lexically identified set of uniformly rendaku-immune nouns. In the above tableaux, blocking of rendaku in non-immune nouns occurs because of the positional markedness constraint COINCIDE, which is only violated in short-short compounds. This means that a single-listeme input with an included [ -voice] feature cannot cause exceptional blocking in long compounds, since it does not interact with the constraint COINCIDE, which only pays attention to prosodic size. Consider, for example, a hypothetical single-listeme input with an added [ -voice] feature for the long compound hituzi gumo 'sheep cloud'. As we see in tableau (51), because the compound is long, and each constituent thus constitutes its own ProsodicWord, the positional markedness constraint is satisfied by a candidate with voicing.

A single-listeme input has no effect on blocking here, since violation of the positional markedness constraint is necessary to derive blocking.

3.3.3. Deriving exceptional voicing among resisters

We turn now to the second type of exception to the regular pattern: exceptional voicing in compounds where we expect blocking. An example is hito de 'starfish' (lit. 'person-hand'), which voices despite the fact that the noun te 'hand' robustly resists voicing in short-short compounds (see (16)–(17)).

A possible listing for such a compound is one in which the floating [ -voice] feature associated with the rendaku-resisting noun is absent from the single-listeme input, as shown in tableau (52). With such an input form, all candidates will equally violate DEP-PATH-FEATURE since there are no linked voicing features in the input form of either the simplex morphemes or the single-listeme input. USE-LISTED eliminates candidates derived from two listings in (c) and (d). Of the two remaining candidates, which are derived from the single-listeme input with no [ -voice] feature, candidate (b) with voicing becomes optimal, because it satisfies DEP-FEATURE, unlike the voiceless candidate (a). Here, the single-listeme input can override the effects of the floating [ -voice] feature in the simplex input form of N2, because USE-LISTED forces the choice of the single-listeme input form.

3.4. Constraining the effects of single-listeme inputs

We have proposed that exceptions to robust blocking of voicing patterns of resisters and non-resisters can be explained through a single-listeme input for the whole compound whose voicing features differ from those of the input form of N2 alone. These features differ either by the presence of a floating [ -voice] feature that is not present in the simplex listing of N2, or by the absence of a floating [ -voice] feature that is present in the simplex listing of N2. However, if compound words can have single-listeme inputs whose features can depart from those of the simplex-morpheme listings of the corresponding constituents of the compound, we must explain why we do not find a compound whose special listing departs even more widely than observed from the input forms of the constituent morphemes. A few compounds exhibit morpheme-final vowel alternations. For example, hune 'ship' surfaces as huna in compounds like huna-ita 'ship-plank' (see Vance 1987:149); however, these kinds of alternations are much less radical than what should be hypothetically possible if we allow single-listeme inputs for compounds.

In the OT framework, Correspondence Theory (McCarthy and Prince 1995) was developed to capture the requirement that two morphologically related structures show correspondence to each other. Even if a morphologically complex word is derived through a single-listeme input, we generally find, except, for example, in cases of suppletion, that its output form will show phonological correspondence to the shape of morphologically related simplex forms.

With respect to Japanese compound words, we find that the possibility of obtaining surface forms that depart from robust patterns is highly constrained. For example, long compounds never block voicing unless the second constituent is itself lexically specified as rendaku-immune. If a single-listeme input of a compound word were allowed to derive outputs in an unconstrained fashion, nothing should prevent the possibility of a compound like hotaru gusa 'firefly grass' being given a single-listeme input with a [ -voice] feature linked to the velar consonant in the input, just as we proposed is the case for the simplex listing of a rendaku-immune noun. Recall that we proposed that the simplex noun kusa 'grass', which resists rendaku in short-short compounds, has a floating [ -voice] feature in its input form that does not have a path to a root node. The proposal here is that compounds with rendaku resisters like hotaru gusa never show exceptional blocking in long compounds because of correspondence between the output form of hotaru gusa and the input form of the simplex noun kusa 'grass', which, unlike a rendaku-immune noun, has no linked voicing feature on the initial obstruent. To account for this kind of correspondence, I propose a faithfulness constraint that is based on correspondence of a single morpheme to its simplex input form. This constraint follows the general framework of Correspondence Theory. In this case, we require correspondence between a morphologically complex output form and the input forms of its constituent morphemes:

(53) DEP-PATH-FEATURE (MORPHEME-CORRESPONDENCE): If P' is a path to F', where F is a feature in morphological domain M_i' in the output, and M_i is a morphological domain in the input, where M_i' corresponds to Mi, then there must exist path P to feature F in the (simplex) input form of M_i, such that P' corresponds to P and F' corresponds to F.

This constraint requires correspondence to the input form of a particular morpheme, rather than to any single-listeme input for a morphologically complex word. It replaces the constraint DEP-PATH-FEATURE in our earlier tableaux. In those tableaux, making the constraint DEP-PATH-FEATURE require morpheme correspondence would have had no added effect, since we were not positing any single-listeme input for a whole compound that differed from the input form of a constituent morpheme with respect to paths from features to root nodes.

3.4.1. Single-listeme inputs cannot force blocking in long compounds

Let us now examine how the faithfulness constraint proposed in (53) will prevent exceptional blocking in a long compound like hotaru-gusa, even if it were given a single-listeme input with a linked [-voice] feature. Since, by hypothesis, kusa has no path to a [ -voice] feature in its simplex listing, the output *hotaru kusa, with no voicing, violates faithfulness that is based on correspondence to the input form of the constituent morpheme kusa, even if it were to satisfy faithfulness to some spurious [ -voice] feature in a single-listeme input. This is because the relevant faithfulness constraint DEP-PATH-F(MORPHEME-CORRESPONDENCE) refers to correspondence between the output and input forms of the simplex morpheme, not the form of the single-listeme entry for the whole compound.

The tableau in (54) illustrates how the output form of the compound hotaru gusa is determined when it has a linked [ -voice] feature in a single-listeme input. DEP-PATH-FEATURE (MORPHEME-CORRESPONDENCE) replaces the previous constraint DEP-PATH-FEATURE in the proposed hierarchy.

Because the DEP-PATH-F(MORPHEME-CORRESPONDENCE) constraint is based on correspondence to the input form of the morpheme kusa, and not to the single-listeme input for the whole compound, this constraint will not eliminate candidates with voicing, the way it does for a rendaku-immune noun, as in (36). The rest of the tableau is the same as that of any other long compound, as in (45).

3.4.2. Single-listeme inputs cannot force voicing among rendaku-immune nouns

To further show that a single-listeme input cannot create "unexpected" exceptions to systematic patterns of blocking and voicing, consider the fact that compounds with immune nouns can never be forced to voice by a single-listeme input that has built-in voicing. Suppose, for example, that there were a single-listeme input with an immune noun in which its linked [ -voice] feature is not present or linked to the obstruent, as shown in (55).

Because of the high ranking of DEP-PATH-F(MORPHEME-CORRESPONDENCE), all candidates that voice are out of contention, since they have no path to a [+voice] feature in the input of any simplex morpheme. By contrast, candidates with no voicing satisfy this constraint, owing to the presence of a linked [ -voice] feature in the input form of the simplex noun. Crucially, because the highest ranked faithfulness constraint is based on correspondence to a simplex morpheme, a single-listeme input cannot force this kind of unexpected output.

4. Conclusion

This article has argued that exceptions to rendaku voicing in noun-noun compounds, while not always fully predictable on an individual basis, show strong lexical and phonological patterning. More specifically, rendaku blocking occurs robustly in two situations: (a) among rendaku-immune nouns that never voice in any compound; (b) among rendaku-resistant nouns when these nouns occur in short-short compounds. These tendencies have been represented in a speaker's knowledge in two ways. One is through identification of rendaku-immune nouns and rendaku resisters in a speaker's lexicon. Formally, this kind of lexical marking occurs, in our analysis, through combinatorial specification of features (Archangeli and Pulleyblank 1994). Specifically, the initial voiceless obstruent of a rendaku immune noun is fully specified in the input with a [ -voice] feature that has a path to a root node. The initial voiceless obstruent of a rendaku resister is specified as [ -voice] but this feature does not have a path to a root node in the input. Nouns that are neither resisters nor immune do not have an initial obstruent specified for a voicing feature. The second is through a grammar that calculates possible outputs with respect to Positional Markedness (Zoll 1998). Specifically, outputs are disfavoured when they have a marked voiced obstruent in a prosodically weak position; in this case, away from the edge of a Prosodic Word. Such a grammar accounts for why blocking among rendaku resisters occurs only in short compounds, where the initial obstruent eligible for voicing occurs in a prosodically weak position.

In the light of recent work on systematic irregularity by Zuraw (2000) and Albright and Hayes (2002), the empirical evidence that rendaku blocking follows systematic patterns further illuminates the question of how systematic irregularity in language might be encoded in a speaker's knowledge of language. Further, the question of why short-short compounds can block voicing in situations where long compounds cannot is relevant to the question of howfaithfulness and markedness interact.

Finally, we observed a parallel between the predictability of pitch accent and the predictability of rendaku voicing with respect to the effects of prosodic size on a compound word. When a compound has at least one constituent that exceeds two moras, both its pitch accent and the voicing of the rendaku-targeted obstruent are predictable from the input forms. When neither constituent attains this size, neither the pitch accent nor the voicing behaviour are completely predictable. In both cases, there is a correlation between prosodic size and irregularity. Further investigation of pitch accent in short-short compounds is needed in order to determine if what we are seeing here is an interaction between size and faithfulness or size and markedness.