2.1. Pitch accent in Japanese

Japanese words may be either accented, in which case they carry a single pitch accent (henceforth accent), or unaccented.Footnote ¹ The presence and location of accents in native and Sino-Japanese nouns have been traditionally considered unpredictable,Footnote ² as exemplified by a well-known triplet: [háɕi] (initial accent) ‘chopsticks’, [haɕí] (final accent) ‘bridge’ and [haɕi] (unaccented) ‘edge’ (acute accent marks indicate accented moras). An exception to this is that the second mora of a heavy syllable – a moraic nasal as in [ka ŋ kokɯ] ‘Korea’, the first element of a geminate consonant as in [ni p poɴ] ‘Japan’, or the second element of a long vowel or a diphthong as in [t͡ɕɯ ː gokɯ] ‘China’ – cannot carry an accent (McCawley Reference McCawley1968; Kubozono Reference Kubozono1993).

This study adopts Ito & Mester’s (2016) conventions for marking the accentuation of a word. That is, in addition to an acute accent mark, a superscript number is assigned to indicate the location of the accented mora counting backward from the end of the word (e.g., ³[ínot͡ɕi] ‘life’, ²[kokóɾo] ‘heart’ and ¹[atamá] ‘head’). Unaccented words are assigned the number ‘0’. Square brackets […] and parentheses (…) indicate prosodic words and metrical feet, respectively. The symbols ‘L’ and ‘H’ denote light and heavy syllables. (In §4.2, additional characters will be introduced to represent gradient syllable weight.) Angle brackets 〈…〉 signify an epenthetic vowel.Footnote ³

2.2. Loanword accentuation in Japanese

The literature on accent in Japanese loanwords dates back to McCawley (Reference McCawley1968), who discovered the ‘antepenultimate accent rule’. This rule states that accented loanwords carry an accent on the syllable containing the antepenultimate mora (e.g., ³[k〈ɯ〉ɾis〈ɯ́〉mas〈ɯ〉] ‘Christmas’ and ⁴[hambáːgaː] ‘hamburger’). More recently, Katayama (Reference Katayama1998) and Kubozono (Reference Kubozono2006) note that loanwords with certain syllable structures deviate from antepenultimate-mora accent. Specifically, they observe that loanwords ending in LH sequences tend to carry an accent on the syllable containing the preantepenultimate mora (e.g., ⁴[dók〈ɯ〉taː] ‘doctor’ and ⁵[béːkaɾiː] ‘bakery’). Kubozono attributes this accent pattern to a diachronic shift in the phonological grammar from the antepenultimate accent rule (which produces ³[LĹH] and ³[HĹH]) to a rule equivalent to the Latin stress rule (Allen Reference Allen1973): stress the penultimate syllable if it is heavy, otherwise stress the antepenultimate syllable (yielding ⁴[ĹLH] and ⁵[H́LH]). Furthermore, Kubozono (Reference Kubozono2006) finds that four-mora loanwords ending in LL syllables tend to be unaccented (e.g., ⁰[ameɾika] ‘America’ and ⁰[abokado] ‘avocado’), unless the word-final syllable is epenthetic (e.g., ⁴[mánmos〈ɯ〉] ‘mammoth’ and ⁴[ák〈ɯ〉ses〈ɯ〉] ‘access’). In a more recent development, Ito & Mester (Reference Ito and Mester2016) put forth the most comprehensive and ambitious formal existing model that attempts to capture Japanese-internal markedness principles of loanword accentuation, which will be reviewed in §4.1 below.

As noted in §1, there is no consensus on the existence of faithfulness effects. Although it is widely recognised that Japanese speakers sometimes mimic the main prominence of source words (e.g., ⁵[ák〈ɯ〉sent〈o〉] (< áccent) and ¹[ɸondjɯ́] (< French fondúe) (e.g., Shinohara Reference Shinohara2000; Kubozono Reference Kubozono2006; Ito & Mester Reference Ito and Mester2016), such faithfulness-driven accents have generally been considered exceptions to the phonological grammar and so have not been subject to serious investigation. For example, Ito & Mester (Reference Ito and Mester2016: 477) acknowledge the occurrence of this phenomenon as an exception, stating that ‘[a]lthough the majority of loans do not take into account the prominence location of the source word, some newer loans preserve the original prominence location of the source word’. This perspective is underscored by the absence of faithfulness constraints in their Optimality Theoretic analyses, a characteristic shared with certain earlier analyses, notably those by Katayama (Reference Katayama1998). Furthermore, in their effort to develop a taxonomy for loanword prosody, Davis et al. (Reference Davis, Tsujimura and Jung-yueh2012) describe English loanwords in Japanese as instances where the prosodic system of the source language has no influence.

One notable exception is Mutsukawa (Reference Mutsukawa2005, Reference Mutsukawa2006), who argues that faithfulness to the location of stress in source words is the dominant factor in determining accent in English loanwords.Footnote ⁴ However, this view might overestimate the impact of faithfulness effects. The basis for Mutsukawa’s conclusion is the observation that the majority of English loanwords in his corpus preserved the English stress, but this group crucially includes words following the antepenultimate accent rule. Kubozono (Reference Kubozono2006) expresses an intermediate view, arguing that while the tendency of English loanwords to be accented (as opposed to unaccented)Footnote ⁵ comes from Japanese speakers’ knowledge that English words are pronounced with a pitch fall in isolation, the location of the accent is determined by the native phonological grammar. Finally, Kubozono explores the influence of epenthetic syllables, observing that loanwords of the form LH tend to carry an accent on the final syllable when the initial syllable is epenthetic (e.g., ²[p〈ɯ〉ɾéː] ‘play’ and ²[b〈ɯ〉ɾɯ́ː] ‘blue’).Footnote ⁶ Ito & Mester (Reference Ito and Mester2016) also mention a similar phenomenon, but such instances are treated as exceptions.

3.1. Data

The data consists of English-based loanwords from the NHK pronunciation and accent dictionary (NHK Broadcasting Culture Research Institute 2016). All loanwords in the dictionary were manually extracted, and their syllable structures and accent patterns were documented. Loanwords with multiple accent variants were counted separately,Footnote ⁷ totalling 7,378 loanwords. Sequences of two vowels ending in a high vowel (i.e., /ai/, /oi/, /ui/, /au/, /eu/ and /iu/) were considered as diphthongs.Footnote ⁸

Many loanwords were excluded to focus on morphologically simple words and loanwords borrowed specifically from English. Excluded categories included compound-like loanwords, truncated ones and acronyms. Also excluded were loanwords derived from phrases, inflected words or those with productive affixes, based on Hayes’s (Reference Ryan2011) English Phonology Search program.Footnote ⁹ To ensure focus on English-based loanwords, ones whose source words are not found in the CMU Pronouncing Dictionary (Weide Reference Weide1994) or the Subtlex Corpus (Brysbaert & New Reference Brysbaert and New2009) were excluded.Footnote ¹⁰ Additionally, loanwords whose adaptations could not be attributed to either auditory or orthographic borrowing from English (e.g., ⁴[bɯ́ijon] rather than ⁴[bɯ́ilon] for ‘bouillon’ and ³[montáːʑ〈ɯ〉] rather than ³[montáːʑ〈i〉] for ‘montage’) were excluded, assuming that such loanwords were borrowed from languages other than English.

Additional loanwords were excluded from the analysis due to various reasons. First, loanwords from English words with two possible stress patterns, as determined by the English Phonology Search, were excluded (e.g., [ˈɪmˌpækt] / [ɪmˈpækt] ‘impact’ (noun vs. verb)). Second, loanwords with superheavy syllables (i.e., sequences of a long vowel or a diphthong followed by a moraic nasal) were excluded (e.g., ³[ɾáin] ‘line’), as there is no consensus on how they should be treated in Japanese accentuation. Third, loanwords with source words involving onset glides that were adapted into Japanese as a high vowel (e.g., ³[iéɾoː] rather than ³[jéɾoː] for ‘yellow’), or ones with source words containing onset palatal affricates or fricatives that were adapted with an [i] inserted after them (e.g., ⁴[maɾéːɕia] rather than ³[maɾéːɕa] for ‘Malaysia’), were excluded, because the status of such loanword syllables is unclear (i.e., full or epenthetic). Finally, loanwords longer than four syllables were excluded, to maintain a manageable size for the models developed in §4. As a result, a total of 3,017 loanwords remained for data analysis.

Each loanword syllable was annotated as either primary-stressed, secondary-stressed, unstressed or epenthetic, based on the CMU Pronouncing Dictionary and the English Phonology Search. When there was a discrepancy between these sources, the annotation from the English Phonology Search was followed. The data likely include both auditory and orthographic borrowings; distinguishing between the two can be challenging as they frequently overlap (Daland et al. Reference Daland, Mira and Kim2015), and the distinction between them is probably not always categorical (Hamann & Colombo Reference Hamann and Colombo2017).

3.2. Overview of the accent pattern in the corpus data

This section provides an overview of the accent patterns found in the corpus data, aiming to give a broad understanding of their distribution across syllable structures before delving into formal modelling. For reasons of space, I will omit the description related to the unaccented pattern, as it is not the primary focus of this study (see Appendix A for more details). Here, loanwords are categorised as either bimoraic/trimoraic or longer. This distinction is necessary because of the limited accent patterns the former can exhibit. Additionally, longer loanwords are further classified based on the syllable structure of the final three moras (i.e., LH, HL, HH, LLL and HLL), which play a key role in determining loanword accent.

Let us begin with a straightforward generalisation: words consisting of two or three moras have an initial accent. Out of 1,087 words in this category, 72% (786 words) have an initial accent (e.g., ²[bós〈ɯ〉] ‘boss’ and ³[kánada] ‘Canada’). In the case of the three-mora words, we may attribute this to the antepenultimate accent rule of McCawley (Reference McCawley1968). Exceptions often involve an epenthetic vowel. Among the 26 LH words beginning with an epenthetic syllable, 73% (19 words) bear penultimate-mora accent (e.g., ²[b〈ɯ〉ɾɯ́ː] ‘blue’, but ³[p〈ɯ́〉ɾan] ‘plan’), consistent with Kubozono’s (Reference Kubozono2006) description.

Turning to words with more than three moras, these roughly fall into two categories. First, words ending in HH or HLL typically obey the antepenultimate accent rule, with 81% (419 out of 515 words) conforming to this pattern (e.g., ⁴[bakéːɕon] ‘vacation’ and ⁴[hambáːgaː] ‘hamburger’). Second, words ending in HL, LLL or LH vary between antepenultimate-mora accent and preantepenultimate-mora accent. This variation even extends to propreantepenultimate-mora accent, with the accent on the syllable containing the mora two positions prior to the antepenultimate one. More precisely, words ending in HL or LLL are more likely to bear antepenultimate-mora accent, with 56% (573 out of 1,020 words) adhering to this accent pattern (e.g., ³[konkóːs〈ɯ〉] ‘concourse’ and ³[t͡ɕokoɾéːt〈o〉] ‘chocolate’) and 28% (284 words) bearing (pro)preantepenultimate-mora accent (e.g., ⁵[táːminaɾ〈ɯ〉] ‘terminal’ and ⁵[ébidens〈ɯ〉] ‘evidence’). In contrast, words ending in LH more often bear (pro)preantepenultimate-mora accent, with 26% (68 out of 395 words) following the antepenultimate accent rule (e.g., ³[kaːdígan] ‘cardigan’ and ³[t〈o〉ɾadíɕon] ‘tradition’) and 66% (262 words) bearing (pro)preantepenultimate-mora (e.g., ⁵[háːmoniː] ‘harmony’ and ⁵[kjáɾak〈ɯ〉taː] ‘character’). The prevalence of preantepenultimate-mora accent in words ending in LH is consistent with the observations made by Katayama (Reference Katayama1998) and Kubozono (Reference Kubozono2006).

Why do words ending in HH or HLL mostly bear antepenultimate-mora accent while ones ending in HL, LLL or LH exhibit a mix of antepenultimate-mora accent and (pro)preantepenultimate-mora accent? I suggest the following explanation. In the case of the former, faithfulness typically aligns with the antepenultimate accent rule. That is, the stress pattern of source words typically assigns stress on source syllables that correspond to loanword syllables including the antepenultimate mora (cf. the Latin stress rule in English). Indeed, this is true of 92% of the words in this category. On the other hand, for the latter structures, the Latin stress rule often stresses source syllables that match loanword syllables containing the (pro)preantepenultimate mora. This pattern is observed in 92% of words ending in LH and 57% of ones ending with HL or LLL.Footnote ¹¹ These proportions do not directly correspond to those of (pro)preantepenultimate-mora accent in loanwords with these structures (92% vs. 66% for LH endings, and 57% vs. 28% for HL or LLL endings). This discrepancy arises because sometimes faithfulness to source words is respected, while at other times, the antepenultimate accent rule is followed. This conflict will be modelled in §4.

4.1. Baseline model: Ito & Mester (Reference Ito and Mester2016)

In this section, I provide a brief overview of Ito & Mester’s (Reference Ito and Mester2016) OT model and explain the adjustments I have made to adapt its structure to MaxEnt modelling.Footnote ¹² Due to space limitations, I focus on the aspects most pertinent to the modelling process; readers should refer to their original paper for full details.

Ito & Mester’s model takes syllable structures as inputs and logically possible foot structures as output candidates, under the three requirements outlined in (1). These are: (1a) accents must align with the head syllable of the head foot; (1b) every prosodic word must have at least one foot; and (1c) feet can have a maximum of two syllables. In their notation, capital letters represent head syllables, and small capitals represent non-head syllables.

These requirements reflect three undominated constraints: Word Prominence to Word Head, Headedness and the maximal version of FootBinarity. In a best-fit MaxEnt model, these undominated constraints, which are never violated by winners, are assigned infinite weight, effectively nullifying the probability of any candidate violating them. This approach helps limit the number of candidates for computational consideration by excluding those that violate these undominated constraints.

To further streamline the output structure, my models also omit candidates that violate Ito & Mester’s MoraicTrochee constraint, which is considered undominated in their analysis. This constraint disallows feet larger than two moras or with an iambic pattern (i.e., (Hh), (hH), (Hl), (hL), (Lh), (lH) and (lL)). Aligning with their approach, my models assume feet are maximally two moras and follow a trochaic pattern. Thus, I avoid using capital and small capital letters to differentiate head from non-head syllables, since all feet in my model are trochaic. For instance, (LL) consistently represents (Ll) rather than (lL).

Ito & Mester rank ten markedness constraints (excluding MoraicTrochee) to capture markedness principles in loanword accentuation in Japanese. These constraints and their rankings are shown in (2), with definitions slightly modified for accessibility to those unfamiliar with Ito & Mester’s analysis.

Ito & Mester’s model predominantly assigns antepenultimate-mora accent to accented loanwords, as illustrated in (3).

The tableaux in (3) show that loanwords with three light syllables (represented by ³[bánana] ‘banana’) and five light syllables (represented by ³[baɾɯséɾona] ‘Barcelona’) bear antepenultimate-mora accent. In brief, this is due to the requirement that syllables are maximally parsed into feet, leaving the final syllable unparsed (i.e., ³[(LL)L] and ³[(LL)(LL)L]).

The key aspect of Ito & Mester’s analysis is that the same ranked constraints result in the unaccented pattern for loanwords with four light syllables (LLLL) and ones ending in HLL.Footnote ¹³ (4) shows a sample tableau for the LLLL structure (represented by ⁰[ameɾika]).

For loanwords with four light syllables (and ones ending in HLL), the optimal foot structure involves parsing the final four moras into bimoraic feet – e.g., [(ame)(rika)], as in (4a) – due to the undominated status of NoLapse and the relatively high ranking of InitFt. Moreover, assigning an accent to either foot critically violates either Rightmost, as in (4b), or NonFin(Ft′), as in (4c), both of which constraints take precedence over WdAcc, making the unaccented candidate (4a) the optimal choice.

Finally, in Ito & Mester’s model, loanwords ending in LLH receive preantepenultimate-mora accent. (5) displays a sample tableau for this pattern (represented by ⁴[dóɾagon] ‘dragon’).

In this context, the undominated NonFin(σ) constraint plays a crucial role. It disqualifies candidates with the final heavy syllable parsed – (5b)–(5d) – making exhaustive footing suboptimal for these structures. Furthermore, FtBin rules out the candidate with antepenultimate-mora accent, (5e), making candidate (5a), with accent on the preantepenultimate mora, the optimal choice. My corpus data, along with descriptions by Katayama (Reference Katayama1998) and Kubozono (Reference Kubozono2006), show that not only loanwords ending in LLH but also ones ending in HLH often bear preantepenultimate-mora accent (i.e., ⁴[…ĹLH] and ³[…H́LH]). This minor discrepancy, however, is less significant in probabilistic models where constraints are not ranked. Additionally, the occurrence of preantepenultimate-mora accent can often be attributed to faithfulness effects to source words, as discussed in §3.2 above.

4.2. Expanding input structure

As in Ito & Mester’s (Reference Ito and Mester2016) model, my MaxEnt models take syllable structures as inputs. However, they require finer distinctions due to the integration of more factors. A key aspect is distinguishing loanword syllables based on their English source syllables: primary stressed (e.g., ˈL), secondary stressed (ˌL), unstressed (L) and epenthetic (〈L〉). Additionally, the models distinguish syllables with devoiced vowels (L̥) from ones with voiced vowels (L), and they distinguish multiple categories of heavy syllables: those with an obstruent coda (as in ³[kápp〈ɯ〉] ‘cup’; these will be labelled G); those with a nasal coda (e.g., ³[dáns〈ɯ〉] ‘dance’, labelled N); and those with a long vowel or diphthong (e.g., ³[páːk〈ɯ〉] ‘park’, labelled V). The rationale for these distinctions will be outlined in §4.4.2. These distinctions lead to a total of 485 inputs, each representing at least one loanword in the corpus, and include 21,503 output possibilities. To enable direct comparisons among the models, this structure is maintained consistently across all model updates. The MaxEnt spreadsheet is provided in the supplementary material.

As in Ito & Mester’s (Reference Ito and Mester2016) model, the inputs in my models are based on segmentally adapted loanword forms rather than the original English source forms. This approach assumes that input forms already incorporate crucial segmental processes of loanword adaptation. Thus, symbols in the input (i.e., L, G, N, V and 〈L〉) represent syllable types as they are adapted from source structures. The model also assumes the input forms encode certain native segmental processes. In the context of the present analysis, the only relevant process is high vowel devoicing (see §4.4.2). Hence, the symbol L̥ in the input denotes a source structure adapted as a light syllable in an environment where the process of high vowel devoicing is applicable.

4.3. Hidden structure

Given the learning data and a set of constraints, a MaxEnt model finds the constraint weights that minimise the difference between observed and predicted probabilities. As in Ito & Mester’s (Reference Ito and Mester2016) model, each output in my models corresponds to a unique foot structure. However, these structures are not directly observable; learners infer them from accumulated surface accent patterns. This is an instance of the ‘hidden structure’ problem (e.g., Tesar & Smolensky Reference Tesar and Smolensky1998; Jarosz Reference Jarosz2015).

In addressing the hidden structure problem here, I assume that the learning data provided to my MaxEnt models only contain surface accent patterns of loanwords, devoid of any information about foot structure. For instance, instead of having access to representations like ³[(bána)na] or ³[(bá)nana], the models receive surface accent patterns such as ³[bánana]. This assumption is implemented by aggregating the predicted probabilities of all possible foot structures for the same surface accent pattern, following the method put forth by Moore-Cantwell (Reference Moore-Cantwell2020) in an analysis of English stress assignment. This approach ensures constraint weights are based on surface accents, reflecting data actually available to learners, while the model determines the distribution of probability across various foot structures.

4.4. Comparing a series of models

In this section, I examine a series of MaxEnt models, gradually integrating more factors into Ito & Mester’s baseline model. This baseline model undergoes three updates to include (i) additional markedness effects, (ii) effects of faithfulness to source words, and (iii) effects of faithfulness to native speakers’ knowledge of the English stress system, termed the JTOE. Each update is justified by improved model accuracy, evaluated using likelihood ratio tests (Wasserman Reference Wasserman2004), with a significance level of 0.05. The difference in log likelihood between the full and subset models is denoted as Δ log likelihood. Additionally, the Akaike information criterion (AIC) and the Bayesian information criterion (BIC) are presented for reference in a summary table (Table 1), although these metrics are not explicitly discussed. While the models are named after their newly added factors, all previously incorporated factors are retained upon integration; a final pruning of non-significant constraints is carried out in §4.4.5.

Table 1

Best-fit constraint weights, log likelihood, AIC and BIC for a series of MaxEnt models: M1, the MaxEnt version of Ito & Mester’s model; M2, the augmented Ito–Mester model; M3, the faithfulness model; M4, the JTOE model; and the final model.

The loanwords from the corpus data are used as learning data for the models. To fit the constraint weights, I use the Excel Solver tool (Fylstra et al. Reference Fylstra, Lasdon, Watson and Waren1998), which employs Conjugate Gradient Descent to find weights in a way as to maximise the likelihood of the model, under the condition that weights must be positive (act as penalties).

4.4.1. A MaxEnt version of Ito & Mester’s model

The first model to examine is a probabilistic version of Ito & Mester’s (Reference Ito and Mester2016) model. The column labelled M1 in Table 1 presents the best-fit constraint weights (along with the log likelihood, AIC and BIC) in this model. The log likelihood of the model, used as a benchmark for later models, is $-$ 2221.48.

Many constraints received weights, indicating their relevance. However, the weight distribution does not fully mirror the constraint system in Ito & Mester’s categorical model. Notably, NoLapse and Rightmost received relatively small weights (0.36 and 0.77), despite being undominated in their original model. The small weight for NoLapse may be due to its overlap with the broader Parse-σ constraint, but there is also an empirical reason. The corpus study uncovered abundant cases of (pro)preantepenultimate-mora accent that are not predicted by Ito & Mester’s model. These instances necessarily violate either NoLapse or Rightmost (e.g., ⁵[(kón)saːt〈o〉] or ⁵[(kón)(saː)t〈o〉] ‘concert’). To allocate some probabilities to such accent patterns, the model requires smaller best-fit weights for these constraints. Additionally, InitialFoot received a zero weight. Since the primary role of this constraint is to ensure exhaustive footing for four-mora structures ending in LL, which is crucial for rendering them unaccented (e.g., ⁰[(LL)(LL)] and ⁰[(H)(LL)]), the absence of weight for this constraint suggests potential inaccuracies in predicting the unaccented pattern, as discussed in the following paragraph.

Figure 1 compares observed corpus probabilities with those predicted by the MaxEnt version of Ito & Mester’s model. Each data point represents a surface accent pattern for an input, with foot structures collapsed as detailed in §4.3. To minimise bias from rare patterns, only inputs with at least five individual words are included in these scattergrams and subsequent ones. The left panel shows aggregate results, while the right panel breaks down the data by accent patterns (Pre2 = propreantepenultimate-mora accent, Pre = preantepenultimate, Ant = antepenultimate, Pen = penultimate, Ult = ultimate and Un = unaccented). The broad scatter of data points indicates the model’s limitations in capturing certain accent patterns in the corpus data. Notably, it tends to underpredict the accent patterns that are not predicted by Ito & Mester’s original model, especially certain instances of (pro)preantepenultimate-mora accent, while overpredicting antepenultimate-mora accent. The model also tends to underpredict the unaccented pattern.

Figure 1

Comparison of observed probabilities from corpus data with predicted probabilities from the MaxEnt version of Ito & Mester’s model.

4.4.2. The augmented Ito–Mester model

The second probabilistic model incorporates two additional markedness effects. First, it addresses vowel devoicing in Japanese. High vowels /i, ɯ/ typically devoice between voiceless consonants (e.g., ⁰[ɕi̥ka] ‘deer’; McCawley Reference McCawley1968), and in this environment, they are less likely to carry an accent (e.g., ²[s〈ɯ̥〉pín] ‘spin’; McCawley Reference McCawley1977; Haraguchi Reference Haraguchi1991; Tsuchida Reference Tsuchida1997, Reference Tsuchida2001).Footnote ¹⁴ Second, it accounts for the finer differences in heavy syllables (distinguishing G, N and V as described above), recognising that syllable weight can be gradient rather than binary (Gordon Reference Gordon2002, Reference Gordon2007; Ryan Reference Ryan2011). This update is crucial for accurately evaluating faithfulness effects in subsequent models. Notably, it is essential to represent vowel devoicing accurately to assess the impact of epenthetic vowels, given many devoiced vowels are epenthetic.

The constraints in (6) formalise the effects of devoiced vowels (in (6a)) and gradient syllable weight (in (6b)–(6d)). The constraint in (6a) is equivalent to *Accented[+s.g.] in Tsuchida (Reference Tsuchida1997, Reference Tsuchida2001), but its name has been altered within the context of this study.

• 

Column M2 in Table 1 shows the best-fit constraint weights in the augmented Ito–Mester model. *DevAcc received a weight of 2.09, confirming the tendency for syllables with devoiced vowels to avoid accent. The gradient versions of WSP received varying weights based on sonority: WSP(V) is high (2.02), WSP(N) is small (0.36) and WSP(G) is zero. This leads to the weight of the original WSP constraint dropping to zero, indicating that the primary effect of WSP arises from heavy syllables with long vowels or diphthongs. Likelihood ratio tests confirm each of these additions significantly improves the model’s fit (*DevAcc: $\Delta\ \textrm{log likelihood} = 43.75$ , $p < 0.001$ ; WSP(G/N/V): $\Delta\ \textrm{log likelihood} = 29.73$ , $p < 0.001$ ). The log likelihood of the model improved to $-2147.72$ from $-2221.48$ ( $\Delta\ \textrm{log likelihood} = 73.76$ ) in the previous model.

Figure 2 shows the predicted-vs.-observed plot for the augmented Ito–Mester model. While the correlation has improved, the same issues remain: underprediction of (pro)preantepenultimate-mora accent and the unaccented pattern, and overprediction of antepenultimate-mora accent.

Figure 2

Comparison of observed probabilities from corpus data with predicted probabilities from the augmented Ito–Mester model.

4.4.3. The faithfulness model

The next step involves adding faithfulness effects to the enriched markedness system. This process takes into account non-native phonological properties in loanword adaptation, including English stress, consonant clusters and word-final (non-nasal) consonants, the latter two of which result in epenthetic syllables in loanword forms (e.g., /pleɪ/ → ²[p〈ɯ〉ɾéi] ‘play’, /ʃɑp/ → ³[ɕópp〈ɯ〉] ‘shop’; note that placing an accent on word-final light syllables is highly marked in any case).

To formalise the faithfulness effects, I introduce three loanword-specific faithfulness constraints that govern the correspondence between English source words and their loanword counterparts, as shown in (7). Dep[Acc] in (7a) and Dep[Acc](PS) in (7b) implement the effect of stress in general and of primary stress specifically. Note that violating the latter constraint implies violating the former as well. Dep[Acc](PS) corresponds to FaithLoc(Accent) in Mutsukawa (Reference Mutsukawa2005, Reference Mutsukawa2006). To my knowledge, however, the broader effects of stress (encompassing both primary and secondary) have not been explored in Japanese loanword accentuation. Dep[V́] in (7c) assesses the claimed tendency of epenthetic syllables to avoid carrying an accent. This is analogous to *v (epenthetic vowel) or Head(Foot)-Dep, introduced by Shinohara (Reference Shinohara2000, Reference Shinohara, Kager, Pater and Zonneveld2004) in the context of Japanese speakers’ online adaptation of French words.

• 

The column labelled M3 in Table 1 presents the best-fit constraint weights in the faithfulness model. Dep[Acc] received a weight of 0.45, indicating that loanword syllables corresponding to English stressed syllables (primary or secondary) attract accent. Dep[Acc](PS) received a weight of 0.54 on top of Dep[Acc], indicating a stronger effect of primary than secondary stress (this result turns out to be less compelling than it may initially appear, however; see §4.4.4). Dep[V́] received a weight of 0.59, confirming the tendency for epenthetic syllables to avoid accent. Notably, the inclusion of Dep[V́] reduced the weight of *DevAcc from 2.09 to 1.20, reflecting the overlap between devoiced and epenthetic syllables. Likelihood ratio tests confirm that each faithfulness constraint significantly improves the model’s fit (Dep[Acc]: $\Delta\ \textrm{log likelihood} = 4.32$ , $p < 0.005$ ; Dep[Acc](PS): $\Delta\ \textrm{log likelihood} = 7.59$ , $p < 0.001$ ; Dep[V́]: $\Delta\ \textrm{log likelihood} = 7.27$ , $p < 0.001$ ). The log likelihood of the model increases to −2007.65 from −2147.72 in the augmented Ito–Mester model ( $\Delta\ \textrm{log likelihood} = 140.07$ ), indicating a significant improvement of the model’s fit to the observed data.

Figure 3 displays the predicted-vs.-observed plot from the faithfulness model. It demonstrates significantly improved correlation, effectively addressing previous inaccuracies, including the underprediction of (pro)preantepenultimate-mora accent and the overprediction of antepenultimate-mora accent. The reason for this improvement was given above in §3.2: in many syllable structures, the observed accent patterns reflect a stochastic conflict of markedness and faithfulness principles. However, the underprediction of the unaccented pattern persists.

Figure 3

Comparison of observed corpus probabilities with predicted probabilities from the faithfulness model.

4.4.4. The JTOE model

Identifying a tendency for English loanwords to preserve the stress patterns of their source words prompts an additional question: beyond merely replicating the stress patterns of individual source words, do Japanese speakers also internalise these stress patterns and apply this knowledge in the adaptation of certain English words? Specifically, when the stress patterns of English words are atypical, do Japanese speakers assign accents that align with the general stress patterns of English, in a manner that cannot be accounted for by markedness principles? This section demonstrates that generalisations about English stress patterns do indeed have a discernible impact on the accentuation of loanwords in Japanese.

Research indicates that even limited exposure to a foreign language enables speakers to develop sophisticated phonotactic knowledge of that language. For example, Oh et al. (Reference Oh, Todd, Beckner, Hay, King and Needle2020) and Panther et al. (Reference Panther, Mattingley, Todd, Hay and King2023) report that New Zealanders who do not speak Māori but are extensively exposed to it can evaluate the well-formedness of Māori-like non-words just as well as fluent Māori speakers. In the context of loanword adaptation, the influence of borrowers’ knowledge of the source language phonology, especially knowledge of its phonotactics, has been relatively understudied. An exception is the work of Kang et al. (Reference Kang, Phạm, Storme, Albright and Fullwood2015), who highlight the significance of this factor in the context of Vietnamese speakers adapting French words.Footnote ¹⁵ Given Japanese speakers’ high level of exposure to English, it is plausible that they develop phonotactic knowledge of English, including stress patterns. I refer to this knowledge as the JTOE. Hence, the model that incorporates this component is referred to as the JTOE model.

One anecdotal source of evidence for such a theory is a process called ‘hyperforeignisation’ (Janda et al. Reference Janda, Joseph, Jacobs, Lima, Iverson and Corrigan1994), where speakers overapply patterns from known non-native forms to novel ones. For example, English speakers often stress the penultimate syllable in borrowed words, regardless of their original pronunciation. Notable cases include pronouncing the Japanese place name Nagasaki as [ˌnɑgəˈsɑki] and the Italian name Cristofori as [ˌkɹɪstəˈfoʊɹi], even though the former has antepenultimate-mora accent in Japanese (³[nagásaki]) and the latter has stress on the antepenultimate syllable in Italian ([kriˈstɔːfori]). In these examples, English speakers overapply the penultimate stress rule, which likely is induced by their exposure to Spanish (and Italian) words ending with a light syllable, in loanwords borrowed from Japanese and Italian.

Japanese speakers, too, appear demonstrate hyperforeignisms. Table 2 presents loanwords with (pro)preantepenultimate-mora accent that deviate from the stress patterns of their source words, leading to their underprediction in the faithfulness model. Crucially, these accent patterns are also not accounted for by the markedness effects as assumed in this study. A detailed examination of the data indicates that their accent patterns mirror typical English stress patterns. This section focuses on modelling the effect of JTOE and its role in improving the accuracy of our loanword accentuation model.

Table 2

Potential hyperforeignisms with (pro)preantepenultimate-mora accent, unsupported by either faithfulness to source words or markedness principles.

To integrate the effect of JTOE, we start by creating a preliminary model of it. The goal here is to develop a model that reasonably represents Japanese speakers’ knowledge of the English stress system, although what exactly such a model should look like must be established empirically. The aim is not to construct an exhaustive model of the English stress system itself, as that would exceed the scope of this study. For this purpose, I modify an existing MaxEnt model of English stress by Hayes (Reference Hayes2020), designed for a class exercise. This model predicts primary stress in about 12,600 English words based on their syllabic and segmental structures, using 28 constraints, mostly extracted from the literature on English stress (e.g., Chomsky & Halle Reference Chomsky and Halle1968; Liberman & Prince Reference Liberman and Prince1977; Hayes Reference Hayes1982).

Unlike my MaxEnt models, which use syllable structures as inputs, the English stress model takes individual words as inputs. As a result, some of the constraints in the model refer to specific types of segments or morphemes. I removed those constraints from the model as they are likely too detailed for the JTOE. For instance, in the original model by Hayes, a constraint requiring penultimate stress when the final syllable has a palato-alveolar onset receives a substantial weight (e.g., [ɪˈnɪʃəl] ‘initial’ and [dɪˈlɪʃəs] ‘delicious’). However, such detailed constraints are unlikely to be internalised by Japanese speakers, as suggested by adaptations like ⁴[íniɕaɾ〈ɯ〉] and ⁴[déɾiɕas〈ɯ〉]. Another crucial difference from my models is that the English stress model does not presuppose separate foot structures for each stress pattern. Therefore, the number of outputs for each input equals the number of stress patterns (or syllables) in the input.

In addition, I replaced the original SupHFin constraint, which requires stress on any word-final superheavy syllable, with three more specific constraints: SupH(vvc)Fin, SupH(vnc)Fin and SupH(vcc)Fin, defined in (8h)–(8j). These differentiate the three types of superheavy syllables in English: VVC, VNC and VCC (where V is a vowel, N a nasal coda and C a non-nasal coda). These distinctions were made because inputs in my MaxEnt models distinguish corresponding adapted forms in Japanese: V〈L〉, N〈L〉 and L〈L〉〈L〉. This modification not only aligns the structure of JTOE with my models but also improves the accuracy of the English stress model, highlighting the significance of differentiating these superheavy syllables in English stress. Notably, the greater weight assigned to SupH(vvc)Fin over SupH(vnc)Fin (2.97 vs. 1.53) mirrors the similar weight difference between WSP(V) and WSP(N), as discussed in §4.4.2.

Following these modifications, a total of ten constraints remained, listed in (8). I ran the model with the revised constraint set, allowing weights to be negative (rewarding violations instead of penalising them) as well as positive.Footnote ¹⁶ The constraint weights in the best-fit model are also shown in (8).

• 

The next step involved calculating the probabilities of stress patterns for English source words in my corpus data. To do this, I established assumptions about how English syllables are typically adapted into Japanese, as shown in Table 3.Footnote ¹⁷ In brief: English light syllables are adapted as Japanese light syllables (L → L); heavy syllables are adapted as Japanese heavy syllables with a nasal coda or a long vowel/diphthong (H → N/V), as sequences of a heavy syllable with an obstruent coda and an epenthetic syllable (H → G〈L〉), or as a light syllable and an epenthetic syllable (H → L〈L〉). English superheavy syllables are adapted as Japanese superheavy syllables (S → S; note that loanwords containing a superheavy syllable are not included in the current analysis), as sequences of a heavy syllable with a long vowel/diphthong and any number of epenthetic syllables (S → V〈L〉), as sequences of a heavy syllable with a nasal coda and any number of epenthetic syllables (S → N〈L〉), or as sequences of a light syllable and multiple epenthetic syllables (S → L〈L〉〈L〉).

Table 3

General assumptions about how English syllables are adapted into Japanese.

Using these assumptions, I inferred the syllable structures of English source words from their adapted loanword forms. Note that the inferred structures may differ from the actual English structures, as real-world adaptations do not always follow these assumptions. For instance, orthographic adaptations like oasis → ⁴[óaɕis〈ɯ〉] may deviate: the inferred structure (LLH) differs from the actual structure (HHH based on /oʊˈeɪsɪs/). In fact, such deviation is favourable as it seems to better reflect how Japanese speakers infer the source pronunciations of orthographic borrowings. That is, they would not typically infer /oʊˈeɪsɪs/ from the orthographic form oasis; the conjectured pronunciations, while probably varying among individuals, are likely to align more closely with the adapted Japanese form rather than the actual English pronunciation in terms of the syllable structure.

Using this approach, I compiled conjectured English syllable structures derived from the adapted Japanese forms. These syllable structures were then input into the best-fit English stress model (i.e., JTOE), which returned probabilities for stress patterns based on their violation profiles. Table 4 shows these predicted probabilities. Note that σ in these structures represents any syllable type, and any number of consonants can precede or intervene between syllables to form consonant clusters, which result in epenthetic syllables in loanword forms. While a thorough evaluation of the English stress model is not the focus here, readers familiar with the literature on English stress will probably recognise that the highest probabilities (highlighted in bold) are assigned to the patterns expected from the existing literature. Indeed, we observe that these patterns generally align with the Latin stress rule, although they are expressed as probabilistic tendencies.

Table 4

JTOE probabilities for English stress patterns (the dominant stress pattern for each phonological shape is shown in bold).

The final step integrates the faithfulness effects with the outputs of JTOE. To this end, I formalise Faith-JTOE[Acc] as in (9). This constraint serves as a bias towards accent patterns that align with the dominant stress patterns in English, by penalising deviations from JTOE probabilities.

• 

The tableau in (10) illustrates how violations of Faith-JTOE[Acc] are assigned to candidates for the input /LˈLL〈L〉/ (e.g., /ɕiˈatoɾ〈ɯ〉/ ‘Seattle’), along with violations of the faithfulness constraints. (Dep[Acc] and Dep[Acc](PS) are collapsed into Dep[Acc], as they assign the same violations in this example.) In this sample tableau, candidates sharing the same surface accent are collapsed for the sake of simplicity.

• 

JTOE probabilities indicate that antepenultimate stress (/ˈLLL〈L〉/) is the most common (0.77), with penultimate stress (/LˈLL〈L〉/) being less frequent (0.23), and ultimate stress (/LLˈL〈L〉/) being nonexistent (0.00) in the data. The Faith-JTOE[Acc] violations are calculated as the difference between 1 and the JTOE probability for each accent pattern, resulting in violation scores of 0.23 ( $=1-0.77$ ) for preantepenultimate-mora accent, 0.77 ( $=1-0.23$ ) for antepenultimate-mora accent, and 1.00 ( $=1-0$ ) for penultimate-mora accent. This creates a conflict between Dep[Acc] constraints, favouring antepenultimate-mora accent, and Faith-JTOE[Acc], favouring preantepenultimate-mora accent. Candidates with ultimate-mora accent and unaccented ones do not violate Faith-JTOE[Acc], but their probabilities are reduced as necessary by Dep[V́] and WdAcc, respectively.

Column M4 in Table 1 shows the best-fit constraint weights in the JTOE model. As the table shows, Faith-JTOE[Acc] receives a weight of 0.98. Note that this addition results in decrease of Dep[Acc](PS) from 0.54 in the faithfulness model to 0.27, as Faith-JTOE[Acc] partly absorbs the effect of Dep[Acc](PS) (i.e., because the JTOE is often correct). The log likelihood of the model increases to −1982.28 from −2007.65 in the faithfulness model ( $\Delta\ \textrm{log likelihood} = 25.36$ ). A likelihood ratio test confirms that the inclusion of this constraint significantly enhances the model’s fit to the data ( $p < 0.001$ ).

Figure 4 shows the predicted-vs.-observed plot based on the JTOE model. The correlation becomes even stronger, although the model continues to underpredict the unaccented pattern.

Figure 4

Observed probabilities based on the corpus data vs. predicted probabilities based on the JTOE model.

Finally, Table 5 compares the observed and predicted probabilities for the accent patterns of potential hyperforeignisms (from Table 2) in the faithfulness and JTOE models. Inputs with only one actual word were excluded to avoid overestimating observed probabilities. Table 5 reveals that the faithfulness model underpredicts all accent patterns, while the JTOE model reduces this underprediction, albeit not completely.

Table 5

Observed probabilities for accent patterns of potential hyperforeignisms from Table 2 (excluding inputs with only one word), compared with predictions from the faithfulness and JTOE models.

4.4.5. The final model

This section presents the final model of loanword accentuation in Japanese, removing constraints whose contribution does not pass the significance test. In addition, I assess the contribution of each component of the grammar: markedness, faithfulness and JTOE.

Likelihood ratio tests on individual constraints reveal that the effects of NoLapse, InitFt, WSP, WSP(G), WSP(N) and Dep[Acc](PS) are not statistically significant. This suggests that the foot structures targeted by NoLapse and InitFt (i.e., successive unparsed syllables and word-initial unparsed syllables) are not notably worse than other unparsed syllables penalised by Parse-σ. This is not surprising given the unaccented pattern remains underpredicted even in the final model: the primary function of these two constraints is to exhaustively parse syllables, which is necessary for producing unaccented words (i.e., ⁰[(H)(LL)] and ⁰[(LL)(LL)]). The insignificance of WSP, WSP(G) and WSP(N) suggests that the widely recognised tendency of heavy syllables to attract stress is primarily due to heavy syllables with long vowels or diphthongs (i.e., WSP(V)). Additionally, the role of Dep[Acc](PS) is largely subsumed by Faith-JTOE[Acc], but Dep[Acc] remains significant, indicating faithfulness to both primary and secondary stress in source words.

The column labelled Final in Table 1 showcases the best-fit constraint weights in the final model. The log likelihood of the final model was −1984.08, a drop of 1.80 from the larger model discussed in the previous section.

To assess the impact of each component in the model, each component was removed from the final model, and the changes in log likelihoods (Δ log likelihoods) were compared. Note that WdAcc is considered a markedness constraint, although it can also be viewed as a faithfulness constraint (i.e., Max[Accent]). The results, summarised in Table 6, show that the markedness effects have the largest contribution (Δ log likelihood $= 1757.16$ ). This is followed by the faithfulness effects to source words ( $\Delta\ \textrm{log likelihood} = 122.57$ ), and the faithfulness effects to the JTOE have the least impact ( $\Delta\ \textrm{log likelihood} = 31.37$ ).

Table 6

Contribution of each component of the grammar.

4.5. Summary

The MaxEnt modelling described in this section demonstrates the significant contribution of faithfulness effects to English source words and Japanese speakers’ implicit knowledge of the English stress system (along with additional markedness effects).

The faithfulness model confirmed the significance of two faithfulness effects: loanword syllables from stressed syllables in English source words, whether primary or secondary, tend to be accented (i.e., Dep[Acc]), and epenthetic syllables, derived from consonant clusters in English, tend to avoid accents (i.e., Dep[V́]). This improvement in the model’s fit to the data ( $\Delta\ \textrm{log likelihood} = 122.57$ from the final model in §4.4.5) suggests that loanwords with faithfulness-driven accents represent a probabilistic interplay between markedness and faithfulness, rather than being idiosyncratic exceptions. However, the impact of faithfulness effects is substantially smaller than that of markedness effects (1757.16 vs. 122.57), highlighting the secondary nature of faithfulness in the overall model.

The JTOE model confirmed the crucial role played by Japanese speakers’ implicit knowledge of the English stress system, particularly in assigning (pro)preantepenultimate-mora accent when the stress patterns of the source words and markedness principles favour other accent patterns. In this case, faithfulness to the JTOE takes precedence over faithfulness to the actual input and markedness effects, resulting in hyperforeignisation. To my knowledge, this model represents the first attempt to incorporate a module that reflects borrowers’ theory of a source language into a model of loanword adaptation.

Finally, the final model presented in this section encounters certain limitations, notably its underprediction of the unaccented pattern. A possible hypothesis is that loanwords with typically unaccented syllable structures (i.e., [HLL] and [LLLL] ending in a full vowel) might frequently be orthographic borrowings, devoid of faithfulness to the stress patterns of their source words. This might also lead to Japanese speakers’ uncertainty about the general stress patterns of source words with such structures, reducing the impact of the JTOE effect as well. A deeper investigation into this issue would require differentiating between auditory and orthographic borrowings and refining the JTOE.