2.1. Weight, prefixation, length, and analogy

The first potential factor to be considered here is syllable weight. Standard accounts of English assume that stress assignment is determined in large part by properties of a syllable´s rhyme – the vowel and any following consonants (Chomsky & Halle Reference Chomsky and Halle1968, Hayes Reference Hayes1982). These units are seen as contributing syllable weight, a property that is relevant for stress in English (e.g. Carr Reference Carr2009). Light syllables are those with a single short vowel in the rhyme (e.g. [ɛ]). Heavy syllables, on the other hand, have a long or diphthongized vowel (e.g. [ɑː]), or a single short vowel followed by one single coda consonant (e.g. [ɪp]). Syllables that contain a long or diphthongized vowel followed by consonant (e.g. [ɑːm]) or a short vowel followed by a consonant cluster (e.g. [ənt]) are termed superheavy (e.g. Hayes, 1989). Verb-final syllables are claimed to exhibit a special behavior in this context, which is typically subsumed under the notion of extrametricality (Hayes Reference Hayes1982): most analyses consider the final consonant of verbs to be invisible to the stress-assigning mechanism. For English verbs, the claim is therefore that final syllables which contain a short vowel may be heavy only if they are followed by at least two consonants. Final syllables that contain a long vowel or diphthong, by contrast, count as heavy regardless of whether or not they are closed.

There are two possible ways in which syllable weight can influence stress assignment. One possibility is that stress placement depends on the structure of one particular syllable. In the case of English verbs, the structure of the final syllable is assumed to be decisive (see, for example, Burzio Reference Burzio1994, Hammond Reference Hammond1999, Hayes Reference Hayes1982). Specifically, the prediction is that the final syllable will be stressed if it is heavy (e.g. tormént, appéar). In this case, the remaining syllables may contain an unstressed full vowel, a reduced vowel, or, alternatively, they may have a secondary stress (Burzio Reference Burzio1994, Guion et al. Reference Guion, Clark, Harada and Wayland2003: 404). Otherwise, main stress is claimed to fall on the penult (as in édit or vómit). Moreover, weight interacts with word length: the patterns described so far hold largely only for short (i.e. disyllabic) verbs. Trisyllabic verbs with a heavy final syllable are claimed to be regularly stressed on the antepenult, except if their final syllable is a complete historical root (for example, éxercise, ánalyze, cómplicate, clárify vs. comprehénd, interséct, introdúce; see the formatives in Chomsky & Halle Reference Chomsky and Halle1968, Liberman & Prince Reference Liberman and Prince1977). Another possibility is that heavy syllables tend to attract stress (which is usually formalized in OT approaches as the “Weight-to-Stress” constraint), irrespective of their position within the word. In other words, the idea is that a prominent syllable will be preferentially stressed even in cases where it does not conform to the regular metrical pattern and/or is predicted by the pertinent rules of the language. Commonly, prominent syllables are those with long or low vowels or sonorant codas (Hayes Reference Hayes1995, Guion et al. Reference Guion, Clark, Harada and Wayland2003 for a discussion of typological evidence).

Beyond weight and length, stress may also be determined by a verb’s morphological structure. One case in point is prefixes. For English, prefixes are said to have stress-repelling properties (e.g. Fudge Reference Fudge1984). In other words, the expectation is that main stress never falls on the prefix, even if stress in this position would be predicted by competing factors such as syllable weight (e.g. *béset, Raffelsiefen Reference Raffelsiefen, Booij and van Marle1999). However, there is an ongoing debate about how to divide the lexicon into words whose stress is affected by morphology and those whose stress is not affected by morphology. This is in large part because there continues to be disagreement regarding the nature of morphological effects, and theories differ vastly in terms of how morphology is defined. For example, while reference to opaque morphological units has been made in the early generative literature (e.g. Chomsky & Halle Reference Chomsky and Halle1968), such words have come to be treated as monomorphemic words in subsequent work (e.g. Burzio Reference Burzio1994). This position has been challenged by work in the Guierrian tradition (e.g. Guierre Reference Guierre1979, Fournier Reference Fournier2007, Dabouis & Fournier Reference Dabouis, Fournier, Ballier, Jean-Michel Fournier and Yamada2023). One prominent example that is often cited in this context is historically prefixed verbs such as beget, converge, or relegate, which involve some combination of an opaque prefix (mostly historical prefixes such as con- that no longer are productive in present-day English) and a bound root (e.g. - verge). From a synchronic point of view, there is no productive morphology involved in these words, which is why they have often been treated on a par with genuinely monomorphemic verbs such as vomit. Specifically, opaque prefixes have often been excluded from phonological analyses on the grounds that they do not have a clear meaning and therefore defy the standard definition of the morpheme as a fully productive form-meaning pairing (e.g. Hammond Reference Hammond1999: 71). That said, work within the Guierrian school has revealed that historical prefixes such as be- or con- do pattern with fully transparent prefixes on a number of phonological characteristics, including vowel reduction in the initial pretonic position in closed syllables (Dabouis & Fournier Reference Dabouis and Jean-Michel2025a). With regard to primary stress placement, Dabouis and Fournier (Reference Dabouis, Fournier, Ballier, Jean-Michel Fournier and Yamada2023) report that initial stress is strongly dispreferred in disyllabic verbs that contain an opaque prefix (e.g. bewítch, submít). Crucially, in their corpus, effects of opaque prefixation were shown to override those of weight in 74 percent of prefixed verbs which have final stress despite their light final syllables (e.g. asséss, *ássess; condémn, *cóndemn; detách, *détach). The patterns reported in these studies have led the authors to advocate a less restrictive approach to the phonology-morphology interface that allows reference to semantically opaque or underspecified constituents.

Again, like syllable weight, opaque prefixation interacts with length: looking at the stress distributions in verbs with a bound base, Dabouis and Fournier (Reference Dabouis, Fournier, Ballier, Jean-Michel Fournier and Yamada2023) report that stress on an opaque prefix is more readily tolerated in long verbs than in disyllables (compare, for example, refér vs. récognize, cómplicate, démonstrate, súbjugate). This is particularly true of suffixed verbs, especially -ate, which make up a large part of long verbs in the lexicon.

There is, however, considerable leakage with regard to the generalization that prefixes repel stress, no matter how morphology is defined. An alternative is to take a processing-based perspective on stress assignment. To illustrate this, let us return to the example of English prefixes. Under a processing-based approach, whether or not the initial syllable in a (new) word receives the main stress might be determined not so much by its abstract morphological status (i.e. is it a prefix or not?) but by the extent to which it is recognized as such. On this view, individual prefixes differ in terms of how well they are recognized and parsed out in a given word by speakers, regardless of their synchronic status in the system, with potential implications for how the word is stressed. This approach is in line with earlier psycholinguistic work showing that, indeed, there is gradience in how language users process prefixed words in English (cf. esp. Wurm Reference Wurm2000). The degree of recognizability of a prefix can hence be quantified with the help of simple prefix ratios that take into account all words that start in a given string (termed “prefix likelihood” by Wurm; see also methodology section further below). Originally developed as a gradient measure for modeling morphological effects in the auditory processing of polymorphemic words, prefix likelihood ratios were shown to influence pseudo-word rejection times in a lexical decision experiment (Wurm Reference Wurm1997, Reference Wurm2000). Applying these measures to predict stress position in English pseudo-verbs, one might hypothesize that dis-, for example, should be relatively more likely to be recognized as a prefix in word-initial position compared to, say, re-, despite both prefixes being highly productive in present-day English – simply because words starting in <dis-> tend to be almost invariably prefixed, whereas there are relatively more words that start in <re-> where re- is not a prefix (as in real). Consequently, one should also expect relatively more word-initial stress in <re->-initial words compared to words starting in <dis->.

The continuing relevance of historical prefixes for synchronic stress assignment raises another, more fundamental question regarding the nature of the alleged prefix effects: to what extent are effects that have hitherto been interpreted as morphological in nature based on pure form, that is, recurrence of phonological strings – and if so, what is the role of other types of recurrence in the lexicon that are not specific to the word beginning, such as (bound) roots or suffixes? These might then be best described as an analogical effect. In this paper, we will discuss analogy as a process by which new words are stressed by analogy with existing, phonologically similar words in speakers’ Mental Lexica. While appeal to analogy has generally had a long tradition especially in the field of morphology, it has often been criticized for a lack of transparency and restrictions (see Bauer Reference Bauer2001: 75–97 for an overview of the limitations of (traditional) analogical approaches). Specifically, it was often left unclear exactly when and under what circumstances alleged analogies take effect, raising questions about the types of features that become relevant in a given context (e.g. semantic? morphological? phonological?), as well as the number of lexemes that are involved in the analogy. Computational analogical models address this criticism by systematically operationalizing notions such as “similarity” and “sets of relevant exemplars.” Importantly, they generate falsifiable hypotheses that can be tested in various ways, for example, in experiments or on corpus data. In this paper we will use the AML algorithm (Skousen Reference Skousen1989). The specifics of this model will be introduced in the methodology section.

Analogy-based approaches are part of a larger group of approaches – exemplar-based approaches – in which large-scale storage of individual occurrences of exemplars plays a prominent role (e.g. Bybee Reference Bybee2010). A key feature of exemplar-based approaches to linguistic categorization is that they reject the common distinction between regular and irregular items (Arndt-Lappe Reference Arndt-Lappe, Müller, Ohnheiser, Olsen and Rainer2015: 827). Instead, the claim is that all patterns emerge bottom-up from comparison over similar exemplars that are stored along with their properties in the lexicon, including, for example, their stress pattern. Consequently, linguistic patterns on this view are necessarily emergent and non-categorical: since classification of new input depends on the stored properties of pertinent exemplars (from now on referred to as “analogues”), the result may be different for each new target. Importantly, analogies are not necessarily always based on just one single lexeme: much more frequently analogues involve whole sets of words in the lexicon (commonly referred to as “analogical sets”). Behavior that is commonly described as “regular” or “rule-based” in other approaches arises exactly in those situations where analogues are particularly numerous. In this regard, analogy-based approaches offer one important theoretical advantage over deterministic approaches: both highly productive and more local, idiosyncratic patterns can be accounted for by one single mechanism.

However, the question of how lexical representations are best conceptualized in an analogy-based framework is still being fiercely debated in the literature. While in some analogical approaches it is assumed that similarity relations between words are largely based on symbolic features (e.g. Becker Reference Becker1990), others contend that analogical reasoning also involves recourse to more detailed and specific information than the symbolic features typically associated with rule-based grammars (e.g. Krott, Baayen & Schreuder Reference Krott, Baayen and Schreuder2001, Krott, Schreuder & Baayen Reference Krott, Schreuder, Baayen, Skousen, Lonsdale and Parkinson2002). In fact, there is a growing body of evidence that supports the relevance of such less abstract representations. Evidence of this kind mainly comes from studies on derivational morphology (e.g. Gagné Reference Gagné2001, Krott, Schreuder & Baayen Reference Krott, Schreuder, Baayen, Skousen, Lonsdale and Parkinson2002, Chapman & Skousen Reference Chapman and Skousen2005), but also from compound stress assignment (Plag, Kunter & Lappe Reference Plag, Kunter and Lappe2007, Arndt-Lappe Reference Arndt-Lappe2011). These studies have typically used features encoding the phonological (i.e. segmental and prosodic) shape of words as a basis for the computation of similarity (e.g. Eddington Reference Eddington2000 on stress assignment in Spanish; Chapman & Skousen Reference Chapman and Skousen2005, Arndt-Lappe Reference Arndt-Lappe2011), sometimes alongside features encoding relevant higher-order information (such as tense in Eddington’s study or the semantic relation between compound constituents in Arndt-Lappe’s study). While it is still an open question as to which types of features result in the best predictive power of a model, a common finding in these studies is that phonotactics-based models not only prove to be highly successful in modeling the pertinent phenomena, but sometimes even outperform traditional rule-based models that implement the structural cues held to be responsible in the literature (e.g. as in Arndt-Lappe’s study on compound stress in English).

This paper explores whether measures derived from an AML implementation are predictive of main stress position in English trisyllabic pseudo-verbs. In order to better understand the relation between traditional structural variables (such as prefixation and syllable weight) and AML measures, we will also compare models that use measures derived from the AML model with models that use the traditional structural variables to predict stress position. This will allow us to assess whether AML measures render the specification of morphological (i.e. prefix versus no prefix) and phonological structure (i.e. the weight configuration of all syllables in the word) dispensable for the prediction of stress position in English verbs.

2.2. Prior research and research questions

A general shortage of empirical studies is a notorious problem in many areas of phonological research. However, the last two decades have seen a surge of studies that seek to assess proposed generalizations against larger datasets, or novel words. Investigating stress placement in disyllabic pseudo-words, Guion et al. (Reference Guion, Clark, Harada and Wayland2003), for example, found that the position of stress was affected by the syllabic structure of both syllables, the syntactic category (i.e. noun vs. verb) of the new word, as well as the stress patterns of phonologically similar words. Specifically, nouns were less likely to take final stress than verbs, heavy syllables tended to attract stress, and the stress distribution among phonologically related words (elicited retrospectively from participants) was significantly related to the choice of stress position in the new words. While their study has been the only one so far to discuss the role of analogy in primary stress assignment to English verbs, effects of syllable weight (Ryan Reference Ryan2014, Turcsan & Herment Reference Turcsan, Herment, Mompean and Fouz-González2015, Ktori, Mousikou & Rastle Reference Ktori, Mousikou and Rastle2018, Dabouis & Fournier Reference Dabouis, Fournier, Ballier, Jean-Michel Fournier and Yamada2023) and syntactic category (Ktori, Mousikou & Rastle Reference Ktori, Mousikou and Rastle2018, Treiman et al. Reference Treiman, Rosales, Cusner and Kessler2020) have been replicated numerous times for English disyllables in both pseudo-word and corpus data. What is more, some studies have also demonstrated effects of prefixation (Ktori et al. Reference Ktori, Tree, Mousikou, Coltheart and Rastle2016, Treiman et al. Reference Treiman, Rosales, Cusner and Kessler2020, Dabouis & Fournier Reference Dabouis, Fournier, Ballier, Jean-Michel Fournier and Yamada2023), orthographic weight (Ktori, Mousikou & Rastle Reference Ktori, Mousikou and Rastle2018), and even onset weight in word-initial position (Kelly Reference Kelly2004, Ryan Reference Ryan2014, Treiman et al. Reference Treiman, Rosales, Cusner and Kessler2020), which contradicts the widely-held view that only syllable rimes contribute to weight. Interestingly, investigating the role of prefixation in English short (i.e. disyllabic) versus long verbs, Dabouis and Fournier (Reference Dabouis, Fournier, Ballier, Jean-Michel Fournier and Yamada2023) report that the stress-repelling effect of transparent prefixes extends also to opaque (i.e. semantically underspecified) prefixes. In long verbs the situation is slightly more complex. In trisyllables with -ate, the prefix is almost categorically stressed, whereas in non-suffixed trisyllabic verbs prefixes usually do not receive primary stress, just as in disyllables, even if less systematically so than in disyllables. When it comes to long words, however, the evidence is more sparse. Ernestus and Neijt (Reference Ernestus and Neijt2008) compare dictionary data to stress judgments elicited in a series of paper-and-pencil experiments but focus on the very specific question of the location of primary stress as affected by the syllables preceding the final three-syllable window. Domahs et al. (Reference Domahs, Plag and Carroll2014) and Moore-Cantwell (Reference Moore-Cantwell2020) both investigate the role of syllable weight in stress assignment in trisyllabic pseudo-words. While Cantwell finds an effect of the quality of the final nucleus (such that a final [i] pushes main stress leftwards to the antepenult, but not other vowels), Domahs and colleagues report that trisyllables with a heavy final syllable more generally show a strong preference for word-initial stress. While Moore-Cantwell finds no effects of syntactic category, Domahs et al. only look at nouns.

Turning to the computational modeling literature, analogical effects on stress assignment have not been discussed much. Pertinent studies have been conducted in Dutch (Daelemans, Gillis & Durieux Reference Daelemans, Gillis and Durieux1994, using TiMBL) and Spanish (Eddington Reference Eddington2000, with AML). In these studies, computational analogical models have been successfully employed to model stress assignment based on the phonemic make-up of the three word-final syllables as well as rudimentary grammatical information such as tense (in Eddington’s study). For English, Arndt-Lappe (Reference Arndt-Lappe2011) looked at stress assignment in English binominal compounds, comparing the performance of two analogical algorithms that were trained on different combinations of symbolic (categorical semantic and syntactic predictors) and low-level features (i.e. sounds and phonotactics). More recently, Arndt-Lappe et al. (Reference Arndt-Lappe, Schrecklinger and Tomaschek2023) and Plag et al. (Reference Plag, Arndt-Lappe, Seiler and Dabouis2026) modeled stress assignment in English using Naive Discriminative Learning (e.g. Baayen et al. Reference Baayen, Milin, Durdević, Hendrix and Marelli2011), with only orthographic bi- and trigram sequences as predictors of stress and no access to higher-level information about the morphological or phonological structure of words. They concluded that “neither directionality nor stratification need to be stipulated as a-priori properties of words or constraints in the lexicon” (Arndt-Lappe et al. Reference Arndt-Lappe, Schrecklinger and Tomaschek2023: 1).

Taken together, these studies suggest that the categories proposed in the literature – syllable weight, morphology, syntactic category and length – do matter when it comes to predicting word-level stress in English. That said, this brief review also suggests that these predictors do not apply in a categorical fashion. Rather, the situation is better characterized as a complex multitude of competing factors that probabilistically affect stress assignment (see Treiman et al. (Reference Treiman, Rosales, Cusner and Kessler2020) for a similar conclusion). The overall picture that emerges here is one where speakers know about statistical patterns in the lexicon, and they consider these patterns when making decisions about where to place the main accent in a (new) word. This begs the question of what exactly these patterns look like and how abstract they are. One possibility is that they are based on speakers´ knowledge of pertinent stress distributions in the lexicon, for example, across different vowel types and different morphological categories. At the same time, these effects might also be taken as a reflex of more fine-grained information that is being used during stress computation, in the form of specific segmental similarities (phonetics, orthography) between exemplars in the lexicon. These might then be best described as an analogical or exemplar-based effect. Assuming there is an intrinsic link between weight and stress in the lexicon (and between prefixation and stress), effects that have mostly been discussed under the header of “morphology” or “phonology” should emerge as a very natural consequence of the cooccurrence of different linguistic properties in the exemplars on which analogies are based. So far, however, the evidence for the use of such less abstract patterns in English stress assignment is limited (Guion et al. Reference Guion, Clark, Harada and Wayland2003, Plag et al. Reference Plag, Arndt-Lappe, Seiler and Dabouis2026).

In sum, we still need studies that systematically test the empirical adequacy of existing generalizations in long verbs. Besides this, we will test if analogy – the patterning of stress across segmentally similar forms in the lexicon – adds any explanatory power to the prediction of stress position above and beyond the structural factors studied here. This leads us to the following research questions for the present study:

3.1. Participants

Fifty native speakers of British English ranging in age from 19 to 53 years took part in the research study (mean age: 29.02). They all were raised monolingual with English as their first language. The study was distributed evenly among male and female participants. Participants were recruited via the online platform Prolific. None of the participants reported any speech or hearing impairments. The data was collected and analyzed with respect to the rights of research participants regarding data protection and confidentiality. Informed consent was obtained from all participants prior to the start of the experiment. On average, the experiment lasted approximately 30 minutes, and participants received a compensation of £4.50.

3.2. Design and stimuli

The purpose of the production task was to assess the potential impact of three factors on stress assignment in the production of trisyllabic pseudoverbs. More specifically, the roles of the factors syllable structure, morphological complexity, and analogy were investigated.

Given their three-syllable window, trisyllabic words provide an excellent test case for analyzing extant claims regarding stress placement in English. Heavy (CVC) and light (CV) syllables were combined in three different constellations within the word to investigate the possible interaction between syllable weight and stress position during stress assignment. Additionally, the morphological status of the pseudo-words was varied such that one-third of the verbs were monomorphemic while the other two-thirds were prefixed. As a first approximation, we started out with a broad classification into productive (e.g. re-) and unproductive prefixes (e.g. ob-), following the classification proposed by Bauer, Lieber, and Plag (Reference Bauer, Lieber and Plag2013). They distinguished these two prefix classes based on their combinatorial possibilities: Productive prefixes are ones that can combine with free bases (e.g. re- in refill) and do so very often, whereas unproductive prefixes generally occur only with bound bases (e.g. sub -mit, ob-struct). Like all other syllable types, the syllabic structure of the prefixes was varied to tease apart the possible impacts of phonological vs. morphological factors.

The trisyllabic pseudo-words used as stimuli in the current study had two different types of syllable structure. Following the weight classification proposed by Domahs, Plag, and Carroll (Reference Domahs, Plag and Carroll2014), open syllables (CV; e.g. ba) were coded as light and closed syllables (CVC; e.g. bat) as heavy when they occurred in word-initial or word-medial position. However, this classification does not consider weight caused by phonological length, i.e., tenseness of the vowel. Some accounts base their discussion of syllable weight on a binary distinction between short (“lax”) and long (“tense”) vowels, classifying short vowels as light and long vowels as heavy, with many long or tense vowels surfacing as diphthongs (e.g. Hayes, Reference Hayes1982). Given that participants only had access to orthographic forms, these were designed in such a way as to elicit short vowel pronunciations. Items for which participants produced diphthongs were excluded from the dataset. Determining whether monophthongs are phonologically long is less straightforward because stressed vowels tend to be phonetically longer than unstressed ones. In a reading experiment like ours, this is a problem that mainly concerns the letter <a>, which can be realized as both phonologically long and short vowels that are close in quality: /æ/vs /ɑː/ (and for some varieties /a/vs /aː/).

To assess whether there is indeed a confound, we perceptually coded vowel length in a random sample of 50 pseudo-words produced by different speakers. We focused on vowels in open syllables in either the antepenultimate or penultimate position that had been coded as stressed and were spelled with <a> (e.g. CAborasp, PAmoltin, subMAtin, conTAsit). With only one case in which coders were unsure, all vowels were realized as short. Overall, then, we refrained from making categorical decisions on the phonological status of monophthongs in our coding.

The weight classification adopted here differs from Domahs and colleagues in one important respect: due to the assumed extrametrical status of final consonants in English verbs, syllables that occurred in the word-final position were defined as heavy if they ended in two coda consonants (CVCC, e.g. - bast). Light final syllables, on the other hand, ended in a single consonant (CVC, e.g. - bat). To avoid syllabifications of consonants as onset consonants of the following syllable, the only word-internal clusters which were selected are clusters that cannot form a well-formed branching onset. The experimental items were pre-read by two native speakers of English (BrE, AE) to test their pronounceability and status as possible verbs. One concern raised by one of the test readers was that some items did not look like English verbs but rather like nouns. However, this should not pose any problems since all verbs were disambiguated as verbs by presenting them in a sentential context (unlike in the test-reading where the words were read from a list).

The morphological status manipulation was intended to ascertain the extent to which the presence of prefixes in pseudo-verbs affects their phonetic realization. An important point to be stressed here is that the current investigation of morphological structures was restricted to prefixed verbs only. Other morphological categories (e.g. compounds, suffixes) were excluded from the present study on grounds of their distinct behaviors (especially suffixed verbs – either with stress-shifting or stress-preserving suffixes – whose effects have been shown to override those of (opaque) prefixes; please refer to Dabouis and Fournier (Reference Dabouis, Fournier, Ballier, Jean-Michel Fournier and Yamada2023) for a comprehensive description of different morphological categories). Here, we distinguished three categories: (1) non-prefixed (i.e. monomorphemic) verbs, (2) verbs with productive prefixes (e.g. de- , re-), and (3) verbs containing prefixes that are no longer productive (e.g. be- , ob-). Both Latinate and Germanic prefixes were included in the study. A complete list of the stimuli for this experiment is provided in Table 1.

Table 1.

Examples of pseudo-verbs in each condition (L = Light, H = Heavy)

Table 1. Long description

This table presents the pseudo-verbs used in each experimental condition, cross-tabulated by weight and prefixation. Rows represent the morphological status of the initial syllable (nonprefixed, unproductive prefixed, productive prefixed) and colums represent the weight configuration (LLH, LHL, HLL). Conditions are numbered. LLH nonprefixed (1): caborasp, pagovast, rucomest. LHL nonprefixed (2): tufontap, bocelpat, pamoltin. HLL nonprefixed (3): pendatip, lencatin, bontilap. LLH prefixed unproductive (4): becarost, emanult, amanost. LHL prefixed unproductive (5): secolpam, ebantoc, befoltip. HLL prefixed unproductive (6): contasit, obserat. LLH prefixed productive (7): precomult, depadest, remigust. LHL prefixed productive (8): rebaltin, deguspat, prenunsip. HLL prefixed productive (9): unpavin, disrupit, mispalip, submatin.

There were nine experimental conditions. Two to three items were used in each condition, making a total of 27 test items. The conditions were created on the basis of structure combinations found in the lexicon, which will be explained below. Conditions are numbered in Table 1 for convenience.

The analyses presented here are restricted to syllable structure combinations involving only one heavy syllable (H) and two light syllables (L). Thus, for each of the three word types described above, three conditions were created: words containing a heavy final syllable (LLH), a heavy penult (LHL), or a heavy antepenult (HLL). The first three conditions allowed us to examine the role of syllable weight on stress assignment, unconfounded by possible effects of morphological structure. In addition, conditions (1)–(3) provided a baseline against which to assess potential effects of morphological structure in conditions (4)–(9). Prefixed words with a heavy first syllable (conditions (6) and (9)) allowed us to test whether the prefix retracts stress to the root (e.g. subMAtin) or whether the heavy antepenult competes with the light penult for main stress (e.g. SUBmatin). In contrast, prefixed words with a heavy final syllable (conditions (4) and (7)) were used to test whether primary stress by default falls on the first syllable of the root (e.g. beCArost) or whether a heavy ultima competes for main stress (e.g. becaROST).

The structure of the stimuli in conditions (7)–(9) was identical to that used in conditions (4)–(6), except that the unproductive prefixes were replaced with productive ones (e.g. be- ca- rost (4) vs. de-pa-dest (7) for verbs of the structure LLH), thus allowing us to see whether there was a difference between words containing productive vs. unproductive prefixes.

There are some critical issues to be addressed here with regard to the construction of the experimental items. First, given that main stress placement in English has been shown to be affected by the stress patterns of phonologically similar words (e.g. Baker & Smith Reference Baker and Smith1976, Guion et al. Reference Guion, Clark, Harada and Wayland2003; discussed in more detail in Section 2.2), potential similarities to existing English words were avoided where possible. In particular, care was taken to ensure that the final two syllables of the experimental items did not rhyme with existing words (based on the OED, e.g. *onvelop ~ develop). Second, we tried to avoid using any graphemic combinations that could potentially be associated with specific pronunciation variants (i.e. full long vowel or diphthongal pronunciations, e.g. <y> is highly likely to be realized as [aɪ] rather than [ɪ] in -by-). Third, with regard to the segmental structure of the nonce words, we made sure that all experimental items respected the phonotactic constraints of English (we specifically avoided unlikely or unidiomatic structural combinations in the word-final consonant clusters used in conditions (1), (4), and (7)).

In addition to the experimental items, 27 filler items were included in the study. Unlike the test items, however, the filler items varied in length: mono-, bi-, and quadri-syllabic words (9 each) served as distractors. The idea here was to present participants with words of different lengths so as to avoid the repetition of identical prosodic structures (for a similar procedure, see Domahs et al. (Reference Domahs, Plag and Carroll2014: 69)). Again, the non-test pseudo-verbs were constructed following the phonotactic rules of English. However, neither the phonological nor the morphological constraints outlined above applied to the filler items. In fact, novel verbs of varying length, different morphological categories (e.g. suffixed verbs such as copromify), and different syllable types (e.g. verbs containing CVVC syllables such as teep) were added to the list in order to increase the diversity of structures in the stimulus set. All nonce words were disambiguated as verbs by embedding them in a syntactic context (to be discussed in more detail below).

Several steps were taken to mitigate many of the problems encountered by previous studies. First, it was important that all target words were produced in the same metrical frame (e.g. Hayes Reference Hayes1984, Kelly Reference Kelly1988, Kelly & Bock Reference Kelly and Bock1988, Kager Reference Kager1989, Baumann & Ritt Reference Baumann and Ritt2017, Matzinger, Ritt & Fitch Reference Matzinger, Ritt and Fitch2021).Footnote ¹ The rhythmic context for the production of the target verbs was kept constant across trials by presenting the stimuli in a context where they were always preceded by a function word (always to, i.e. [tə]) and followed by an article (the [ðə] or a [ə]) which are typically realized as weak forms in this context. Second, it was necessary to control for the sentence position of the target words, thus allowing us to minimize the potential influence of post-lexical production processes (such as sentence prosody) on pronunciation. Prior research indicates that words before syntactic boundaries are subject to phrase-final lengthening, increased pitch and pitch range as well as the realization of a boundary tone (e.g. Conwell & Barta Reference Conwell and Barta2018, Baumann & Winter Reference Baumann and Winter2018). Therefore, all target words were presented in sentence-internal position, as exemplified below:

1. (1) He was extremely lucky to secolpam the destruction of Hurricane Catherina.

In this regard, our design differs from that used by Guion et al. (Reference Guion, Clark, Harada and Wayland2003) in which the target words were presented at the end of carrier sentences, thereby avoiding a potential confound between word-level stress and phrase-level prominence. Another issue concerns the relation between accentuation and focus. Specifically, words tend to receive focal stress in some contexts, for example, when preceded by focus-sensitive particles such as let alone or if not in the case of verbs (e.g. Gordon & Roettger Reference Gordon and Roettger2017, Baumann & Winter Reference Baumann and Winter2018). While it may be difficult, if not impossible, to predict or control the position of focal stress in isolated utterances, the effect of this factor was minimized by avoiding the use of explicit focus particles or constructions signaling contrast (e.g. not to __ but to__).

3.3. Procedure

All pseudo-words were embedded in carrier context sentences. The reason for this is twofold: first, to ensure a natural intonation, and second, to disambiguate the items as verbs. Participants were also explicitly instructed to regard them as verbs. Test items were randomized to avoid any effects of order of presentation. To facilitate the acceptance of the novel verbs, participants were provided with a short definition of each verb, each described in 15-20 words (see Schlücker & Plag Reference Schlücker and Plag2011 for a similar procedure). Both sentences were presented visually to the participants (font size 16, Times New Roman font). For each trial, participants were first presented with a short definition of the pseudo-word, followed by the target sentence which was marked in bold. An example of a definition of a novel verb and its corresponding context sentence is given below:

1. (2) If you caborasp something, you build something without anyone’s help.

   He managed to caborasp a small cabin.

One additional advantage of this presentation format was that participants saw each target item twice, once in the definition and once in the target sentence. Importantly, they were instructed to read through the definition first before turning to the target sentence. They were then asked to read out the carrier phrase, in the way they would say it and consider most natural. They were also told to take their time and read each sentence including the nonce verb silently to themselves before reading it out loud. Additionally, they were asked not to change the words in the sentences they read. Participants were also encouraged to try again if they were not happy with their reading. The recording of their voice started automatically once the target sentence had appeared on the screen. When they had completed the recording, participants clicked a button to continue to the next trial.

The experiment was distributed online. Stimulus presentation and voice recording were controlled by experimental software (Labvanced; Finger et al. Reference Finger, Goeke, Diekamp, Standfoß and König2017). The use of a microphone or headset was recommended. At the beginning of the session each participant completed one training trial using non-test words to familiarize them with the procedure. In total, the experiment lasted approximately 35 minutes.

3.4. Data coding

All responses were transcribed according to their stress patterns by three trained raters. Interrater reliability (i.e. agreement between at least two raters) was 95 percent (with all three raters showing perfect agreement in 57 percent of cases). Each recording was listened to no more than three times. The identification of primary stress positions was based on perceptual cues to stress, such as perceived loudness, pitch, and relative duration. When a stress position could not be identified after listening three times, the item was discarded. For the most part, however, the identification of primary stress positions was rather unambiguous, largely due to the reduction of unstressed syllables to schwa syllables in English. Some of the words were produced with a segmental pattern different from that of the stimulus. In case the substitution did not alter the syllabic structure of the target word (e.g. catorasp instead of caborasp), the response was included in the analysis. If, however, the substitution did change the syllabic structure, the response was discarded. This included, for example, cases where subjects left out or added consonants (e.g. subsmatin rather than submatin) as well as cases where subjects did not produce a short vowel (i.e. they realized a full long vowel or a diphthong instead). If a participant made several attempts, only the final answer was counted.

Only responses displaying an unambiguous stress pattern (i.e. responses for which at least two of the three coders agreed in their ratings) were included in the analyses. Responses which violated the criteria for inclusion set out above were discarded. Also excluded were cases where a) the audio quality was very bad (i.e. noisy environment, etc.), b) the response contained reading errors, c) the target word was not produced in its carrier sentence, and d) cases where all three raters disagreed in their ratings, leading to the exclusion of 9.54 percent of all responses. In total, 1081 items were included in the final analysis.

3.5. AML – Coding and set-up

We used AML to predict stress in the pseudo-verbs. AML (short for: Analogical Modeling of Language) is a computational implementation of exemplar-based analogical theory that classifies novel input based on a training set (henceforth: lexicon). It has been quite successfully employed in the past to model a variety of linguistic phenomena especially in morphology (e.g. Arndt-Lappe Reference Arndt-Lappe2014 on suffix rivalry between -ity and -ness nominalizers) and phonology (e.g. Arndt-Lappe Reference Arndt-Lappe2011 on compound stress assignment).

AML basically performs a classification task. For each new word that is given to the algorithm as a test item, the model decides between different outcomes (in our case stress type) based on the distribution of these values among exemplars in the lexicon. Within the dataset, each item is represented as a vector of features. Our coding comprises 13 features, 12 of which represent the phonological form of the word. Specifically, each word was coded as a string of sounds that are grouped by syllable position, with each feature corresponding to a syllabic constituent (e.g. onset/nucleus/coda of the final, penult, etc. syllable). Another feature encodes the length of the word, measured in terms of the number of syllables it contains. The coding of items in the test and lexicon files is identical, except that exemplars in the lexicon file are also coded for the outcome category (i.e. stress type in our case, with the levels ante, pen, fin) on top of their phonological form and length. The basic format of the coding is exemplified in Figure 1 below. When presented with a novel item, AML algorithmically extracts a set of similar words from the lexicon that resemble the target in an informative way. Based on the distribution of stresses in this set, the stress pattern for the new word will be predicted. An exhaustive, technical description of AML´s functional architecture can be found in Skousen (e.g. Reference Skousen1989, Reference Skousen, Skousen, Lonsdale and Parkinson2002). The important point here is that the outcome is probabilistic, reflecting a majority choice among relevant exemplars. What this means in our concrete case is that for each pseudo-word in our test set AML gives a probability of each possible outcome being selected (i.e. the probabilities of antepenultimate, penultimate, and final stress being assigned).

Figure 1.

Overall distribution of stresses in the experiment. All numbers represent absolute counts.

As a lexicon we used all verbs extracted from the English Cambridge Pronouncing Dictionary (Jones Reference Jones2006) that are longer than one syllable (N = 2,919). Very rare verbs and compounds were excluded. Importantly, each verb occurred only once in the lexicon file. In case the dictionary provided two or more pronunciation variants, we settled on the variant that was marked as the main pronunciation for British English. The 27 pseudo-verbs from the experiment served as the test set. All data were coded as strings of sounds, with vowels being replaced by orthographic vowels. The phonological coding was based on IPA transcriptions provided by Jones which were then converted into machine-readable DISC transcriptions (following the guidelines used for the CELEX Lexical Database, Baayen, Piepenbrock & Guilkers Reference Baayen, Piepenbrock and Guilkers1995; guidelines laid out in Burnage Reference Burnage1990). It is important to point out here that the coding of the phonological features was position-specific: Syllables were aligned by the right edge of the word, ensuring that onsets, nuclei, and codas of the final, penultimate, and antepenultimate syllables were always compared only with their respective counterparts in the same position. The coding is illustrated in Table 2. This was mainly due to two reasons: First, this type of alignment is more in line with traditional phonological approaches which rest on the assumption that stress is assigned from right to left (e.g. Hayes Reference Hayes1995), thus allowing us to compare the model´s predictions with previous claims from the formal literature about structural factors influencing stress assignment. Second, it allows the model to incorporate potential effects of position in the computation of similarity. If a syllable position was not occupied by a segment, it was coded as “=”. Empty spots were treated as valid feature values in the simulation. As mentioned above, vowel transcriptions were replaced by orthographic vowels. The rationale behind this decision was twofold: Apart from allowing us to test the role of orthography, using orthographic as opposed to phonetic vowel transcriptions has the added advantage that more detailed information about vowel quality (i.e. length and tenseness, generally taken to be major determinants of stress) is not available to the algorithm, thus making it harder for the model to arrive at its predictions. This is most desirable here, given that we want the model to remain largely unbiased by potential effects that occur on a higher level of structural representation (e.g. vowel length) than its pure phonemic make-up alone.

Table 2.

Example coding of the words “duplicate”, “improvise”, and “disbar”

Table 2. Long description

The table has eleven columns and three rows. Columns from left to right are labeled Stress, Oante, Nante, Cante, Open, Npen, Cpen, Ofin, Nfin, Cfin, and Length. The first row codes duplicate with ante stress, dj in Oante, u in Nante, equals sign in Cante, pl in Open, i in Npen, equals sign in Cpen, k in Ofin, a in Nfin, t in Cfin, and 3-syll in Length. The second row codes improvise with ante stress, equals sign in Oante, i in Nante, m in Cante, pr in Open, o in Npen, equals sign in Cpen, v in Ofin, i in Nfin, z in Cfin, and 3-syll in Length. The third row codes disbar with fin stress, equals sign in Oante, equals sign in Nante, equals sign in Cante, d in Open, i in Npen, s in Cpen, b in Ofin, a in Nfin, equals sign in Cfin, and 2-syll in Length. The equals sign indicates no segment present in that position.

An obvious question here is where to place syllable boundaries in word-medial positions. While the Jones Dictionary does provide syllabifications for each word, these are rather inconsistent. Therefore, all words were re-syllabified according to the Maximal Onset Principle as well as the principle of rising sonority in English onsets. Consonants were regularly assigned to the syllable onset (e.g. be.stir, where the dot indicates the syllable boundary), unless it would result in an illegal onset cluster in English (*co.nceive). In this case, any remaining segments were assigned to the coda of the preceding syllable (e.g. conceive would be divided as follows: con.ceive). Another issue to be addressed here is the presence of triphthongs in some words (e.g. “aɪə” in transpire). Sound sequences of three adjacent vowels were parsed as a diphthong followed by a regular vowel. For example, transpire was coded as /træn.spaɪ.ə/.

Given that some studies documented effects of word length (e.g. Ernestus & Neijt Reference Ernestus and Neijt2008, Dabouis & Fournier Reference Dabouis, Fournier, Ballier, Jean-Michel Fournier and Yamada2023), we also included length as a feature on top of the formal features, with the values 2-, 3-, and 4-syllables. With regard to the outcome category, all verbs in the lexicon set were coded for their stress type (i.e. as fin, pen or ante), which was taken directly from Jones.

Prior to running a simulation on the pseudo-words we tested the lexicon on itself. The parameters were set in such a way as to ensure that for each new item to be classified, its copy in the lexicon was excluded from the training set, also known in the AM literature as the “leave-one-out” procedure. This essentially allows us to simulate the classification of words that are treated as novel by the algorithm. Classification accuracy ranges between 84 percent for penultimate stress and 94 percent for antepenultimate stress, thus justifying its use as a valid measure of lexical support in the subsequent analyses. In a second simulation, we set AML the task of predicting stress in the pseudo-verbs based on the stress distribution in the dictionary database.

3.6. Statistical analysis

The production data was analyzed using generalized mixed-effects logistic regression. For the regression analyses we used the lme4 package (Bates et al. Reference Bates, Mächler, Bolker and Walker2015). We fitted logistic regression models with the position of stress as the response variable and weight, prefixation, and lexical support (analogy) as predictor variables. The weight variable takes one of three values (initial, medial, and final), specifying the position in the word that contains the heavy syllable. For example, initial means that the test item has a heavy first syllable, the final and penult syllables being both light (i.e. HLL). Prefixation and analogy, on the other hand, were coded numerically. Regarding the latter, we used the lexical support measures derived from AML which represent the predicted probabilities of antepenultimate stress being assigned to each pseudo-verb. With regard to prefixation, we calculated prefix likelihood scores for each individual prefix (this measure was adapted from Wurm (Reference Wurm2000)), giving the proportion of prefixed verbs among all verbs that start in a given letter string. To address the controversies raised in the literature regarding the morphological status of semantically opaque prefixes (see Section 2.1 for a discussion), we compared two types of scores. One score considers only instances of fully transparent prefixation (e.g. rephrase) in its prefix count. The other also treats cases of opaque prefixation (e.g. relegate) as “prefixed” in addition to fully transparent constructions. The morphological coding was based on all words starting in the pertinent prefix-like strings for which there was an entry in the Wikimorph database (Yarbro & Olney Reference Yarbro, Olney, Roll, McNamara, Sosnovsky, Luckin and Dimitrova2021); words were extracted with the help of a software tool created by the authors of this paper (Morph2Excel: Preidt et al. Reference Preidt, Sonja, Jakob, Anika, Lilly-Kristin, Charlie and Arndt-Lappe2024). Specifically, words were assigned to one of three morphological classes (noprefix, intransparent, transparent) based on morphological and etymological information provided by Wikimorph. If the word was broken down into Prefix + Verb (or alternatively, Prefix + Noun/Adverb, etc.), the word was coded as “prefixed”. Otherwise, Wikimorph’s etymology section was checked: If the prefix was listed there, the word was coded as “intransparent”. If the prefix was not listed there, the word was coded as “noprefix”. Each word was coded only once. For most words, however, the model returned multiple entries. In case the individual entries differed in their analyses, the transparent combination was chosen.

Prefix likelihoods can take on any value within a range from zero to positive infinity. Values above 1 indicate there are relatively more transparently prefixed than non-prefixed verbs among all verbs starting in a given letter string in the database. Conversely, values below one indicate a relatively higher proportion of non-prefixed verbs, and a score of one indicates perfect balance between the two. Values for the prefixes used in this study are listed in Table 3. They ranged from 0.062 (con-) to 44 (mis-) for the prefix count that only considers transparent structures and between 0.071 (ob-) to 44 (mis-) for the count that includes opaque structures on top of the fully transparent cases, averaging 4.852 and 6.27, respectively. Non-prefix-like strings (e.g. ca-) were coded as 0.

Table 3.

Prefix likelihood ratios for the prefixes used in this study. Score 1 includes only transparent forms in its prefix count; Score 2 treats both opaque and transparent forms as prefixed

Table 3. Long description

The table has three columns: Prefix, Score 1, and Score 2. From top to bottom, the prefixes and their values are as follows. mis- has 44 for both Score 1 and Score 2. un- has 22 for both. re- has 0.982 for Score 1 and 6.2 for Score 2. pre- has 0.773 for Score 1 and 5.5 for Score 2. sub- has 0.214 for Score 1 and 4.67 for Score 2. be- has 1.89 for Score 1 and 2.67 for Score 2. dis- has 1.02 for Score 1 and 2.54 for Score 2. con- has 0.0615 for Score 1 and 2.45 for Score 2. de- has 0.565 for Score 1 and 1.6 for Score 2. e- has 0.667 for Score 1 and 1.1 for Score 2. a- has 0.25 for Score 1 and 0.667 for Score 2. se- has 0.0833 for both. ob- has 0.0714 for both. The highest values are for mis- and un-, while the lowest are for ob- and se-.

First inspection of the score distributions reveals a strong positive skew: for both measures, the vast majority of transparency scores cluster toward the lower end of the scale near or below 1 (especially a-, e-, ob-, se-, and also con-, de-, pre-, re- if we consider only the fully transparent cases in the prefix count). At the same time, we also identify a few outliers extending the upper end of the scale, most notably un- (22) and mis- (44). For this reason, the ratios were logarithmically transformed. However, this type of transformation returns negative values for those scores that are below one. To avoid division by zero, we added 1 to both the nominator and the denominator before computing the logarithm. The resulting prefix likelihood scores were scaled and centered.

In a first round, we performed three separate analyses, one for each type of stress (i.e. antepenultimate, penultimate, and final) as the dependent variable. First we fitted a model with the three predictors and antepenultimate stress as the dependent variable. This variable had two levels, i.e. ante and no_ante, depending on the presence or absence of stress on this syllable. Moreover, to control for potential sources of random variation in the data, the test word and the participant ID were included as random intercepts. We will focus our discussion below only on the model for antepenultimate stress. Given the extremely small number of observations of final stresses in the experiment, virtually all non-antepenultimate stresses are penultimate stresses.

We started out with a regression model that included all three predictors as well as random intercepts for both word and participant. A series of models was then compared by removing and adding predictor variables to see if that improved the quality of the model – the AIC served as a measure of relative model quality here.