The Japanese accent system exhibits culminativity but not obligatoriness: at most one prominence is allowed within a word, whether simple or complex. The language allows for contrastive accent specification, both in terms of presence vs. absence and location, yielding minimal pairs such as (1). Japanese shows a so-called ‘N+1’ accent pattern for words with N syllables, where N is an integer greater than or equal to one (McCawley 1968) (2).

While native nouns exhibit unpredictable lexical accent, loanwords generally follow a default accentuation pattern that places the accent on the syllable containing the antepenultimate mora (3-a) (McCawley 1968)1. When a word lacks an antepenultimate mora, the accent falls on the penultimate syllable (3-d). In cases where the antepenultimate mora is prosodically deficient, such as a moraic nasal, the first half of a geminate, or the second part of a long vowel, the accent shifts leftward to the preantepenultimate mora, which is the head of the syllable (3-b,c).

A brief note on surface tonal patterns in Japanese is warranted here. The standard representation of Japanese accent is a HL tonal complex (Kawahara 2015; Ito & Mester 2018). As noted by Kawahara (2015: 449–450), aside from tones assigned by lexical accent, the first two syllables of a Japanese word typically bear a LH tonal sequence, often referred to as initial lowering or initial rise, unless the first syllable is accented, in which case the word exhibits the accentual HL fall instead. When syllables do not receive a tonal specification from either accent or initial rise, their tonal realisation is derived through spreading from the rightmost specified tone. This spreading is often analysed as phonetic interpolation, suggesting that such syllables are phonologically toneless even at the surface level (Pierrehumbert & Beckman 1988). Since this study focuses on suffix-induced accent patterns, it does not attempt to model surface tonal interpolation.

The accentuation of suffixed words in Japanese is largely unpredictable and depends on the accentual type of the suffix. While roots are generally classified as either accented or accentless, suffixes resist this binary distinction and fall into a range of accentual classes. Crucially, suffixes contribute to word accentuation in ways that cannot be reduced to their inherent accent. They have been categorised into dominant, recessive, preaccenting, subtractive (referred to here as usurper), and accent-shifting, as summarised in Table (4). The classification of suffixes is based on previous work, including Poser (1984), Alderete (1999), and Kawahara (2015). It should be noted, however, that the suffixes presented in these sources do not always form a homogeneous morphological class, and their status as true suffixes is sometimes debated. This issue is orthogonal to the central claim of the present paper, as the proposed analysis treats all forms uniformly within word-level phonology, regardless of their precise morphological classification (see Section 4). For the sake of consistency with the existing literature, I retain the traditional terminology here.

I now turn to each suffix class and the accent patterns that emerge from their interaction with accented and accentless roots. We begin with the pure recessive, which comes in two types: accentless and accented. The nominative marker /-ga/ represents the accentless subtype. It is ‘recessive’ in that it cannot modify a preexisting accent and ‘accentless’ in that it does not surface with an accent when combined with an accentless root (5-a). The conditional suffix /-tára/ represents the accented subtype: it realises its accent when combined with an accentless root, while the root retains its accent over that of the suffix (5-b)2.

The next class is the dominant accented suffix class, exemplified by /-ppói/ ‘-ish’. This suffix realises its accent regardless of whether the root is accentless (6-a) or accented (6-b). In cases where the root is already accented, the root accent is deleted to satisfy the culminativity restriction. This behaviour contrasts with that of the recessive accented suffix, where the suffix accent is deleted instead. Another example of a dominant suffix is /-gúrai/ ‘at least’, which deletes the root accent to retain its own (Kawahara 2015: 468).

Preaccenting suffixes comprise the third suffix class and are divided into two subtypes: dominant and recessive, represented by /-ke/ ‘family of’ and /-si/ ‘Mr.’, respectively. When combined with an accentless root, both types induce an accent on the syllable immediately preceding the suffix (7a–c). However, when combined with an accented root, only the dominant subtype induces preaccentuation (7-b), while the recessive subtype has no effect, leaving the root accent unchanged (7-d).

The usurper (i.e. deaccenting) suffixes, represented by /-kko/ ‘native of’ and /-teki/ ‘-like’, exhibit an accent-subtraction effect. These suffixes are underlyingly accentless and do not assign an accent when combined with an accentless root (8-a). However, when attached to an accented root, they delete the existing accent without contributing one of their own (8-b,c). Since the suffix itself lacks an underlying accent, this subtraction effect cannot be attributed to culminativity (as is the case with pure dominant suffixes). Other suffixes exhibiting similar subtractive behaviour in Japanese include /-iro/ ‘colour’, /-tama/ ‘ball’, and /-too/ ‘(political) party’ (Kawahara 2015: 470).

The final suffix class comprises accent-shifting suffixes, divided into two subtypes: the attractive suffix /-mono/ ‘thing’ and the adsorbent suffix /-te/ ‘who Vs’. These suffixes are typically analysed as not bearing an underlying accent but shifting a pre-existing one; see Poser (1984: 68), Kawahara (2015: 469), and Alderete (1999: 229). That is, they do not contribute a new accent when attached to an unaccented root (9-a,c), but shift the root’s underlying accent (9-b,d)3. The difference between the two subtypes lies in the target of the accent shift: the adsorbent attracts the root accent onto itself (9-d), whereas the attractive suffix shifts the root accent to the syllable immediately preceding the suffix—that is, the root-final syllable (9-b).

I begin with the analysis of the recessive suffix effect. The tonal representation of the suffix /tára/ ‘Conditional’ is given below. As shown, tones and moras are annotated with activity levels: the root H₁ has default activity 1, whereas the suffix H_0.5 has a lower activity of 0.5. Only the root H is realised in the output, and the suffix H becomes floating, as indicated by the crossed-out association line8.

The activity difference between the root and the suffix reflects their different accentual behaviours: the root H-tone with default activity overrides the weaker tone of the suffix. This follows from gradient versions of standard constraints on tone-TBU association, originally formulated as well-formedness conditions on association in Autosegmental Phonology (Goldsmith 1976). The conditions include (i) every TBU must bear a tone; (ii) every tone must be associated with a TBU; (iii) associations proceed in a one-to-one, directional manner, and (iv) association lines must not cross. As noted by Yip (2002: 83), these conditions can be reformulated in OT as violable markedness constraints: (i) *Float, (ii) Specify T, (iii) Align L/R, and (iv) *Crossing. In addition, tone is subject to general faithfulness constraints such as Dep T and Max T, which penalise tone insertion and deletion, and *Associate and *Disassociate, which penalise insertion and deletion of association lines. For clarity and consistency with recent work, I adopt simpler labels for the constraints here. For instance, *Float and Specify T, which demand association between H-tones and moras (µs), are termed H ⊳ µ and µ ⊳ H, respectively.

The assumption of GSR, combined with standard input–output correspondence constraints on tone, entails that constraint violations are sensitive to the activity of elements and are incurred in proportion to those activity levels. For example, H ⊳ µ is violated proportionally to the activity of the H-tone, as shown in (19-b). This sensitivity to gradient activity is crucial for modelling the competition between root and suffix tones in Japanese. In combinations of a toned root and a toned recessive suffix, only one H-tone can surface, as Japanese exhibits culminativity, i.e. each prosodic word may bear at most one tone. This is enforced by the high-weighted Culm(inativity) H constraint (19-a). As a result, the two H-tones, from the root (activity 1) and the suffix (activity 0.5), compete for realisation. The outcome of this competition is determined by the constraint H ⊳ µ (19-b), which favours association of the more active tone. Thus, the root H wins over the weaker suffix H. Other potential repair strategies for satisfying Culm are ruled out by other constraints. Dissociation of a tone-bearing mora is prohibited by V(owel) ⊳ µ (19-c), which bans dissociation of underlyingly associated µs from the vowel. In addition, the highly weighted Max H constraint (19-d) requires that all H-tones in the input have output correspondents, banning deletion of H-tones.

Tableau (20) illustrates how the constraints discussed above interact to predict the tonal behaviour of a recessive suffix when combined with a toned root. The data to be analysed is tábetara ‘if he eats’, where the recessive suffix loses its accent to that of the root. The underlying tone of the verb root is assumed here to be penultimate, which surfaces faithfully, while the recessive suffix loses its initial tone9. Candidate (20-a) shows this pattern and incurs only a 0.5 violation of H ⊳ µ, compared to a full violation of 1 in candidate (20-b), making it the optimal output. Candidate (20-c) faithfully realises both H-tones and thus satisfies H ⊳ µ, but fatally violates the high-weighted Culm H. Candidate (20-d) avoids this violation by deleting the H-tone, but is ruled out by Max H, which penalises deletion of H-tones. Candidate (20-e), which dissociates the tone-bearing µ, is excluded by V ⊳ µ. The optimal strategy is therefore dissociation of one H-tone, namely the partially active suffix H. Note that certain suboptimal candidates are omitted from the tableau for brevity. For example, any candidate that changes the activity level of an element between input and output is excluded by the faithfulness constraints Dep Activity and Max Activity, which ban activity insertion or deletion.

In total, the accentual recessiveness of the suffix is captured by the weak H-tone in its underlying representation, which surfaces only in the absence of a preexisting tone, but loses to the root H-tone with activity 1. This is ensured by H ⊳ µ, which favours association of the more active tone. The tone system of Japanese provides an interesting counterpoint to recession in the form of suffix dominance effects, which are addressed in the following subsections.

Most accentual suffix types in Japanese are dominant in that they induce root tone deletion to satisfy their accentual demands. I start with the simplest dominance effect in Japanese, induced by the pure dominant suffix /ppói/ ‘-ish’. This suffix always realises its tone even at the cost of root tone deletion. As shown in (21), the dominant suffix is represented with a H-tone with default activity 1, which is always realised, whether combined with a toneless or a toned root.

The dominant suffix is tonally stronger than the recessive (discussed above); it however, shares the same H activity as the root, as shown in (21). With a toneless root, the only H-tone in the word, that of the suffix, surfaces faithfully to satisfy H ⊳ µ as well as other faithfulness requirements such as (22-a). Once combined with a toned root, two underlying H-tones are available with the same activity, equally satisfying H ⊳ µ. However, only one tone is permitted due to the culminativity restriction. The requirement for rightmost H realisation becomes decisive and selects the suffix tone; this is captured by the Alignment constraint (22-b). AlignR H is ‘counting cumulativity’ (Jäger & Rosenbach 2006); it counts each intervening µ between the toned µ and the right edge of the word. Note that only ‘properly integrated’ µs (i.e. associated with σ) are taken into account, whereas the constraint could alternatively be formulated over all intervening µs, including the floating.

The constraint interactions that derive pure dominance are illustrated in tableau (23). Candidates (23-a,b) are equally harmonic with respect to H ⊳ µ; however, (23-a), which maintains the suffix H, minimally violates AlignR H and therefore outweighs (23-b). (23-c) perfectly satisfies this constraint, but is nevertheless excluded by the higher-weighted Dep |; this shows that a minimal violation of AlignR H is tolerated to satisfy Dep |. The faithful realisation of both H-tones in (23-d) is also excluded by the high-weighted Culm H.

Preaccenting suffixes do not surface with a tone themselves but assign a H-tone to the final syllable of the root. This effect comes in two types: dominant and recessive. Dominant suffixes always impose preaccentuation, whereas recessive suffixes do so only in the absence of a preexisting tone. The suffixes /-ke/ ‘family of’ and /-si/ ‘Mr.’ exemplify the dominant and recessive subtypes, respectively. The analysis proposed here departs from earlier accounts in that it does not attribute the dominant/recessive distinction to different phonologies or diacritic markings (see Section 5.1 for discussion of preaccentuation in the TAF account of Alderete 1999). Instead, this distinction follows from the principles underlying gradient autosegmental representations. Preaccentuation is analysed here as triggered by a floating H-tone in the suffix’s representation. The H-tone is not underlyingly associated to any mora, but may dock onto a root mora. Dominant suffixes have stronger tone assignment requirements than recessive ones. This is represented in (24) by the difference in activity levels of the suffixes’ H-tones: the dominant suffix has a strong H with activity 2 (24-b) while the recessive suffix has a weak H with activity 0.5 (24-a).

The association of the floating H to a mora is governed by the derived environment restriction proposed by Van Oostendorp (2007), formalised as the constraint Alternation (25). This constraint penalises epenthetic associations between tautomorphemic elements, thereby triggering the association of a floating tone to a mora belonging to a morpheme other than the one that sponsors the tone10. This constraint has been employed in analyses of various tonal processes involving floating tone associations across morpheme or word boundaries (see, e.g. Trommer 2022).

Tableau (26) illustrates the role of Alternation in deriving dominant preaccentuation. Candidate (26-a) emerges as optimal: H ⊳ µ, AlignR H, and Alternation jointly favour association of the strong H-tone to the rightmost heteromorphemic µ. In candidate (26-b), the suffix docks its floating H-tone to its own µ, violating Alternation, which penalises such tautomorphemic associations. The floating H-tone must therefore associate with the root instead, where Align-R H determines its docking site: it attaches to the rightmost root µ (26-a), since additional violations of this constraint are not tolerated (26-c). The gradiently evaluated constraint H ⊳ µ induces the dominance effect: associating the suffix’s strong floating H (activity 2) incurs less cost than faithfully realising the root H (activity 1) (26-d). This asymmetry is reflected in the weighting relation H ⊳ µ > Dep |11. Finally, faithful realisation of both the root and suffix H would best satisfy H ⊳ µ, but this candidate is excluded by the high-weighted Culm H.

The situation differs in the case of the recessive preaccenting suffix. This class does not exert a strong enough demand for H-tone assignment to override the root’s H-tone. This is captured by the suffix’s weak H with an activity level of only 0.5 and by the constraint H ⊳ µ, which favours the faithful realisation of the root H with default activity. This is illustrated in (27-a), where the root H is merely penalised by Align-R H, a constraint whose weight is too low to be decisive in this context. In contrast, candidate (27-b), which dissociates the root H, not only incurs a greater violation of H ⊳ µ (by 0.5) but also violates Dep | due to the insertion of a new association line. It is thus apparent that only in the absence of a root tone does the suffix’s H-tone associate optimally. Since this outcome is analytically straightforward and less informative than other, more complex patterns, it is not discussed in detail here (see tableau 61 in Appendix).

As shown above, the preaccentuation effect arises from the integration of the floating H-tone, whose association is restricted to a derived environment. The locality of preaccentuation, i.e. the tone surfacing adjacent to the sponsoring suffix, results from the rightmost alignment requirement, rather than from any intrinsic structural adjacency; cf. the morpheme-contiguity assumption proposed in alternative analyses of preaccentuation (Revithiadou 2007; Kurisu 2001). See Section 5.4 for further discussion of this assumption.

Three accentual suffix classes in Japanese, the adsorbent, the usurper, and the attractive, exhibit non-replacive dominance, analysed here as tone subtraction. All three classes eliminate the root’s H-tone when one is present, yet show no tonal effect when attached to a toneless root. This shared subtractive behaviour, which unifies these suffix classes, poses analytical challenges. I reserve discussion of the attractive class for the next subsection, as it involves a combination of subtraction and preaccentuation, and thus merits separate treatment. This section focuses on two suffix types: (i) the adsorbent suffix, represented by /-te/ ‘who Vs’, which removes the root H-tone and surfaces with that tone itself, but remains toneless when attached to a toneless root; and (ii) the usurper suffix, represented by /-teki/ ‘-like’, which deletes the root H-tone and renders the entire word toneless, while having no effect on toneless roots. In both cases, the culminativity restriction is orthogonal to the deletion effect, since the suffixes themselves are underlyingly toneless. Dominance is therefore not due to the insertion of a competing H-tone. Root tone deletion cannot be attributed to a preassociated dominant H-tone (as in the pure dominant class; Section 4.2) or to an overriding strong floating H-tone (as in the preaccenting class; Section 4.3). Below, I show that the subtractive patterns follow from a representational analysis that uses the same general mechanism responsible for replacive dominance discussed earlier, namely a preference for association with more active elements, whether tones or moras.

Gradient autosegmental representations allow phonological elements to differ in activity not only on the tonal tier, but also on the moraic tier. Under this assumption, tone subtraction follows from a suffix’s strong requirement to bear a H-tone. In the case of the adsorbent suffix /-te/ ‘who Vs’ and the usurper suffix /-teki/ ‘-like’, both are analysed as underlyingly toneless, but as strongly requiring a tone. This requirement is encoded in their representation as a highly active mora (µ₂) that is not linked to a H-tone and demands one. This shared representational ingredient unifies the two suffix classes and distinguishes them from other suffix types. The difference between the two suffix classes lies in the association status of this strong mora: in the adsorbent class, it is underlyingly associated (28-a), whereas in the usurper class it is floating (28-b).

Capturing the tone attraction effect of the strong µ requires a constraint sensitive to moraic activity, namely the gradiently violated constraint µ ⊳ H in (29) (the reverse counterpart of H ⊳ µ). It requires every mora to be associated with a H-tone and evaluates violations proportionally to moraic activity, thereby favouring association of more active moras to tone.

Tableau (30) illustrates the role of µ ⊳ H in enforcing root H deletion with the adsorbent suffix. In the optimal candidate (30-a), the root’s underlying H-tone dissociates and reassociates with the suffix’s strong mora. This reassociation incurs a Dep | violation, but this is tolerated since it avoids a more costly violation of the gradient constraint µ ⊳ H, as shown in (30-b). This trade-off reflects the weighting relation µ ⊳ H >Dep |. In contrast, candidate (30-c) incurs a fatal violation of Culm by maintaining the root H and inserting an additional H-tone (highlighted in the tableau) on the strong suffix mora in order to satisfy µ ⊳ H. This strategy is also ruled out by a high-weighted Dep H (not shown in the tableau), which penalises tone insertion. This constraint is also operative for the adsorbent suffix’s interaction with toneless roots, as in /kiki-te/ → kikite ‘narrator’ (Poser 1984). The relevant representation (28-a) and the constraints in (30) derive this outcome straightforwardly: the high-weighted Dep H ensures that no H-tone is inserted, even on strong moras. However, when a H-tone is available in the input, it is reassigned to the suffix’s strong mora, since Dep | is not strong enough. The crucial asymmetry between toned and toneless roots is therefore captured by the weighting Dep H>Dep |, which explains both the suffix’s dominance effect in the presence of a root tone and its lack of effect when attached to a toneless root.

The usurper suffix, represented by /-teki/ ‘-like’, is the second subtractive suffix addressed in this subsection. Like the adsorbent suffix, it has no effect on a toneless root, yet deletes the root’s H-tone. However, it does not surface with a H-tone itself. The same set of constraints used above derives this subtractive effect, although the surface result differs. Here, the root H is usurped by the suffix but not realised overtly (cf. the adsorbent suffix)12. I analyse this usurpation effect as arising from the underlying representation of the suffix, specifically from a floating strong mora that lacks a H-tone but strongly demands one. As shown in Tableau (31), the optimal candidate (31-a) associates its floating strong µ with the only available H-tone, namely the root H, thus optimally satisfying µ ⊳ H. This is the same mechanism seen in the adsorbent case, cf. candidate (30-a). However, in the usurper suffix, the strong µ is floating and remains so due to the high-weight constraint Dep | µ-S (see (31-d)). This constraint bans the insertion of a new association between a mora and a segment, ensuring that the floating µ does not trigger vowel lengthening or shortening13. In candidate (31-b), the root H associates with the rightmost integrated µ, satisfying AlignR H, but this candidate is suboptimal because the high activity of the floating strong µ and the weight of µ ⊳ H jointly favour association with the suffix µ. The faithful realisation of the root H, as seen in (31-c), also fails as it results in an additional violation of µ ⊳ H by the floating strong µ. The output is a toneless word, as the H-tone cannot be realised on any of the integrated µs, which are less active than the suffix floating µ. The H-tone is instead associated with the strong floating µ in (31-a), although this association is not phonetically realised due to its non-integration into the prosodic structure. Finally, note that the non-association of the floating µ does not trigger a violation of V ⊳ µ. This constraint penalises newly floating moras, that is, moras that have been dissociated in the output; the strong µ in this case is underlyingly floating.

The final interaction I briefly address is the suffix’s lack of effect on a toneless root, as in /bungaku-teki/ → bungakuteki ‘literature-like’. This case is not illustrated with a tableau, as it follows straightforwardly: due to the high weight of Dep H, all µs, including the strong floating µ, remain toneless in the absence of a preexisting H-tone. The apparent contrast between dominance and neutrality is resolved by the weighting of the two Dep constraints. Specifically, the suffix is strong enough to override Dep | and usurp an existing H-tone under the pressure of µ ⊳ H, but not strong enough to override Dep H and trigger tone insertion.

The final suffix class is the attractive class, exemplified by /-mono/ ‘-like’. Attraction is a codependent effect: the surface pattern depends crucially on the interaction between the suffix and the root. The suffix has no effect when attached to a toneless root; however, when a root H-tone is present, it induces tone shift, analysed here as the combination of preaccentuation and usurpation. The pattern follows directly from combining the representational ingredients proposed for these two phenomena: a floating strong µ, responsible for usurpation, and a floating H, responsible for preaccentuation, as illustrated in (32). When a toned root is present, the floating H and the floating strong µ associate heteromorphemically. The strong floating µ attracts the root H, resulting in root tone deletion, while the floating H docks onto the root-final mora, yielding preaccentuation. The overall effect is apparent attraction of tone towards the suffix (32-a). When the root is toneless, the floating elements associate tautomorphemically. They crucially do not integrate into the prosodic structure, yielding no surface H-tone and no accentual effect (32-b).

Tableau (33) illustrates the interaction of constraints in deriving the tone shift in a toned root followed by an attractive suffix. The crucial constraint set H ⊳ µ and µ ⊳ H sets two core requirements in this context: (i) both H-tones (whether underlyingly associated or floating) must be hosted by a mora, and (ii) the stronger µ must be associated with a H-tone. Both of these conditions are satisfied by the close contenders (33-a) and (33-b): in each, the strong µ bears a H-tone, and the floating H-tone is associated with a mora. The decision between these candidates is determined by the constraint Alternation. Candidate (33-b) performs better on AlignR H, as it realises the root H on the rightmost available mora. However, to satisfy µ ⊳ H, it must associate the remaining H-tone to the tautomorphemic strong µ, thereby violating Alternation, which is weighted higher than AlignR H. This renders (33-a) optimal. Other candidates that might satisfy both µ ⊳ H and H ⊳ µ are not shown, as they incur fatal violations of Alternation. For the remaining candidates (33-c) and (33-d), the exclusion is straightforward: both fail to associate a H-tone to the strong µ, resulting in severe violations of µ ⊳ H. Realisation of both H-tones is independently excluded by the high-weighted constraint Culm H, omitted from the tableau for space reasons. Regarding the crossing association lines in the optimal candidate (33-a), note that the root H associates to the suffix strong µ, which remains floating and does not integrate into the prosodic structure. This suggests that the ban on crossing lines, enforced by a high-weighted constraint, is violated only when the crossing involves integrated elements.

When the attractive suffix is concatenated with a toneless root, the output remains toneless; that is, the suffix has no surface effect. The question is how the grammar resolves the association of the floating H-tone and the strong µ, which ideally should satisfy both H ⊳ µ and µ ⊳ H. This interaction is illustrated in Tableau (34). In candidate (34-d), both elements remain unassociated, incurring fatal violations of the two high-weighted gradient constraints, despite perfect satisfaction of the lower-weighted constraints. Given the high activity of the suffixal µ, associating the floating H-tone with any µ of default activity, as in (34-b) and (34-c), results in a more serious violation of µ ⊳ H. Consequently, in the absence of a root H, tautomorphemic association between the suffix’s strong µ and its floating H yields the most harmonic outcome, as shown in (34-a). Although this candidate violates Alternation, the violation is tolerated because it minimises the overall harmony score by satisfying the more highly weighted constraint µ ⊳ H.

Alderete (1999) develops Transderivational Antifaithfulness Theory (=TAF) by building on Benua (1997) to account for non-concatenative morphology and affix-controlled accentual phenomena, including dominance effects in Japanese. This section shows that the TAF account runs in a ranking paradox in addition to theoretical complications.

A key assumption of TAF is root-controlled accentuation, whereby the root’s accentual demand necessarily overrides that of the suffix. To account for cases of affix dominance, the theory proposes a new constraint family, namely antifaithfulness constraints. These constraints include a negative version of faithfulness constraints each of which demands unfaithfulness with respect to a particular phonological feature in a derivative with respect to its base. For example, ¬OO-Max-Prom demands deletion of accent, ¬OO-Dep-Prom mandates insertion of accent not present in the base, and ¬OO-Flop-Prom requires shifting of base accent. TAF constraints hence demand a change in the base of affixation. Crucially, however, the ultimate realisation of this change is governed by the language’s general grammar. Within this account, each dominant affix subcategorises for an antifaithfulness constraint. Thse assumptions are applied to Japanese suffix classes and illustrated below. The pure dominant suffix /ppói/ ‘-ish’ is assumed to subcategorise for ¬OO_Dom-Max(Accent) that outranks its faithfulness counterpart Max(Accent). The AF constraint is also ranked higher than the antifaithfulness constraint indexed to the recessive suffix /tára/ ‘Cond. This is shown in the schematic ranking (35).

(36) illustrates the ranking of TAF and faithfulness constraints in deriving the different behaviours of the pure dominant and the recessive suffix. The dominant suffix subcategorises for ¬OO_Dom-Max(Ac) that outranks OO-Max(Ac) and induces root accent deletion (36-ib). The opposite ranking with ¬OO_Rec-Max(Ac) that is activated by the recessive suffix results in root dominance (36-iib). As shown by the shadings, the TAF constraints are only sensitive to the affixes that trigger them.

The next suffix class to be discussed is the usurper /-kko/ ‘native of’. Recall from Section 4.4, that this suffix is accentless but deletes the root accent, hence the word is rendered accentless. The suffix has no effect on an accentless root. In the tableau below, (37-i) shows how ¬OO_Dom-Max(Ac) subcategorised by the suffix captures its subtraction effect. The TAF constraint is responsible for the root’s accent deletion and excluding the faithful candidate (37-ia). A high-ranked general faithfulness constraint, namely IO-Dep(Ac), is required to exclude accent insertion on the root or the suffix (37-ib), ensuring the word’s unaccentedness (37-ic). Alderete (1999: 223–24) asserts that a new TAF constraint is required to derive the peculiar effect of the other subtractive suffix, namely the adsorbent /-te/ ‘who’. Recall that this suffix has no effect on an accentless root but induces root accent deletion. However, unlike the usurper, it bears the surface accent itself. This difference requires the root faithfulness constraint IO-Max(Ac)_Root that is ranked high in the grammar of the adsorbent but is crucially ranked below the TAF constraint of the usurper (¬OO_Us-Max(Ac)). This is shown in (37-ii) where the root faithfulness constraint is lower than the usurper’s TAF constraint but not strictly ranked with respect to ¬OO_Ads-Max(Ac). (37-iia) shows how the TAF constraint enforces deletion of the root accent. The optimal candidate (37-iic) shows that IO-Max(Ac)_Root is crucial to ensure the surfacing of the underlying accent on the suffix by excluding the complete deletion of accent (37-iib).

The TAF treatment of preaccentuation requires more complex mechanisms, namely the local conjunction of TAF constraints with anchoring constraints. (38) is proposed by Alderete (1999: 201) that is a conjunction of ¬OO-Dep(Accent) with the already conjoined constraint Anchor-R/L. The constraint prohibits the preaccenting suffix from de-aligning the stem through suffixation and at the same time triggers an epenthetic accent onto the stem at the de-aligned edge.

Tableau (39) illustrates the effect of this rather complex constraint in deriving dominant and recessive preaccentuation. The dominant suffix /-ke/ ‘family of’ subcategorises for ¬OO_DomPre-Dep(Ac)_Edge-σ that outranks the faithfulness constraints and induces accent assignment to an adjacent position to the preaccenting suffix; this is shown in (39-ib) that is chosen over (39-ia) that maintains the root accent. The opposite ranking of IO-Max(Ac)_root and ¬OO_RecPre-Dep(Ac)_Edge-σ activated by recessive preaccenting suffix (/-si/ ‘Mr.’) results in preservation of the root accent (39-iib). The suffix only induces preaccentuation on an accentless root which is captured by the TAF constraint outranking IO-Dep(Ac)_Root (39-iid).

A brief look at the summary rankings proposed by Alderete (1999) reveals a ranking paradox. The crucial constraints involved are the faithfulness constraints IO-Dep(Ac) and IO-Max(Ac) (highlighted in red). Although two versions of each constraint are posited, one general and one local root-restricted, they stand in a subset relation, giving rise to a ranking paradox. For instance, in (40-b), the general IO-Dep(Ac) must be dominated by IO-Max(Ac)_Root to permit recessive preaccenting suffixes to assign accent to an accentless root. However, in (40-c), the reverse ranking is required: IO-Dep(Ac) must dominate IO-Max(Ac)_Root to derive accent usurpation. One possible workaround would be to allow IO-faithfulness constraints to be lexically subcategorised by suffixes. However, this move is not entertained in Alderete (1999), which explicitly restricts lexical subcategorisation to TAF constraints.

Revithiadou (2023) offers a GHG account of dominance effects centred on so-called dependent accents: accents that surface only if at least one underlying accent is present elsewhere in the word. In Japanese, attractive and adsorbent suffixes are considered dependent, as they show co-dependency with the roots in determining the accent pattern of the word. Revithiadou’s account builds on two core ingredients: (i) autosegmental representations in which accent is modelled as a prosodic root node π (following Spahr 2016), and (ii) constraint indexation, formalised as a scaling factor s introduced by affixation (Gouskova & Linzen 2015), which increases the penalty of indexed constraints. Dominance is modelled as a violation of the scaled constraint, multiplied by both the constraint’s weight and the affix-induced scaling factor, i.e. –n × w × s, (see Mcpherson & Hayes 2016). Revithiadou (2023) illustrates the account by a case study of Lithuanian accent, where dominant suffixes are associated with s = 2, so deletion of their accent doubles the penalty of Max[π], which penalises deletion of a prosodic root node π in proportion to its activity. Non-dominant suffixes have s = 1 and exert no scaling effect on constraint evaluation. To account for morpheme codependency in accentuation, each morpheme bears a latent weak accent with activity 0.5. On its own, such an accent remains unrealised, due to Dep[π] that penalises activity insertion and crucially outweighs Max[π]. Weak accents can only surface when their activities merge into a stronger π-node, following the mechanism proposed in Kushnir (2019).

Although Revithiadou (2023) focuses on Lithuanian accentuation, the account is claimed to extend to Japanese, for which a brief sketch is provided here. As mentioned above, central to the proposal is a scalar constraint Max[π], which penalises deletion of the prosodic root node (π) proportionally to its activity and a scaling factor s introduced by dominant suffixes14. Codependency effects observed in attractive and adsorbent suffixes are accounted for by the weighting relation Dep[π] >Max[π], which blocks the realisation of weak accents unless they are coalesced. To see the predictions of this account for Japanese, I construct an analysis based on the core assumptions outlined in Revithiadou (2023)15. Since output gradience is not permitted in her account, it is avoided here. Scaling is likewise not applied, to allow a fair comparison with the present account|particularly as the aim is to evaluate whether an input-only gradient approach can capture Japanese suffix accent patterns using only general constraints.

The gradient representations in (41) are proposed for suffix classes in which the accent (π) exhibits gradient activity at or below 1, following the input-only gradient approach of Revithiadou (2023). Accents may be either associated or floating. All roots are assumed to bear a weakly active associated accent with activity 0.5, enabling the accent coalescence mechanism to derive the codependency effects.

I adopt the constraints in (42), following Revithiadou (2023). Dep[π] and Max[π] are crucially sensitive to the activity level of accents, and RightMost[π] is gradient, assigning penalties based on the number of moras between the π-node and the right edge of the word.

Tableau (43) illustrates how the constraints interact to capture the accentual behaviour of the adsorbent suffix /-te/ ‘agentive’, which surfaces with an accent only when the root is underlyingly accented. Both the root and suffix bear weak accents (π_0.5), which coalesce into a single accent with activity 1 under the weighting Dep[π]>Max[π] (43-a). Compare (43-c), which inserts activity on the root accent and deletes the suffix accent. Candidate (43-b), which realises the merged accent on the leftmost mora, is ruled out by Rm[π]. Finally, the unaccented candidate (43-d) is excluded by Max[π], which is violated by the combined activity of root and suffix (0.5 + 0.5 = 1). The same mechanism accounts for attractive suffixes, represented with a floating weak accent (π_0.5). As shown in (44-a), merging the accents and realising them on the rightmost root syllable is preferred over alternatives such as leftmost accentuation (44-b), ruled out by the gradient Rm[π], or accent deletion (44-c,d). A crucial candidate not included in the tableau is one where the merged accent surfaces on the suffix; this is ruled out by the constraint Alternation (introduced in (25), Section 4.3), which bans epenthetic association of a floating accent with its sponsor. When the adsorbent or attractive suffix combines with an accentless root, the optimal output is an unaccented form. The suffix’s weak accent is not realised on its own as it would require high activity insertion which is ruled out by Dep[π] (the tableaux showing this evaluation is given in Appendix B to save space).

The usurper suffix bears a very weak floating accent (π_0.1) whose deletion is preferred under the constraint weightings. As shown below, the optimal candidate is always accentless with this suffix, whether the root is accentless (45-a) or accented (46-a). This follows from the high weighting of Dep[π], which disfavors activity insertion (45-b,c). With accented roots, merging the suffix accent with the root’s would require an activity insertion of 0.4, which is sufficient to rule out candidates (46-b,c).

The dominant suffix bears an accent with full activity (1), which always surfaces faithfully on the suffix; the root accent which is weaker, is deleted, as shown in (47-a). However, this mechanism mispredicts the behaviour of recessive suffixes. For example, with a recessive suffix bearing slightly weaker activity (e.g. π_0.9), the optimal candidate would incorrectly show accent on the suffix, as in (48-b). Slightly reducing the suffix’s activity still results in incorrect rightmost accentuation, while lowering it too much yields an accentless form, i.e. a usurpation effect, see (46). This shows a conflict in the system: the mechanisms underlying dominance and recession require opposite constraint weightings and therefore cannot be straightforwardly unified. The system also fails to capture the behaviour of accentless recessive suffixes (e.g. /-ga/ Nom): here, a root accent (π_0.5) remains unrealised in the absence of a coalescing accent, again resulting in a usurpation effect. Modelling the recessive suffix with a floating accent does not resolve the issue either, as it incorrectly yields preaccentuation, discussed below.

A floating accent (π_0.9) in the representation of the dominant preaccenting suffix captures the preaccentuation pattern observed with both accentless and accented roots. As shown in (49-a), the suffix accent surfaces on the root-final mora due to the interaction of Rm[π] and Alternation (the latter not shown in the tableaux). Its activity is high enough to block deletion, ruling out (49-b). When combined with an accented root, the suffix accent merges with the root’s accent and surfaces on the root-final mora, as shown in (50).

It remains unclear how the account could capture recessive preaccentuation. For example, a floating weak accent (π_0.6) yields either usurpation or attraction (see tableaux (45) and (44)), while a weak associated accent results in adsorption (see (43) and (66) in the Appendix). Adjusting activity levels does not resolve the issue. Although the activity scale is formally gradient, distinct accentual patterns do not emerge with each incremental change. Instead, different outcomes arise only at specific threshold values. When implemented in the grammar, the coalescence and deletion mechanisms can enforce only one of the conflicting patterns (e.g. dominance vs. recession, or dominant vs. recessive preaccentuation) at a time. This limitation suggests that the representational assumptions of Revithiadou’s proposal, particularly the reliance on gradient input activity, are insufficient to capture the full range of attested suffix behaviours in Japanese. While the account offers an elegant treatment of codependency effects, it does not generalise to suffix types that do not involve morpheme codependency. A further issue concerns the treatment of root accent: roots must necessarily be weak in order to participate in coalescence-based realisation, on the assumption that weak accents surface only when merged to form a strong one (activity = 1). It remains unclear how roots with full activity would behave, as such cases fall outside the scope of the coalescence mechanism. Revithiadou (2023) addresses some of these issues by introducing constraint indexation. The alternative pursued here is to allow output gradience, while preserving a simple grammar built from independently motivated, general constraints.

This section discusses two alternative approaches to Japanese suffix accentuation: Kurisu (2001) and Trommer (2019a). Both expand the standard OT machinery and rely on comparatively complex representational or constraint-based mechanisms. I begin with Trommer’s (2019a) representational model within Coloured Containment theory (Trommer 2011), which introduces the proposal of two-dimensional concatenation. Under this proposal, concatenation of floating affixes may apply not only horizontally but also vertically. Floating material may attach to a pivot on its own tier (Zimmermann 2014; Trommer & Zimmermann 2014), but it may also associate upward or downward to a pivot on an adjacent tier. Vertical association requires an additional complex autosegmental primitive: a floating element bearing a coloured association that enters the phonology as a pre-specified structural unit, represented here as H—.

Table (51) illustrates the underlying representations that Trommer (2019a) proposes for seven suffix classes in Japanese. Note that the suffix names are adapted for terminological consistency. The adsorbent suffix is neither included in the table nor analysed in detail. It therefore remains unclear how the account would derive its attraction effect. In the representations in (51), both H and L tones are lexically specified. An accented position is represented by a H-tone followed by a L-tone on the subsequent syllable. The analysis crucially relies on different association statuses of these tones to derive the various suffix effects. An underlyingly associated H not followed by L encodes recession (51-ii), whereas the presence of an underlyingly associated L in (51-iii, vi, viii) is intended to induce different dominance effects. A floating H in (51-v) is intended to derive subtraction, while the complex floating H— is required to derive preaccentuation.

The constraints implementing these representations come in two versions: generalised and phonetic, the latter underlined in the tableaux.The general constraints refer to all output material, whereas the phonetic ones exclusively reference phonetically realised (visible) material. The crucial constraints include 1H (corresponding to Culm in my analysis), T→µ, Max T₁, which penalises phonetically invisible PW-initial tones, H⇔L, which requires a phonetic H to precede a phonetic L, and Fth L, which enforces faithful realisation of an underlying L. Fth L plays a central role in deriving the dominance effect of /-ppói/ ‘-ish’ (52) and the strong preaccenting suffix /-ke/ ‘family of’ (53). As shown in the tableaux below, Fth L excludes candidates (52-b) and (53-b), which preserve the root tone at the expense of the suffix tone.

Trommer (2019a) further introduces a specific constraint, , which penalises a phonetic H–L sequence whose H bears a nonmorphological (i.e. epenthetic) association. It thus functions as a hybrid constraint: it simultaneously references phonetic tones and morphological association lines. This constraint is crucial for deriving the interaction of a recessive suffix with a toneless root. It ensures faithful realisation of the suffixal H-tone by excluding candidate (54-b). The formulation of must be sufficiently complex to avoid incorrectly excluding optimal candidates such as (55-a), where an attractive suffix shifts the root tone. This candidate satisfies only because the root H retains a morphologically invisible association line.

While Trommer’s account offers an elegant extension of Coloured Containment and provides a unified representational treatment of multiple suffix types, excluding the adsorbent class,, it does so at the cost of introducing (i) complex floating primitives, (ii) vertically specified concatenation, and (iii) hybrid constraints that reference both phonetic and morphological structures. These mechanisms considerably expand the theoretical apparatus of OT and raise questions about restrictiveness and learnability. In contrast, the present account derives the same effects using independently motivated constraints without recourse to unmotivated concatenation operations.

The other account of suffix accentuation in Japanese is Kurisu’s (2001) Realise Morpheme Theory (RMT), which is based on the assumption that the presence of a morpheme in the underlying representation must be expressed through some phonological change in the output. A central constraint in this model is Realise Morpheme (RM), defined below.

To capture dominance effects in Japanese, the RM account is supplemented with Sympathy Theory (McCarthy 1999), an extension of OT that evaluates inter-candidate correspondence. For example, the usurpation effect of /kko/ ‘native of’ is by designating a sympathy candidate via a selector constraint (marked ❀), rendering the suffix opaque. An accentless root, e.g. koobe ‘Kobe’ is selected as a sympathy candidate. The optimal candidate koobe-kko ‘native of Kobe’ must imitate the absence of accent in the sympathy candidate due to the selector constraint Dep-❀O-Accent. Accent deletion thus emerges as the strategy for satisfying RM. Although this mechanism captures the usurpation effect, it does not derive the adsorption effect (as also noted for Trommer 2019a). Moreover, non-local affix-induced accentuation is in principle excluded under the architecture of the theory. Kurisu (2001) further proposes the Morpheme-Contiguity assumption, according to which all segmental and suprasegmental elements expressing a given morpheme must be realised contiguously. This assumption is captured by the constraint Morph-Contiguity, which is crucially undominated. As a result, preaccentuation must be strictly local, which poses a challenge for suffixes such as root-initial-accenting /-zu/ in Japanese. To capture the remaining suffix accentuation patterns, Kurisu (2001) resorts to constraint indexation, further increasing the complexity of the analysis. As Kurisu (2001: 213) explicitly states, “each faithfulness constraint must inherently bear the type marking of the relevant affix,” since the ranking of constraints such as Max-IO-Accent varies across affixes.

The brief review of alternative accounts shows that each relies on substantial extensions of OT, raising concerns about theoretical economy and restrictiveness, while also exhibiting more limited empirical coverage than the present analysis. The TAF account (Alderete 1999) introduces a new family of antifaithfulness constraints which require direct reference to morphological information through indexation. Moreover, the system encounters a ranking paradox when attempting to derive multiple accentual suffix types within a single grammar. Trommer (2019a) expands the representational apparatus of autosegmental OT by introducing complex floating primitives (e.g. H|) and vertical concatenation. Tt remains unclear how this machinery generalises to suffixes such as the adsorbent, which combine subtractive and attractive properties. Realise Morpheme Theory (Kurisu 2001) shares similar empirical limitations and relies on Sympathy Theory and constraint indexation. Its Morpheme-Contiguity assumption further restricts accentual effects to strictly local domains, making non-local accentuation difficult to accommodate. In contrast, the GSR account developed here captures a broader range of Japanese while maintaining the Indirect Reference Hypothesis and employing only independently motivated constraints within Harmonic Grammar.