2.1. Portuguese

Primary stress in Portuguese is constrained by a trisyllabic window, as mentioned above: marítimo ‘maritime’, martélo ‘hammer’, papél ‘paper’; pre-antepenultimate stress is illicit in the language: *máritimo. Even though both verbs and non-verbs respect this trisyllabic window, stress in these two classes of words is driven by different factors: stress in verbs is heavily influenced by morphological factors (see Wetzels (2007) for a review), while stress in non-verbs relies mostly on phonological factors like weight (Bisol, 1992; Garcia, 2017b; Lee, 2007; Wetzels, 2007). In this paper, we focus on stress in non-verbs, which is typically assigned as per Rule (1)—H stands for a heavy syllable, L for a light syllable, and X for any syllable (H or L).

Traditionally, all stress patterns that deviate from (1) are considered to be irregular. These include all words in the Houaiss Dictionary (Houaiss, Villar, and de Mello Franco, 2001) with antepenultimate stress (13% of non-verbs; Garcia (2014)), regardless of the weight profile involved (H́LL: fósforo ‘match’ (noun), ĹHL: pénalti ‘penalty’, ĹLH: júpiter ‘Jupiter’). Irregular cases also include words with penultimate stress that have a heavy final syllable (XX́H; esténcil ‘stencil’, nível ‘level’; 11% of non-verbs), and words with final stress that have a light final syllable (XXĹ: jacaré ‘alligator’, tatú ‘armadillo’; 3% of non-verbs). Approximately 72% of the Portuguese lexicon can be accounted for by the algorithm in (1) (Garcia, 2017b), which makes it relatively robust. In spite of this, researchers do not agree on what type of feet Portuguese builds, as well as what extra machinery must be employed to account for the various irregular patterns found in the language.

Bisol (1992) proposes that regular stress in Portuguese requires both moraic and syllabic trochees: moraic trochees capture XXH́ words, which contain a heavy final syllable and bear final stress; syllabic trochees capture XX́L words, which contain a light final syllable and bear penultimate stress. Bisol also assumes that both syllabic and moraic trochees are employed in irregular patterns. She proposes that the final syllable in words with antepenultimate stress is exceptionally marked as extrametrical.³ As a result, these words have the same foot structure as regular penultimate stress (i.e., they involve syllabic trochees).

Bisol’s approach does not straightforwardly extend to cases of word-final stress on light syllables. To capture this pattern, some scholars assume that Portuguese builds both trochees and iambs (Bonilha, 2004; Lee, 2007). Bonilha (2004), for example, proposes that iambs are built in words such as abacaxí ‘pineapple’ and urubú ‘vulture’ due to the word-final vowels in question (/i, u/): underlyingly high vowels in word-final open syllables virtually always attract stress. According to Bonilha, /i, u/, “when positioned at the end of the prosodic word, are considered good peak elements” (p. 41). This assumption, however, is inconsistent with the fact that, among vowels, high vowels have the lowest sonority (de Lacy, 2006; Ladefoged & Johnson, 2011). Indeed, in Portuguese, /e, o/ in final unstressed syllables are lenited to [i, u], respectively, consistent with the low sonority of the latter. If final stress on light syllables were truly driven by sonority, then XXL words ending in /a/ would be the best candidates for final stress—but this is not the case.

Lee (2007) also assumes that both iambic and trochaic feet play a role in the grammar of Portuguese. Lee proposes an optimality-theoretic account and, thus, takes advantage of the position that constraints that strive for outputs conforming to both foot types will be present in every grammar: FtForm = Iamb and FtForm = Trochee. Iambic feet are built in XLĹ words such as jacaré ‘alligator’: ja(caré), and trochaic feet are built in XĹL words such as caválo ‘horse’: ca(válo). Note that, under this approach, having two foot types active in a single language results in indeterminacy. For example, a XLH́ word could be parsed with a final moraic trochee (i.e., XL(H́)) or bisyllabic iamb (i.e., X(LH́)).

Some analyses of Portuguese assume that theme vowels play a role in stress assignment (e.g., Lee, 2007; Pereira, 1999, 2007). Theme vowels are always unstressed, and consist of {e, a, o} in Portuguese—in a word such as gát-o ‘cat’, the theme vowel (-o) indicates gender (masculine). In these analyses, stress is assumed to fall on the last vowel of the stem: gát]-o, jornál] ‘newspaper’, café] ‘coffee’. As a result, feet are not necessary to account for regular stress patterns: if theme vowels are never stressed, then the final vowel in words like café cannot be thematic, and must instead be part of the stem. A significant challenge for this approach is that we only know if a given vowel is thematic if it is not stressed. As the following pairs of words show, {e, a, o} can be thematic or not, depending on where stress falls: cón-e ‘cone’ versus café ‘coffee’; cóbr-a ‘snake’ versus sofá ‘sofa’; gát-o ‘cat’ versus robó ‘robot’ (see critique in Garcia 2017b). In addition, no mechanism (e.g., extrametricality) is available to capture words with antepenultimate stress, which predicts that no generalizations hold for words of this shape. However, Garcia’s (2017b) examination of the Portuguese lexicon suggests otherwise: antepenultimate stress is statistically more likely to occur in LLL words than in HLL words.⁴

This observation follows if Portuguese prioritizes the building of moraic trochees and if final syllables are extrametrical in words with antepenultimate stress (e.g., Bisol, 1992). On this view, ĹLL should be preferred to H́LL because in the latter word shape, both possible parses are non-optimal: the medial syllable could be left unparsed, as in (2a), but this results in a cross-linguistically marked metrical configuration; alternatively, an uneven trochee could be built, as in (2b), but this yields a typologically disfavored foot type (Hayes, 1985; Prince, 1990). In contrast, ĹLL words face neither of these problems, as shown in (2c), which could thus explain why words of this shape are more frequent than H́LL words in the Portuguese lexicon.

There are, of course, rankings that can generate (2a) or (2b) as optimal relative to (2c) if other metrical constraints are factored in, notably Stress-to-Weight (stressed syllables are heavy). In Chimalapa Zoque, stressed syllables undergo trochaic lengthening yielding (H́L) from (ĹL) (Hyde, 2011), indicating that Stress-to-Weight is dominant. Uneven trochees are also optimal in tone languages like Mandarin, where duration is the main correlate of stress because pitch is earmarked for lexical contrasts (Qu, 2013). However, in most trochaic languages, as Portuguese is described here, Stress-to-Weight does not play a determining role, in contrast to the role that it typically plays in iambic languages.

Moraic trochees have been posited to regulate other patterns in Portuguese as well. Nevins (2012), for example, proposes this for patterns of vowel lenition. In a word such as moléque /molɛke/ ‘boy’, lenition occurs post-tonically ([molɛ́ki]) or both pre- and post-tonically ([mulɛ́ki]), but not just pre-tonically (*[mulɛ́ke]). Nevins explains the asymmetry in question with foot structure: the post-tonic syllable is in the weak position of a trochee, whereas the pre-tonic syllable is unfooted. Obligatory post-tonic vowel lenition would therefore serve the purpose of augmenting the contrast between the strong and weak positions of a foot. Unfortunately, all of the words considered in Nevins’s study are trisyllables with penultimate stress. Crucially, Portuguese lenites vowels word-finally, regardless of whether stress falls on the penult (with the final syllable in foot-dependent position) or on the antepenult (with the final syllable outside the foot). Thus, we can conclude that, instead of lenition being regulated by foot structure, it occurs because the affected syllable is word-final, a position often associated with weakening across languages (see, e.g., Gordon, 2016). No conclusion regarding the status of the foot in Portuguese can therefore be drawn from the pattern in question.

Returning to stress, we have seen that proposals for Portuguese differ in how much emphasis is placed on the foot and on the type or types of feet that the language employs. Stem-based approaches assume that the foot plays, at best, a minor role. In other approaches, the foot plays a definitive role, but no general consensus about foot type emerges across proposals. Indeed, a particularly worrisome finding is that more than one foot type has been proposed to be simultaneously active in the language, either different types of trochees (Bisol, 1992; Wetzels, 2007), or different types of feet, i.e., trochees and iambs (Bonilha, 2004; Lee, 2007; Wetzels, 1992). Although the preference for ĹLL over H́LL observed in the lexicon may appear to support moraic trochees over other foot types, as mentioned, we will see that this finding is not mirrored in experimentally collected data. The lack of consistency within and across proposals stems from conflicting patterns in the language, which are more intricate than what has traditionally been assumed. We return to the question of footing in Portuguese in Section 5.1. We first examine stress in English non-verbs.

2.2. English

As already mentioned, the stress systems of Portuguese and English are, on the face of it, very similar. Unlike for Portuguese, however, there has not been much debate as to which metrical structures best characterize the English system (see Chomsky & Halle, 1968; Halle & Idsardi, 1995; Halle & Vergnaud, 1987b; Hayes, 1982; Liberman & Prince, 1977; Selkirk, 1980 among others). In English non-verbs, stress falls on the penultimate syllable if that syllable is heavy, and on the antepenultimate syllable otherwise, as can be seen in (3b) and (3c). The examples in (3b) show that heavy penultimate syllables contain a coda consonant (veránda) or a long vowel (oppónent). Final stress in nouns and adjectives tends to be avoided (Giegerich, 2005, p. 185), but can be found in words ending in VV(C) syllables (Halle & Vergnaud, 1987a); see (3a).⁵

Given the behavior of word-final syllables shown in (3b) and (3c), extrametricality has played a central role in metrical accounts of English stress. Hayes (1982), for example, proposes that the final syllable in nouns is extrametrical, and is therefore invisible during stress assignment: ve(rán)⟨da⟩. For words with final stress, such as those in (3a), Hayes proposes a rule (Long Vowel Stressing) that assigns a foot to such syllables. We can thus summarize the stress algorithm in English non-verbs as follows: starting at the right edge of the word, stress is final if this syllable contains a VV(C) rhyme. Else, the penultimate syllable is checked. If a heavy syllable is found, stress is penultimate. Else, stress is antepenultimate. This algorithm can account for over 80% of the words in a subset of the Carnegie Mellon University Pronouncing Dictionary (Weide, 1993) containing 6,531 nouns and adjectives (excluding disyllables). Although this might not seem substantially higher than the 72% observed earlier for Portuguese (Section 2.1), unlike in Portuguese, the vast majority of exceptions in English involve words where extrametricality exceptionally does not hold (e.g., va(nílla)).

Because English is a weight-sensitive language, and because final stress is typically avoided in non-verbs, the foot type standardly assumed for the language is the moraic trochee. Indeed, processes such as Trochaic Shortening (e.g., Hayes 1995) support an analysis based on moraic trochees: the long vowel in roots like sane [seɪn] is shortened when the suffix -ity is attached, enabling an optimal parse of the resulting string, namely (2c) above: [(sǽnɪ)⟨ti⟩]. This highlights an important difference between English and Portuguese: there is wide agreement in the literature that moraic trochees, and not syllabic trochees (or iambs), capture stress in the former language but not in the latter.

2.3. Summary

In sum, we have established that English supports (a particular type of) footing, while the evidence from Portuguese is less clear. Although weight effects are robustly observed in both languages, a question that must be addressed is whether weight effects, in and of themselves, motivate footing and whether the same conclusion can be reached for both languages on this matter.

Both languages show that heavy syllables (of particular profiles) attract stress in final and penultimate positions. Concerning antepenultimate position, we saw earlier (in (2c)) that if weight regulates footing, there should be a preference for light syllables in this position. Seemingly consistent with this, a preference for ĹLL over H́LL is observed in the Portuguese lexicon (Garcia, 2017b). Trochaic Shortening in English similarly favors ĹLL over H́LL. Although such patterns may appear to provide solid support for moraic trochees in both languages, Garcia (2019) shows that the frequency effect in the Portuguese lexicon is not upheld in the judgments of native speakers. In light of this finding, it is essential that we test the judgments of English speakers as well to verify whether the lexical patterns in English are reflected in speakers’ grammars.

In the next sections, we experimentally examine whether weight effects in antepenultimate position are observed for native speakers of English as well as Portuguese, and if so, what profile they take. Evidence that footing is sensitive to weight would come either from a negative weight effect in antepenultimate position, where stressed light syllables are preferred over stressed heavy syllables (ĹLL ≻ H́LL), as seen earlier in (2c), or from a null weight effect in said position (ĹLL = H́LL). In contrast, a positive weight effect in antepenultimate position, where speakers generalize the effects of weight from penultimate and final positions and thereby prefer stressed heavy syllables over stressed light syllables (H́LL ≻ ĹLL), would challenge the role of the foot in regulating stress; this is because the parse for H́LL would either involve a medial unfooted syllable or an uneven trochee, as seen earlier in (2a) and (2b), respectively. Indeed, Weight-to-Stress (heavy syllables are stressed), a constraint that does not require footing, could capture positive weight effects in all positions in the stress window.

4.1. Portuguese results

Figure 1 presents the main results from the Portuguese experiment. Overall, we can see that the preference for antepenultimate stress is below 50% (y-axis) for all three weight profiles. This isn’t surprising—recall from Section 2.1 that antepenultimate stress is a minor pattern in the language. Crucially, though, antepenultimate stress is preferred more often in HLL words (H́LL vs. HĹL) than in LLL words (ĹLL vs. LĹL). This suggests a positive weight effect in antepenultimate syllables, which we model below. Finally, LHL words have the lowest rate of preference for antepenultimate stress, which is what we would expect considering that ĹHL words are extremely rare in the language.

Figure 1

Overall weight effects on Portuguese stress: Percentages of preference for antepenultimate stress as a function of weight profile.

To statistically model the data shown in Figure 1, a Bayesian hierarchical logistic regression was run that included weight profile as a main effect and as a random effect in the form of by-participant random slopes. LLL was used as the reference level for weight, and is represented by the intercept of the model. Accordingly, we interpret the effect size of HLL and LHL in relation to LLL (our baseline). The model also includes by-word random intercepts—we used the default priors provided in the brms package (Bürkner, 2021), which means that no bias regarding the relationship between weight and stress preference is included in the model. Both effective sample sizes (ESS) and R̂ were checked to confirm convergence of the four chains used in the model. The model was run using Stan (Carpenter et al., 2017) through the brms package in R.

The model estimates are shown in Figure 2, which also displays the region of practical equivalence (ROPE) for the parameters in question—LLL*, with an asterisk, represents the intercept. Simply put, if the entire posterior distribution is found within said region, we can statistically conclude that the effect is null (see, e.g., Kruschke, 2015)—this is clearly not the case here. Unlike frequentist approaches, which provide a single parameter estimate and the probability of the data given such an estimate, Bayesian approaches provide an entire distribution of credible parameter values. Thus, even though the mean $\widehat{\beta }$ s discussed below are the most probable parameter values,⁶ the actual effects of interest are represented by entire distributions of parameter values.

Figure 2

Parameter estimates and associated posterior distributions for Portuguese. The intercept is indicated with an asterisk. Credible intervals are shown in the figure and detailed on the right (mean [95% highest density interval, or HDI]). The percentage of each HDI that is contained within the ROPE (shaded area around 0.0) is shown on the left.

In Figure 2, we see that the 95% highest density interval (HDI) of the posterior distribution for HLL is entirely positive (i.e., it doesn’t include zero as a credible parameter value). The effect in question must be interpreted relative to the intercept (LLL), whose posterior distribution is entirely negative, confirming that antepenultimate stress is disfavored (compared to penultimate stress) in LLL words, consistent with Figure 1. The mean effect size in the posterior distribution for HLL is $\widehat{\beta }$ = 0.34 (shown to the right of the distribution in Figure 2 along with the 95% HDI). We therefore conclude that, relative to the reference level, HLL words display a statistically higher level of preference for antepenultimate stress (vs. penultimate stress), thus confirming a positive weight effect in antepenultimate position. Table 3 presents the same results along with R̂ and ESS values.⁷ R̂ is used to inspect the convergence of the model (an R̂ of 1 indicates the model has converged). ESS (n_eff) refers to the number of sampling steps assuming an uncorrelated chain.

The positive effect of weight observed in the results confirms the earlier findings of Garcia (2019). In his study, Garcia found effect sizes of $\widehat{\beta }$ = 0.26 and $\widehat{\beta }$ = 0.16 in his two experiments for HLL (the second experiment being a replication of the first). Thus, we not only replicate the results for HLL, but also capture a stronger effect size ( $\widehat{\beta }$ =0.34)—note that all three effect sizes come from the same experimental design and from the same stimuli. It is thus safe to conclude that the empirical evidence for weight effects in antepenultimate syllables in Portuguese is robust. We refer the reader to Kobrock & Roettger (2022) for important discussion on the scarcity of replication studies in experimental linguistics.

We consider finally certainty levels and reaction times. Certainty levels did not show any particular trends for antepenultimate stress preference in HLL versus LLL words in our data. There were, though, two trends that were observed overall: first, participants were more certain about penultimate stress regardless of weight profile, a finding that is not surprising given the high frequency of penultimate stress in the language. Second, participants were least certain when choosing antepenultimate stress in LHL words, which is again consistent with frequency effects, given the extremely low frequency of ĹHL words in the language. Reaction times similarly revealed no trends related to the crucial question examined. However, we observed faster reaction times overall when participants chose penultimate stress and, once again, ĹHL words can be singled out as they displayed the slowest reaction times in the data. We do not discuss these two variables further.⁸

4.2. English results

We turn now to the English experiment. We start by analyzing the main variable of interest, namely, weight. In Figure 3, we can see the mean percentage of preference for antepenultimate stress across LLL, HLL and LHL words. Standard errors from the mean are also provided. Preference for antepenultimate stress is above 50% for LLL and HLL words, but below 50% for LHL words.

Figure 3

Overall weight effects on English stress: Percentages of preference for antepenultimate stress as a function of weight profile.

As with the Portuguese data discussed earlier, to statistically model the data shown in Figure 3, a Bayesian hierarchical logistic regression was run with by-word random intercepts, as well as by-speaker random slopes (weight) and intercepts. Like in the Portuguese model, LLL was the reference level for weight and default priors were used. The model’s estimates confirm what is observed in Figure 3, namely, that relative to LLL words, LHL words affect speakers’ responses, reducing the probability of preference for antepenultimate stress. HLL words have a slightly negative effect on speakers’ preference for antepenultimate stress, but the posterior distribution is too close to zero for us to confidently conclude that antepenultimate heavy syllables have an effect—as shown in Figure 4, zero is not only included in the 95% credible interval, but also in the 50% credible interval of the posterior distribution.

Figure 4

Parameter estimates and associated posterior distributions for English. The intercept is indicated with an asterisk. Credible intervals are shown in the figure and detailed on the right (mean [95% highest density interval, or HDI]). The percentage of each HDI that is contained within the ROPE (shaded area around 0.0) is shown on the left.

Table 4 lists the mean estimates (Mean $\widehat{\beta }$ ) as well as the 95% credible intervals in the posterior distributions estimated. For example, the mean $\widehat{\beta }$ for the intercept (0.55) represents the mean of the posterior distribution for this parameter (in log-odds)—the positive intercept captures the preference for antepenultimate stress in LLL words, as observed in Figure 3. The 95% credible interval in question goes from 0.28 (2.5%) to 0.81 (97.5%).

As we can see from the table, the mean estimate for HLL is negative (–0.09). The 95% credible interval of its posterior distribution ranges from –0.43 to 0.25. As a result, the distribution includes zero as a credible parameter value. Given that our reference level is LLL, this means that the effect of HLL words is mostly negative relative to LLL words, but that we cannot conclude that these two weight profiles have reliably different effects on the basis of our model. The results for English thus show no credible effect of weight in antepenultimate syllables—indeed, 68% of the HDI of HLL is located within the ROPE in Figure 4.

On the other hand, the posterior distribution of LHL excludes zero, and has a mean of –0.97. This is the log-odds of antepenultimate stress given a LHL word (relative to a LLL word). Simply put, a LHL word lowers the odds of antepenultimate stress by a factor of 2.6 ( $(e^{\left|\widehat{\beta }\right|})$ . If we again pick mean( $\widehat{\beta }$ ) as a single point estimate, the overall probability of choosing antepenultimate (over penultimate) stress in LHL words is 27% (compared to 63% for LLL words).

We point out finally that, as with the Portuguese data discussed earlier, certainty levels and reaction times did not show any particular trends for antepenultimate stress preference in HLL versus LLL words in our English data.

4.3. Language comparison

In this section, we compare the main findings for the two languages. To facilitate this, we combine both languages into a single figure of posterior distributions of effect sizes. Figure 5 overlays the two central models discussed thus far (Portuguese is shown in white; English is shown in violet). Both LLL* and HLL effects are included, given that we must interpret the HLL distributions relative to those of LLL* in each language.

Figure 5

Posterior distributions of HLL: Portuguese (white) versus English (violet). Portuguese displays a clear positive posterior distribution for HLL (relative to LLL*); English displays a distribution centered around zero.

First, we point out that the LLL* distributions are on opposite sides of the null effect line ( $\widehat{\beta }$ = 0). This is consistent with the patterns in each language: antepenultimate stress is much less common than penultimate stress in Portuguese, but much more common in English.⁹ Second and more crucially, the distributions of HLL are fundamentally different: positive in Portuguese, but centered around zero in English. Note that we cannot directly compare the HLL distributions in both languages in the figure, since the effects are always relative to the posterior distribution of the intercept, LLL*.

4.4. Sonority effects

We have seen that the results for Portuguese and English parallel each other insofar as both languages display a positive weight effect for penultimate position. The languages differ for antepenultimate position: Portuguese exhibits a positive effect for weight in this position while English does not, as shown in Figure 5. In this section, we show that the difference between the two languages extends to sonority effects, which can be detected in HLL words in Portuguese but not in English.

4.4.1. Sonority effects in Portuguese

In the Portuguese experiment, we find that type of coda (sonorant vs. obstruent) in antepenultimate position has an effect on speakers’ preferences for antepenultimate stress. This is shown in Figure 6, which plots the preference for antepenultimate stress (y-axis) by weight profile and coda type—means and bootstrapped confidence intervals are provided.

Figure 6

Effect of coda type on stress preference in Portuguese.

To confirm the effect of coda type observed in Figure 6, a model was run with main effects as well as the interaction of weight and coda type (sonorant or obstruent). The model includes by-speaker random slopes for sonority and its interaction with weight, and by-item random intercepts. LLL words, where no coda effect can be assessed, were excluded. The model is shown in Table 5. We can see a statistically credible effect of coda (son): $\widehat{\beta }$ = 0.40. This effect refers to HLL words (the reference sonority level in the model), and confirms what we observe in Figure 6. Of interest, the sonority effect we detected remains if we exclude stimuli with coda /l/, which is vocalized to [w]: $\widehat{\beta }$ = 0.42. Thus, if participants were interpreting vowel + /l/ sequences as diphthongs, this possibility would not change the overall sonority effect found in the data.

Table 5

Sonority model for Portuguese. Mean parameter estimates and associated estimated errors, credible intervals, R̂, and ESS (n_eff).

Parameter	Mean  $\widehat{\mathit{\text{β}}}$	Est Error	2.5%	97.5%	R̂	neff
Intercept (HLL, obs)	–0.68	0.27	–1.20	–0.16	1.0	2434
LHL	–0.60	0.29	–1.15	–0.03	1.0	2757
coda (son)	0.40	0.21	0.00	0.82	1.0	2826
LHL:coda (son)	–0.31	0.31	–0.93	0.30	1.0	2758
model: stress ~ weight * coda + (weight*coda|participant) + (1|word)

4.4.2. Sonority effects in English

Given the sonority effect observed for Portuguese, we must question whether a similar effect might be found in English. Any effect in the latter language would likely be relatively smaller, given the absence of a positive correlation between syllable weight and antepenultimate stress.

Parallel to Figure 6 for Portuguese, Figure 7 plots the preference for antepenultimate stress (y-axis) by weight profile and coda type for English—means and bootstrapped confidence intervals are provided. As already seen in Figure 3, antepenultimate stress is not statistically affected by the weight of an antepenultimate syllable. Here, we see that the sonority value of antepenultimate (and penultimate) codas does not appear to affect the preference for antepenultimate stress in the data.

To model the effect of coda type in HLL and LHL words in English, a model was run with main effects as well as the interaction of weight and coda type (sonorant or obstruent). This is the same model that was used to examine sonority in the Portuguese data above. Thus, the model excludes LLL words, where no coda effect can be assessed. In addition, by-speaker random slopes for the interaction and random intercepts, as well as by-word random intercepts, were added.

Figure 7

Effect of coda type on stress preference in English.

The results are shown in Table 6, and can be summarized as follows: for HLL words with a sonorant coda (coda (son) in the table), we have a posterior distribution with mean $\widehat{\beta }$ = 0.30, although $\widehat{\beta }$ = 0 is included in the 95% credible interval of the distribution, which is consistent with what we see in Figure 7. Only one effect is statistically credible given all the 95% credible intervals in question (LHL), which is neither surprising (given what we have seen thus far) nor relevant to our discussion: LHL words with an obstruent coda reduce the probability of preferring antepenultimate stress relative to HLL words with an obstruent coda.

Table 6

Sonority model for English. Mean parameter estimates and associated estimated errors, credible intervals, R̂, and ESS (n_eff).

Parameter	Mean  $\widehat{\mathit{\text{β}}}$	Est Error	2.5%	97.5%	R̂	neff
Intercept (HLL, obs)	0.30	0.20	–0.09	0.69	1.0	2354
LHL	–0.66	0.25	–1.14	–0.17	1.0	2383
coda (son)	0.30	0.19	–0.08	0.68	1.0	2737
LHL:coda (son)	–0.44	0.28	–0.99	0.09	1.0	2563
model: stress ~ weight * coda + (weight*coda|speaker) + (1|word)

4.5. Predictions revisited

In sum, we have seen that weight effects are present in antepenultimate syllables in Portuguese, but not in English. We have also detected a sonority effect in antepenultimate codas in Portuguese, but not in English, which provides a more nuanced picture of weight effects in this position. Although sonority effects are not the main object of examination in the present paper, these findings contribute to the discussion on sonority-driven stress, which has typically focused on sonority (height) in vowels (e.g., McCollum, 2020; Shih, 2015; Shih & de Lacy, 2019). We will consider the implications of sonority more concretely in the discussion (Section 6).

Given the cross-language difference concerning weight effects in antepenultimate position, we address the main results discussed above in light of our predictions (see Section 3.4). Recall that English is a language for which there is strong evidence for moraic trochaic footing (Section 2.2). On the basis of this, we predicted that English would either show a negative weight effect in antepenultimate position, where stressed light syllables are preferred over stressed heavy syllables (ĹLL ≻ H́LL), or a null weight effect in said position (ĹLL = H́LL). Consistent with this prediction, we experimentally confirmed a neutral weight effect in antepenultimate position for English.

In contrast, following earlier work by Garcia (2019), we predicted that Portuguese would show a positive weight effect in antepenultimate position, where speakers generalize the effects of weight from penultimate and final positions and thereby prefer stressed heavy syllables over stressed light syllables (H́LL ≻ ĹLL). We suggested that such a result would challenge the role of the (moraic trochaic) foot in regulating stress in Portuguese because there is no optimal parse that prioritizes foot well-formedness for H́LL strings. Recall in this context that earlier research does not agree on the type of feet that Portuguese builds (Section 2.1). Consistent with the stated prediction, we experimentally confirmed a positive weight effect in antepenultimate position for Portuguese.

5.1. Portuguese

5.2. English

The experimental results summarized above for Portuguese contrast with what we observed for English, namely, a neutral effect for weight in antepenultimate position. The latter finding is consistent with English building feet, a proposal that has not, to our knowledge, been challenged in the literature on stress in this language. As the cross-linguistic difference we have experimentally demonstrated may reflect a different status for the foot in the two languages, we must determine whether English shows minimal word, truncation, and hypocorization effects distinct from those in Portuguese, specifically, patterns that are consistent with the language building binary moraic trochees. In the following section, we show that this is indeed the case.