Cues to Stress Assignment in Reading Aloud
Nelson Cowan, Editor
1Department of Psychology, Royal Holloway, University of London
Petroula Mousikou is now at the Max Planck Institute for Human Development, Berlin, Germany.
Stimulus and data files for all experiments and simulations reported in this article are openly accessible at https://osf.io/5736m/.
This research was supported by The Leverhulme Trust (RPG 2013-024) and the Economic and Social Research Council (ES/L002264/1). We are grateful to Clare Lally, Hannah Harvey, and Benedetta Cevoli for research assistance, and to Debra Jared and Marcus Taft for helpful comments on an earlier version of this manuscript.
*Correspondence concerning this article should be addressed to Maria Ktori, Cognitive Neuroscience Sector, Scuola Internazionale Superiore di Studi Avanzati (SISSA), via Bonomea 265, 34136, Trieste, Italy email@example.com
Author information ►Article notes ►Copyright and License information ►
Received 2016 May 27; Revised 2017 Aug 14; Accepted 2017 Sep 8.
Copyright © 2018 The Author(s)
This article has been published under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/3.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Copyright for this article is retained by the author(s). Author(s) grant(s) the American Psychological Association the exclusive right to publish the article and identify itself as the original publisher.
Research seeking to uncover the mechanisms by which we read aloud has focused almost exclusively on monosyllabic items presented in isolation. Consequently, important challenges that arise when considering polysyllabic word reading, such as stress assignment, have been ignored, while little is known about how important sentence-level stress cues, such as syntax and rhythm, may influence word reading aloud processes. The present study seeks to fill these gaps in the literature by (a) documenting the individual influences of major sublexical cues that readers use to assign stress in single-word reading in English and (b) determining how these cues may interact with contextual stress factors in sentence reading. In Experiments 1, 2, and 3 we investigated the effects of prefixation, orthographic weight (i.e., number of letters in a syllable), and vowel length on stress assignment by asking participants to read aloud carefully-constructed nonwords that varied on the presence of these cues. Results revealed individual effects of all three cues on the assignment of second-syllable stress. In Experiment 4, we tested the effects of these cues on stress assignment in the context of sentence reading. Results showed that sublexical cues influenced stress assignment over and above higher-level syntactic and rhythmic cues. We consider these findings in the framework of extant rule-based, distributed-connectionist, and Bayesian approaches to stress assignment in reading aloud, and we discuss their applications to understanding reading development and acquired and developmental reading disorders.
Keywords: reading aloud, stress assignment, sublexical cues, computational modeling
One of the fundamental insights of psycholinguistic research over the past 40 years is that the computation of sound-based (phonological) codes is central to skilled reading and reading acquisition (see e.g., Frost, 1998; Melby-Lervåg, Lyster, & Hulme, 2012; Perfetti, 2003; Rastle & Brysbaert, 2006; Share, 1995 for reviews). This insight has motivated an extensive body of empirical research (e.g., Ferrand & Grainger, 1992; Lukatela & Turvey, 1994) and the development of computational models (e.g., Coltheart, Rastle, Perry, Langdon, & Ziegler, 2001; Harm & Seidenberg, 2004; Perry, Ziegler, & Zorzi, 2007), which seek to explain how we translate printed letter strings into their corresponding sounds. It has also supported major shifts in approaches to reading instruction, so that children’s learning of the relationship between letters and sounds (i.e., phonics) is given high priority (e.g., Rayner, Foorman, Perfetti, Pesetsky, & Seidenberg, 2001; Rose, 2006).
Despite widespread acceptance of the central role of phonology in reading, research seeking to uncover the mechanisms by which we translate orthography to phonology has focused almost exclusively on monosyllables. Critically, the focus on monosyllables has allowed reading research to ignore major challenges that emerge when considering polysyllables, the most important of which is the assignment of stress. Stress at the level of the single word refers to the phonetic accentuation of a particular syllable (as in camel vs. canal). Evidence from eye-movement research suggests that the computation of stress facilitates access to lexical information in silent reading (Ashby & Clifton, 2005). Stress is also used to disambiguate words phonologically in sentence contexts (e.g., record as a noun or verb), and at the word level (e.g., trustee vs. trusty). More importantly, polysyllables make up the vast majority of words in most languages (e.g., Baayen, Piepenbrock, & Gulikers, 1995 for English, Dutch, and German languages); thus, failing to understand how phonology is computed for polysyllables presents a major impediment to any theory of reading that aspires to completeness.
The present study provides a substantial advance in our understanding of how we read letter strings with more than one syllable. In a series of three reading aloud experiments, we investigate the major factors that influence stress assignment at the single-word level. The data from these experiments are then used to assess the performance of three computational accounts of reading disyllables aloud—the rule-based disyllabic algorithm of Rastle and Coltheart (2000), the connectionist dual process (CDP++) model (Perry, Ziegler, & Zorzi, 2010), and the connectionist print-to-stress network of Ševa, Monaghan, and Arciuli (2009). We also consider these results in the context of a Bayesian approach to understanding stress assignment (Jouravlev & Lupker, 2015a). In a fourth experiment, we test the influence of higher-level contextual cues on stress assignment, and critically, assess how these sentence-level cues impact on the word-level cues revealed in the first three experiments. This fourth experiment is especially interesting because it provides a means to begin to bridge the empirical evidence from single-word reading aloud into the domain of sentence reading. Research on single-word reading aloud and sentence reading span vast literatures; however, we show that these literatures rarely overlap. This experiment will reveal whether the mechanisms underlying stress assignment in single-word reading aloud are fundamentally altered when printed stimuli are placed in sentence contexts. Our data will thus provide vital new constraints on the development of computational models of reading aloud as they move beyond the monosyllabic domain, and beyond the domain of words read in isolation.
Stress Assignment in Single-Word Reading
Rastle and Coltheart (2000) considered how an existing model of reading aloud, namely the dual-route cascaded model (hereafter referred to as DRC; Coltheart et al., 2001), could accommodate reading of disyllables in English, including the assignment of stress. One possibility is that stress could be retrieved lexically, using whole-word information stored in memory. This mechanism would allow the reader to pronounce familiar words such as camel and canal, which look similar but are stressed on different syllables. However, this mechanism would not account for readers’ ability to assign stress to unfamiliar words or nonwords that are not stored in lexical memory (Rastle & Coltheart, 2000). Thus, another possibility is that stress could be computed on the basis of sublexical information, much as the dual-route theory asserts that people are able to compute the phonemes of a printed nonword. So how might stress be computed without recourse to lexical information?
Distribution of Stress Patterns in the Language
In considering what might be the nature of the sublexical cues used to assign stress, one of the first proposals put forward was that readers assign stress on the basis of the simple distribution of stress patterns in the language (Colombo, 1992). For example, about 80% of Italian polysyllabic words are stressed on the penultimate syllable (Sulpizio, Burani, & Colombo, 2015), so a sublexical process applicable to polysyllabic words could implement penultimate stress as a default. Similarly, around 75% of English disyllables are stressed on the first syllable (calculated from CELEX, Baayen et al., 1995), so a sublexical process that implemented initial stress as a default would stress a high proportion of words correctly. Based on this type of language-specific distributional rule, words that follow the rule would be considered regular, whereas words that fail to adhere to the default stress rule would be considered irregular.
However, the evidence to support this hypothesis is weak. Although some studies in Italian and Russian report a processing advantage for stress-regular words based on the distributional rule (Colombo, 1992; Colombo & Tabossi, 1992 in Italian; Jouravlev & Lupker, 2014, in Russian in the by-subjects analysis) others fail to do so (Burani & Arduino, 2004; Burani, Paizi, & Sulpizio, 2014, in Italian; Jouravlev & Lupker, 2014, in Russian in the by-items analysis). In English, the strongest evidence for a distributional rule comes from a study carried out by Brown, Lupker, and Colombo (1994). Using items developed by Monsell, Doyle, and Haggard (1989), they reported that disyllabic words stressed on the first syllable were read aloud faster than disyllabic words stressed on the second syllable. However, analyses were conducted only by subjects, so it is unknown whether this effect held across items. Similarly, although the earlier study by Monsell et al. (1989) had reported a numerical advantage for disyllables stressed on the first syllable, no statistics relevant to this comparison were reported. In a more recent study of English reading aloud, Rastle and Coltheart (2000) failed to find any evidence for a stress regularity effect based on the distributional rule. Stress regularity effects were absent even for low-frequency words, in which sublexical information is thought to play a more potent role in the translation of orthography to phonology (Jared & Seidenberg, 1990; Seidenberg, Waters, Barnes, & Tanenhaus, 1984). Hence, the empirical support for an account of stress assignment based on the distribution of stress patterns in the language is weak.
Distribution of “Word Endings” and “Word Beginnings”
Some authors have suggested instead that distributional information about the relationship between smaller sublexical units and stress patterns may influence stress assignment. There is reasonably good evidence that the “endings” of words can serve as an indicator of stress position in Italian. The ending in this research is defined as the orthographic unit that includes all the final letters of a word starting from the nucleus of the second syllable (e.g., -ola in picola). Words that contain the same endings are said to be part of the same “stress neighborhood.” Several studies have now indicated that the reading aloud of Italian words and nonwords is influenced by the consistency of items within their stress neighborhood (see Sulpizio et al., 2015, for a review). Words and nonwords with many stress friends (i.e., items with a stress pattern that is consistent with the stress pattern of the majority of words in the same neighborhood) are read faster and more accurately than those with many stress enemies (i.e., items with a stress pattern that is inconsistent with that of most words in the same neighborhood). Furthermore, this effect appears to be modulated by the number of stress neighbors that are available in the language for a given ending, with a greater influence being observed for stress neighbors that are widely represented in the Italian language (Sulpizio, Arduino, Paizi, & Burani, 2013). Finally, Jouravlev and Lupker (2014, 2015b) provided evidence that the orthographic endings of words can serve as a stress cue in Russian.
The evidence that distributional properties of word endings can influence stress assignment is less plentiful in English. The most rigorous study investigating this hypothesis was conducted by Arciuli and Cupples (2006). They reported a linguistic analysis of 340 endings in disyllabic English words and showed that these are associated to varying degrees with particular stress patterns (e.g., the endings -ock and -ibe are associated with first- and second-syllable stress, respectively). They then demonstrated that the presence of these endings in nonwords biased stress decisions when adults were asked to underline the part of a nonword that they would emphasize had they been reading the nonword aloud. Subsequently, Arciuli and Cupples (2007) provided preliminary data on how distributional properties of word beginnings (i.e., letter string up to and including the first vowel or vowel cluster) might also impact adults’ stress decisions.
Despite recent enthusiasm for the notion that simple units like word beginnings and word endings may provide reliable cues to stress assignment, it is unlikely that such an account would work in English. To learn the statistical relationship between a word beginning or word ending and stress assignment, the learner needs to represent the orthographic input in such a way as to identify the word beginning and word ending. This turns out to be a challenge in English. For example, when the definition of word beginning used by Arciuli and Cupples (2007) is applied to the whole lexicon of disyllables, many hundreds of words are exposed in which the word beginning includes the vowel of the second syllable (i.e., part of the word ending; e.g., quiet, ruin, dial, react, triumph, stoic). Similarly, there are many hundreds of words that do not have a word ending (on the definition given by Arciuli & Cupples, 2006) as a result of syllabic “l” (e.g., apple, drizzle, bubble), syllabic “m” (e.g., schism, spasm, rhythm), or falling diphthongs (e.g., scour, squire). It is also unclear how to classify the letter Y; this must be treated as a vowel in abyss but as a consonant in beyond. Arciuli and Cupples (2006, 2007) avoided these problems because they selected only a very small proportion of the many thousands of possible word beginnings and word endings, in which these units could be unambiguously identified. However, these challenges would need to be solved for this type of account of English reading aloud to be viable.
Orthographic, Phonological, and Morphological Cues
In addition to these distributional cues to stress, other forms of phonological and orthographic information have been argued to provide sublexical cues to assigning stress in English polysyllables. Several researchers have argued that vowel length and the phonological weight of a syllable are important determinants of stress (Baker & Smith, 1976; Chomsky & Halle, 1968; Hayes, 1982; Kelly, 2004; Mousikou, Sadat, Lucas, & Rastle, 2017; Smith & Baker, 1976), such that syllables containing many phonemes (e.g., consonant clusters in the coda, as in collapse, elect) and/or long vowels (e.g., atone, divine) tend to attract stress in pronunciation tasks. Morphological units are also thought to provide important cues to stress assignment. Rastle and Coltheart (2000) provided evidence that prefixes (e.g., re-, mis-) repel stress when typical adults read disyllabic nonwords aloud (Rastle & Coltheart, 2000). The association between prefixes and final stress is also evident when patients with acquired surface dyslexia attempt to read disyllabic prefixed words aloud. In such cases, these patients tend to make stress errors (e.g., reading the word reflex with second-syllable stress; Ktori, Tree, Mousikou, Coltheart, & Rastle, 2016). Finally, it has been argued across a number of experiments that syllables with greater orthographic weight (i.e., as defined by the number of letters) and/or syllables with redundant letters tend to attract stress (e.g., the final “te” in roulette, Kelly, Morris, & Verrekia, 1998; Mousikou et al., 2017).
Confounding of Cues in Empirical Work
Thus far, there is evidence that a large variety of sublexical cues may influence stress assignment in reading aloud. However, studies in this domain have typically confounded some or all of these cues. For example, Rastle and Coltheart (2000) argued that participants reliably assign second-syllable stress to nonwords such as misbane because of the presence of prefixes (in this case, mis-). However, misbane also has high orthographic weight and a long vowel in the second syllable, possibly biasing the reader toward second-syllable stress. Similarly, Arciuli and Cupples (2006) claimed that readers are likely to assign second-syllable stress to certain nonwords, such as aject, because the ending, -ect, is typically associated with second-syllable stress. However, this nonword also contains a prefix (i.e., a-) and its second syllable has high orthographic weight, thus potentially biasing the reader toward second-syllable stress. Likewise, in the Arciuli and Cupples’ (2007) preliminary study of the impact of word beginnings on stress decisions, all of the beginnings associated with second-syllable stress were also prefixes. In contrast, the beginnings that were not associated with second-syllable stress were never prefixes. Finally, Smith and Baker (1976) argued that a nonword like gevesp should receive more second-syllable stress than a nonword like nodud, because the former contains more letters in the second syllable. However, in addition to more letters, gevesp contains more phonemes in the second syllable than nodud, which could also bias stress assignment toward the second syllable. These authors also inadvertently included prefixes in some of their stimuli, which is another uncontrolled potential cue to second-syllable stress.
In a recent megastudy of disyllabic reading, Mousikou et al. (2017) sought to disentangle some of these cues. Although they found evidence for individual contributions of vowel length and orthographic weight, they were unable to determine whether prefixation contributed to stress assignment due to its confounding with other sublexical cues to stress. Consequently, we do not have a clear understanding of the individual influences of word-level cues on stress assignment. Mousikou et al. (2017) suggested that one possibility would be to conduct factorial experiments with nonwords purposely designed to pull these interrelated cues apart. Therefore, in the present work, we conducted a series of carefully-constructed reading aloud experiments (Experiments 1 through 3), in which we sought to delineate the individual influences of prefixation, orthographic weight, and vowel length on stress assignment in English. In order to strengthen our conclusions about the contribution of these sublexical cues, we also included in our analyses the potential confounding variables of word ending frequency (Sulpizio et al., 2013), the association between word endings and stress assignment (Arciuli & Cupples, 2006), lexical similarity (orthographic Levenstein distance [OLD] 20; Yarkoni, Balota, & Yap, 2008), and bigram frequency.1
Sentence-Level Cues to English Stress Assignment
In addition to the sublexical cues discussed earlier, stress assignment appears to be strongly influenced by contextual factors that operate beyond the level of the single word. Most importantly, the grammatical category to which a word belongs is associated with specific stress patterns (Chomsky & Halle, 1968; Howard & Smith, 2002; Kelly & Bock, 1988; Liberman & Prince, 1977). In particular, the majority of English disyllabic nouns (approximately 90%) in CELEX take first-syllable stress, while most disyllabic verbs (approximately 67%) tend to be stressed on the second syllable (Howard & Smith, 2002). Accordingly, Arciuli and Cupples (2003) showed that grammatical class judgments are faster and more accurate for typically stressed words (i.e., nouns with first-syllable stress and verbs with second-syllable stress) than for atypically stressed words (i.e., nouns with second-syllable stress and verbs with first-syllable, stress). More recently, Breen and Clifton (2011, 2013) embedded stress-alternating noun/verb homographs in sentences (i.e., (a) The brilliant abstract the best ideas from the things they read; (b) The brilliant report the best ideas from the things they read; (c) The brilliant abstract was accepted at the prestigious conference; (d) The brilliant report was accepted at the prestigious conference). Longer reading times were found for “from the things they read” (sentences a and b) than for “at the prestigious conference” (sentences c and d), thus showing that readers had to shift their (preferred) syntactic representation of the words abstract and report from noun to verb. Interestingly, this association between grammatical class and stress has also been observed with nonwords, whose stress assignment is influenced by the syntactic context in which they are placed (Baker & Smith, 1976; Kelly & Bock, 1988; Smith & Baker, 1976; Smith, Baker, & Groat, 1982). Readers are more likely, for example, to assign first-syllable stress to a disyllabic nonword such as pralel when this is placed into a noun context (e.g., “The pralel caught the bird”) compared with when it is placed in a verb context (e.g., “The hunter pralel the bird”; Kelly & Bock, 1988).
Rhythm, and specifically the alternation between strong and weak beats, is another sentence-level factor that has been shown to affect stress placement in reading aloud (Kelly & Bock, 1988; Kentner, 2012). For example, in both of these sentences, “Use the pralel proudly” and “Planes will pralel pilots,” the nonword pralel is preceded by a weak beat and its stress is thus biased toward a strong beat (i.e., trochaic context). Conversely, in the sentences, “The proud pralel proposed” and “The pins pralel balloons,” the nonword pralel is preceded by a strong beat and its stress is thus biased toward a weak beat (i.e., iambic context). Kelly and Bock (1988) found that a nonword placed in a trochaic-biased context (i.e., strong–weak) was more likely to receive first-syllable stress compared with the same nonword placed in an iambic-biased context (i.e., weak–strong), irrespective of syntactic context. In a more recent study, Kentner (2012) constructed German sentences in which syntactic ambiguity could be resolved by the stress assigned to a target word, which could function either as an adverb or as a comparative quantifier. Kentner (2012) reported that during reading aloud and silent reading, participants appeared to generate an implicit prosodic representation based on the principle of rhythmic alternation of syllables (i.e., avoiding stress clashes due to adjacent stressed syllables). Critically, this implicit prosodic representation biased syntactic analysis of the target word, even though this led to integration difficulties on some trials.
The work of Kentner (2012) suggests that prosodic and syntactic cues arising at the sentence level can interact with one another. However, research is virtually silent on how these types of sentence-level cues impact on the sublexical cues typically studied in the domain of reading aloud. This is an important shortcoming, as any theory of reading that aspires to completeness must consider how factors operating at the level of single words may be influenced when words are presented in the context of whole sentences. If we step away from the specific problem of stress assignment, we are unaware of any theory that describes the mechanisms that underpin reading aloud in sentences. Further, although there is a small amount of empirical research comparing reading behavior when single words are read aloud versus read silently in sentences (Kuperman, Drieghe, Keuleers, & Brysbaert, 2013), we are unaware of any work that has compared reading aloud in single-word versus sentence-reading contexts.
In the domain of stress assignment, two studies conducted 40 years apart have begun to address this gap. The first of these studies was conducted by Baker and Smith (1976; see also Smith & Baker, 1976), who asked participants to read aloud nonwords appearing in noun or verb contexts within sentences. They observed that sublexical cues thought to operate at the word level (e.g., vowel length) were observed in these sentence contexts, suggesting that higher-level contextual cues do not completely override lower-level sublexical cues. However, there are a number of serious problems with this work, which undermine the conclusions that can be drawn. First, no data for nonwords read aloud in isolation were presented, making it difficult to ascertain the impact of grammatical category on the sublexical cues tested. Second, in addition to the confounding of different sublexical cues discussed earlier, these studies included very close analogies to existing words (e.g., zeranda, estonish, thrombossis), thus making it difficult to distinguish between potential effects of sublexical cues and word neighbors. They also included nonwords that were sufficiently difficult as to promote a “very cautious, syllable by syllable” reading strategy, which does not resemble natural reading (e.g., tupaivend, ollanteam; Baker & Smith, 1976, p. 23). Finally, the sentence contexts in which these nonwords were placed were not controlled for rhythm, as the impact of this factor on stress assignment only became apparent later through the work of Kelly and Bock (1988).
More recently, Spinelli and colleagues (Spinelli, Sulpizio, Primativo, & Burani, 2016) have returned to this issue by investigating the impact of contextual information on the effect of stress neighborhood consistency, which is frequently observed on single-word reading in Italian (see Sulpizio et al., 2015, for a review). They found that when contextual information (e.g., gender and number in the case of nouns; person in the case of verbs) is present, this information substantially overrides stress neighborhood consistency at the level of the single word. This result suggests a complex interaction between contextual cues and cues operating at the level of single words, perhaps where context contributes to the prosodic structure of a phrase, which in turn constrains the processing of individual words within the phrase. However, Spinelli et al. (2016) argued that further research is necessary to draw firm conclusions. They suggested that research in which words and nonwords are placed in sentence contexts would be particularly desirable, “as the presence of a context may affect stress processing in non-obvious ways” (Spinelli et al., 2016, p. 9)
For all of these reasons, our empirical work includes a fourth experiment, in which we investigate how the higher-level contextual cues of grammatical category and rhythm interact with the lower-level sublexical cues identified through Experiments 1, 2, and 3, to influence stress assignment in reading aloud English sentences.
Models of Stress Assignment in English
The problem of stress assignment has thus far been considered in three computational accounts of single-word reading in English: the rule-based algorithm proposed by Rastle and Coltheart (2000), the connectionist dual process model (CDP++; Perry et al., 2010), and the distributed-connectionist network proposed by Ševa et al. (2009). These accounts adopt different approaches to stress assignment during disyllabic word reading, and to the way sublexical cues may affect this process. These models are restricted to the processing of isolated letter strings: they have nothing to say about the impact of higher-level contextual cues that may arise in sentence reading contexts. There are, of course, many computational models of sentence reading, including models of eye-movement control (e.g., Reichle, Pollatsek, & Rayner, 2012), discourse processing (Kintsch, 1988), and syntactic parsing (e.g., Frazier & Rayner, 1982; see Rayner & Reichle, 2010; Reichle, 2015, for reviews). However, none of these models has anything to say about the computation of phonology required for reading aloud. Thus, while there is a substantial theoretical basis for understanding stress assignment at the level of the single word, there is as yet no model to offer predictions about how higher-level contextual cues may constrain processing at the level of the single word. In the following, we therefore describe the three models of single-word reading aloud presently available. The three models are shown in Figure 1.
The Rastle and Coltheart algorithm (hereafter referred to as RC00; see Figure 1a) is a partial implementation of the sublexical pathway of the DRC model (Coltheart et al., 2001), which computes the phonology of a word using a set of grapheme-to-phoneme rules. In this rule system, the spelling-to-sound conversion follows the grapheme-to-phoneme correspondence rules identified and used by the DRC model (Rastle & Coltheart, 1999; and subsequently, Coltheart et al., 2001), while stress placement is determined following the identification of orthographic strings that correspond to prefixes and suffixes. Specifically, the identification of a prefix (e.g., pre-, de-, dis-. re-, mis-) results in the assignment of second-syllable stress, whereas the identification of a suffix results in the assignment of first-syllable stress (except for a small group of stress-taking suffixes, such as -een, -ique, -oo, identified by Fudge, 1984). In the absence of an identifiable affix, the default first-syllable stress pattern of English disyllables is assigned. Evaluation of the RC00 revealed very good performance in stressing disyllabic words, as the algorithm assigned stress correctly to almost 90% of all English disyllabic words present in the CELEX database (Baayen, Piepenbrock, & van Rijn, 1993). Furthermore, when compared against human stress assignments to disyllabic nonwords, the algorithm predicted the modal human stress for 93% of items that received first-syllable stress and 75% of items that received second-syllable stress (Rastle & Coltheart, 2000).
The CDP++ model (Perry et al., 2010; see Figure 1b) is a dual-pathway model of disyllabic reading aloud that is built on its predecessor, the CDP+ model of monosyllabic reading aloud (Perry et al., 2007). Unlike the RC00 algorithm, the CDP++ model is a full processing model that produces a pronunciation, stress marker, and reaction time (RT). The CDP++ is very similar to the CDP+ model except for an increase in the number of letter and phoneme slots to accommodate longer words, an expanded input coding template to accommodate disyllables, the introduction of the schwa phoneme to deal with vowel reduction, the inclusion of stress nodes to represent the position of stress, and the use of a larger training corpus and lexicon. The lexical procedure of the CDP++ model is identical to that of the DRC model, storing item-based knowledge about the orthography and phonology of a known word. Accordingly, the lexical procedure of the CDP++ model directly activates the stress that is associated with a familiar word’s spoken form. The sublexical procedure of this model incorporates a two-layer associative (TLA) network for mapping graphemes onto phonemes, unlike the set of rules used by the DRC model, and by extension, the RC00 algorithm. In the TLA network, the orthographic input is organized along a graphosyllabic structure, which segments each syllable into onset (i.e., the initial consonant graphemes) and rime (i.e., the vowel and any following consonant graphemes) units that make direct contact with the corresponding phonological units and two sublexical stress units. During a training phase, the TLA network uses a connectionist algorithm to learn grapheme-to-phoneme and grapheme-to-stress associations based on statistical regularities across the model’s training set. Accordingly, the sublexical procedure of the CDP++ model activates the stress pattern that it learnt to associate with certain graphemes in order to assign stress to unfamiliar letter strings. The performance of the CDP++ model on stressing disyllabic words was evaluated against several databases (English Lexicon Project database; Balota et al., 2007; Chateau & Jared, 2003; Yap & Balota, 2009) and in all cases, the model was highly successful in predicting the correct stress pattern of words, with the majority of stress errors resulting from the model assigning first-syllable stress to words with second-syllable stress. The CDP++ model’s performance was also assessed against the human nonword reading aloud data of Rastle and Coltheart (2000). The model predicted the modal human stress for approximately 92% of items that received first-syllable stress and 51% of items that received second-syllable stress.
The Ševa et al. (2009) network (hereafter referred to as SMA09; see Figure 1c), also uses a distributed-connectionist framework to simulate stress placement in English disyllables during reading aloud, but provides no pronunciation or RT. This model uses the statistical regularities available in its training set to learn how to map an orthographic input onto a stress pattern but it differs from the CDP++ in three key aspects. First, though the CDP++ model uses a structured template representing onsets and rimes in each syllable, the SMA09 model organizes its orthographic input using a simple left-aligned, slot-based coding scheme. Second, though the orthographic and phonological representations are directly connected in the CDP++, the SMA09 includes an additional intermediate layer of hidden units between its orthographic input and the phonological output layers, which could potentially contribute to learning more complex relationships between orthography and phonology. Finally, the training set of the SMA09 model includes only disyllabic words, whereas the CDP++ is trained on both monosyllabic and disyllabic words. The SMA09 model’s performance on assigning stress to a subset of disyllabic words drawn from the CELEX database proved to be slightly better than the RC00 algorithm (87% and 84% of correct stress predictions, respectively). However, the model performed less well than the RC00 algorithm and the CDP++ model in predicting the human modal stress assigned to the group of disyllabic nonwords used in the Rastle and Coltheart (2000) study. Specifically, the SMA09 model predicted the correct modal stress pattern for almost 88% of the first-syllable stressed items and 50% of the second-syllable stressed items. Similarly to the CDP++ model, the SMA09 network’s inferior performance was due to assigning first-syllable stress to nonwords that were given second-syllable stress by the majority of human readers. Ševa et al. (2009) argued, however, that the superior performance of the RC00 algorithm on this set of nonwords could have been due to an overrepresentation of items containing affixes, which were readily identified by the RC00 algorithm.
In summary, despite their different approaches, all three models described appear relatively successful in simulating the assignment of stress in English disyllables. Further, they account for the role of sublexical cues on stress assignment from different standpoints. On one hand, the RC00 algorithm employs a set of all-or-none rules that are governed by the morphological structure of a word. On the other hand, the CDP++ model and the SMA09 network adopt a statistical learning approach, which allows for the discovery of print-to-stress regularities in the language. In the present study, we compare simulation results from these models against data from Experiment 1 through 3 (isolated presentation) to examine which of these alternative approaches to stress assignment best captures human reading aloud performance. Though these models are unable to simulate higher-level contextual factors on stress assignment (e.g., grammatical category, rhythm), we also assess how well they predict the impact of lower-level cues when nonwords are placed in sentences (Experiment 4).
Experiments 1 Through 3
Experiments 1 through 3 aimed to establish the effects of three sublexical cues on stress assignment during reading aloud, namely, prefixation, vowel length, and orthographic weight. In all three experiments, we asked participants to read aloud a series of carefully-constructed disyllabic nonwords, in which we systematically varied the presence of these cues and examined their effects on second-syllable stress. In Experiment 1, we examined the effect of prefixation and vowel length by presenting participants with four sets of nonwords varied factorially on the inclusion of a prefix and the length of the second vowel (e.g., prelel, pralel, preleal, praleal). In Experiment 2, we examined the effects of prefixation and orthographic weight by presenting participants with four sets of nonwords varied factorially on the inclusion of a prefix and the number of letters contained in the second syllable (e.g., prelel, pralel, prelell, pralell). In Experiment 3, we reexamined the effects of prefixation and vowel length while controlling for orthographic weight. We achieved this by presenting participants with four sets of nonwords that contained the same number of letters in the second syllable and were varied factorially on the inclusion of a prefix and the length of the second vowel (e.g., prelell, pralell, preleal, praleal). This experimental approach allowed us to ascertain the influence of each one of these sublexical cues, independently and in combination with another cue, on the assignment of second-syllable stress.
Experiment 1: Prefixation and Vowel Length
Experiment 1 examined the effects of prefixation and vowel length on stress assignment during nonword reading aloud. We predicted that readers would be more likely to assign second-syllable stress to prefixed nonwords compared with non-prefixed nonwords, and to nonwords with a long vowel in their second syllable compared with nonwords with a short vowel in their second syllable.
Twenty undergraduate students from Royal Holloway, University of London, were paid £5 to participate in the study. Participants were monolingual native speakers of Southern British English and reported no visual, reading, or language difficulties.
Stimuli and design
Stimuli comprised 80 phonotactically legal nonwords, ranging in length from five to seven letters. All nonwords were disyllabic with at least one medial consonant flanked by two vowels. Nonwords were varied factorially on (a) the inclusion of a prefix and (b) the length of the second vowel, thus yielding 20 prefixed nonwords with a short vowel in the second syllable (prefixed short vowel), 20 prefixed nonwords with a long vowel in the second syllable (prefixed long vowel), 20 non-prefixed nonwords with a short vowel in the second syllable (non-prefixed short vowel), and 20 non-prefixed nonwords with a long vowel in the second syllable (non-prefixed long vowel).
Nonword construction was performed in such a way so that items were pairwise matched between the different conditions, with prefixes (de-, mis-, pre-, re-) changing into non-prefixes (do-, mes-, pra-, ro-) and short vowels (a, e, o, u) changing into long vowels (ai, ea/ee, oa, oo/ou) in a consistent manner. For example, nonwords beginning with the prefix pre- in the prefixed conditions would be matched to nonwords beginning with the letter sequence pra- in the non-prefixed conditions (e.g., the prefixed item prelel was matched to the non-prefixed item pralel). Likewise, nonwords containing the short vowel “e” as a second vowel in the short-vowel conditions would be matched to nonwords containing the long vowel “ea” or “ee” as a second vowel in the long-vowel conditions (e.g., the items prelel and pralel were matched to the items preleal and praleal, respectively). This pairwise item matching was undertaken to ensure that differences across conditions were minimized apart from the experimental manipulations of interest. Also, stimuli were constructed in such a way that prefixes and their control orthographic counterpart units were likely to overlap with the first syllable of the nonword (e.g., in nonwords such as misdut and mesdut, the letter sequence “sd” results in a phonotactically illegal second-syllable onset, hence these nonwords would be most likely syllabified as mis.dut and mes.dut, respectively). Finally, care was taken to ensure that neither of the two syllables of the nonwords contained a monosyllabic English word.
Nonwords in the four conditions were group-wised matched on orthographic neighborhood size (Coltheart’s N; Coltheart, Davelaar, Jonasson, & Besner, 1977): prefixed short vowel (M = 0.15, SD = 0.37); prefixed long vowel (M = 0, SD = 0); non-prefixed short vowel (M = 0.10, SD = 0.31); non-prefixed long vowel (M = 0.05, SD = 0.22); F(1, 76) < 2.87, p > .05). Further, nonwords across the four conditions did not differ in terms of the number of their orthographic neighbors that take second-syllable stress, F(1, 76) < 2.11, p > .05.
Nonwords were further divided into four blocks of 20 experimental trials, so that all experimental conditions were equally represented in each block but no pairwise-matched nonwords appeared in the same block. Nonwords in each block were presented in a random order, while the order of presentation of the blocks was counterbalanced across participants. Following a practice session of five trials, each participant was presented the four blocks. The nonword stimuli used in Experiment 1 are listed in Appendix A.
Apparatus and procedure
Participants were tested individually in a quiet room. Each participant sat approximately 40 cm in front of a CRT monitor. Stimulus presentation and data recording were controlled by DMDX software (Forster & Forster, 2003), and verbal responses were recorded by a head-worn microphone. Nonword stimuli were displayed in white on a black background in a 14-point Courier New font. Each trial began with the presentation of a fixation cross in the center of the screen for 500 ms. The fixation cross was replaced at the same location with a nonword stimulus, which was displayed for 3,000 ms. Participants were asked to read aloud the nonword into the microphone as quickly and as clearly as possible, as if it were a real word. The next trial followed after a 850-ms blank interval.
The analyses included responses to a total of 76 items per participant (due to an oversight, a group of four pairwise-matched stimuli was not presented to the participants). Nonword responses were excluded if they were pronounced with anything other than two syllables, or if they were characterized by hesitations or other articulatory dysfluencies (1.8% of the data). The remaining responses were classified as being stressed on the first or second syllable. Stress judgments were undertaken by one of the authors (K.R.) who had previous training and experience in such a task (Mousikou et al., 2017; Rastle & Coltheart, 2000).2 The proportions of second-syllable stress in the four different conditions are presented in Table 1.
Human and Model Mean Proportions of Second-Syllable Stress as a Function of Prefixation and Second-Vowel Length in Experiment 1 (Mean Standard Error for the Human Data in Parentheses)
In this and all following experiments, we analyzed the impact of stimulus factors on stress assignment (a binary variable) using generalized linear mixed-effects models. These analyses were performed using the packages lme4 (Bates, Maechler, Bolker, & Walker, 2015), car (Fox & Weisberg, 2011), multcomp (Hothorn, Bretz, & Westfall, 2008), and lsmeans (Lenth, 2016) implemented in the statistical software R (Version 3.3.1; R Core Team, 2016). In each experiment, a logit mixed model (Jaeger, 2008) was created using stepwise backward model comparison and model selection was performed on the basis of chi-squared log-likelihood ratio tests. The significance of the fixed effects was determined with Type III model comparisons using the Anova function provided by the car package.
Nonword items were matched across conditions on orthographic neighborhood size (Coltheart’s N). However, many of our items had neighborhoods of zero on this measure. Hence, we included two more sensitive measures of orthographic similarity, namely the OLD20 (Yarkoni et al., 2008) and mean bigram type frequency, as covariates in the analyses. Furthermore, given the evidence that the endings of words can carry orthographic cues to stress assignment in English (Arciuli & Cupples, 2006), we included two additional covariates in our analyses, the total number of disyllabic words in CELEX that share the same endings as our stimuli (ending frequency) and the proportion of these words that take second-syllable stress (ending-to-second-syllable-stress proportion). Following Arciuli and Cupples (2006), word ending was defined as the orthographic unit that includes all the final letters of a word starting from the nucleus of the second syllable (e.g., -ow in follow, -ark in embark, and -upt in erupt). These steps were taken to ensure that any differences between our conditions of interest would not be due to the frequency of occurrence of the items’ sublexical orthographic units in the lexicon, or to the association of these units with second-syllable stress in English disyllabic words. However, in the following analyses, we focus only on the results pertaining to our factors of interest.
Our model included stressed syllable (first vs. second) as the dependent variable, prefixation (non-prefixed vs. prefixed), second-vowel length (short vs. long) and their interaction as fixed effects, and participants and items as crossed random effects. OLD20, χ2(1) = 7.36, p = .007, and ending-to-second-syllable-stress proportion, χ2(1) = 12.83, p < .001, were the only covariates that contributed significantly to the model’s goodness of fit and were thus included in the final model.
Results revealed a main effect of prefixation, χ2(1) = 16.78, p < .001, with prefixed nonwords receiving more second-syllable stress than non-prefixed nonwords. As predicted by the fitted logistic regression model, the probability of second-syllable stress for non-prefixed items was 28%, whereas the corresponding probability for prefixed items was 45%. There was also a main effect of second-vowel length, χ2(1) = 32.00, p < .001, with nonwords containing a long second vowel receiving more second-syllable stress than those with a short second vowel. The predicted probability of second-syllable stress for items with a short second vowel was 23%, whereas the corresponding probability for items with a long second vowel was 52%. The interaction between prefixation and vowel length was not significant, χ2(1) = 0.17, p = .679.
Results revealed that both prefixation and vowel length influenced the assignment of stress in nonword reading aloud. Readers were more likely to assign second-syllable stress to disyllabic nonwords that contained a prefix compared with non-prefixed nonwords. Similarly, readers were more likely to assign second-syllable stress to disyllabic nonwords with a long second vowel compared with nonwords with a short second vowel. The lack of a significant interaction between prefixation and vowel length suggests that these sublexical cues constitute independent sources for predicting the assignment of stress. The additive effects of prefixation and vowel length can be clearly seen in the prefixed long vowel condition, where prefixed nonwords with a long second vowel (e.g., preleal) received the highest proportion of second-syllable stress.
In this experiment, however, the influence of vowel length was confounded with the potential effect of another sublexical cue, that is, the orthographic weight of a syllable. This is because nonwords with a long second vowel (e.g., praleal, preleal) also contained more letters in their second syllable compared to nonwords with a short second vowel (e.g., pralel, prelel), as is typical in English spelling-to-sound mappings. Therefore, it is possible that the observed effect of vowel length is driven by the combination of vowel length and orthographic weight, or that it simply reflects a pure effect of orthographic weight, with syllables containing more letters being more likely to attract stress. Following the same factorial design, Experiment 2 was designed to establish whether the orthographic weight of a syllable influences the assignment of stress and whether it interacts with the observed effect of prefixation.
Experiment 2: Prefixation and Orthographic Weight
Experiment 2 examined the effects of prefixation and the orthographic weight (i.e., number of letters) of the second syllable on stress assignment during nonword reading aloud. To avoid confounds with potential effects of vowel length on stress assignment (as those observed in Experiment 1), all nonword stimuli contained a short vowel. We predicted that readers would be more likely to assign second-syllable stress to prefixed nonwords compared with non-prefixed nonwords, and to nonwords that contain more letters in their second syllable compared with nonwords that contain fewer letters in their second syllable.
Twenty newly recruited undergraduate students from Royal Holloway, University of London, were paid £5 to participate in the study. Participants were monolingual native speakers of Southern British English and reported no visual, reading, or language difficulties.
Stimuli and design
Stimuli comprised 80 phonotactically legal disyllabic nonwords, which ranged in length from five to seven letters. Nonwords were varied factorially on (a) the inclusion of a prefix and (b) the orthographic weight of the second syllable (operationalized as containing three or four letters), yielding 20 prefixed nonwords with a short second syllable (prefixed light syllable), 20 prefixed nonwords with a long second syllable (prefixed heavy syllable), 20 non-prefixed nonwords with a short second syllable (non-prefixed light syllable), and 20 non-prefixed nonwords with a long second syllable (non-prefixed heavy syllable). Nonwords with a light second syllable were the same stimuli as those used in the prefixed and non-prefixed short vowel conditions in Experiment 1. Nonword stimuli with a heavy second syllable were pairwise matched to those with a light second syllable, and were constructed by adding an additional consonant letter to their second syllable. Specifically, added letters that represent a single phoneme were placed either at the beginning or at the end of the second syllable to form a digraph such as “ph,” “sh,” and “th,” or a double consonant cluster such as “ng,” “ss,” and “ll.” For example, the item prelel in the prefixed light syllable condition was matched to the item prelell in the prefixed heavy syllable condition, the item pralel in the non-prefixed light syllable condition, and the item pralell in the non-prefixed heavy syllable condition. Thus, all nonwords had the same number of phonemes and consisted of a short vowel in their second syllable. As with Experiment 1, care was taken so that prefixes and their control orthographic counterpart units comprised the first syllable of the nonword. Further, neither of the two syllables of the nonwords corresponded to an English monosyllabic word.
Nonwords in the four conditions were group-wise matched on Coltheart’s N: prefixed light syllable (M = 0.15, SD = 0.37); prefixed heavy syllable (M = 0, SD = 0); non-prefixed light syllable (M = 0.10, SD = 0.31); non-prefixed heavy syllable (M = 0.05, SD = 0.22); F(1, 76) < 2.87, p > .05. Further, nonwords in the four conditions did not differ in terms of the number of their orthographic neighbors that take second-syllable stress, F(1, 76) < 2.11, p > .05.
As in Experiment 1, nonwords were divided into four blocks of 20 items, so that all experimental conditions were equally represented but no pairwise-matched nonwords appeared in the same block. Nonwords in each block were presented in a random order, whereas the presentation order of each block was counterbalanced across participants. Following a practice session of five trials, each participant was presented the four blocks. The nonword stimuli used in Experiment 2 are listed in Appendix B.
Apparatus and procedure
The apparatus and procedure were the same as in Experiment 1.
Nonword responses were classified as having stress on the first or second syllable, while hesitations and responses with anything other than two syllables were marked as erroneous and discarded (1.1% of all responses). Table 2 displays the proportion of second-syllable stress across the four conditions.
Human and Model Mean Proportions of Second-Syllable Stress as a Function of Prefixation and the Orthographic Weight of the Second Syllable in Experiment 2 (Mean Standard Error for the Human Data in Parentheses)
As for Experiment 1, stress assignment data were analyzed using a generalized linear mixed-effects model. Our model included stressed syllable (first vs. second) as the dependent variable, prefixation (non-prefixed vs. prefixed), orthographic weight of the second syllable (light vs. heavy) and their interaction as fixed effects, and participants and items as crossed random effects. OLD20, χ2(1) = 8.55, p = .003, and mean bigram frequency, χ2(1) = 19.19, p < .001, contributed significantly to the model’s goodness of fit and were thus included in the final model as covariates.
Results revealed a main effect of prefixation, χ2(1) = 15.64, p < .001, with prefixed nonwords receiving more second-syllable stress than non-prefixed nonwords. As predicted by the fitted logistic regression model, the probability of second-syllable stress for non-prefixed items was 10%, whereas the corresponding probability for prefixed items was 22%. There was also a main effect of orthographic weight, χ2(1) = 8.46, p = .004, with nonwords containing a heavy second syllable receiving more second-syllable stress than those with a light second syllable. The predicted probability of second-syllable stress for items with a light second syllable was 10%, whereas the corresponding probability for items with a heavy second syllable was 22%. The interaction between prefixation and orthographic weight was not significant, χ2(1) = 0.55, p = .460.
Correlational analyses examined the consistency with which participants assigned stress to the subsets of nonword items that appeared both in Experiments 1 and 2 (i.e., the nonwords in the prefixed and non-prefixed short vowel conditions of Experiment 1 and the nonwords in the prefixed and non-prefixed light syllable conditions of Experiment 2). Results established high item-based reliability (r = .73, p < .001) on the assignment of stress for these nonwords across the two experiments.
Our results from Experiment 2 revealed an effect of prefixation on the assignment of stress in nonword reading aloud. Readers were more likely to assign second-syllable stress to disyllabic nonwords that contained a prefix than to non-prefixed nonwords, replicating the results of Experiment 1. The results from Experiment 2 also showed an effect of orthographic weight on stress assignment, as readers were more likely to assign second-syllable stress to disyllabic nonwords that contained a higher number of letters in their second syllable compared with nonwords that contained fewer letters in their second syllable. The effects of prefixation and orthographic weight were independent, as evidenced by the lack of a significant interaction between these two sublexical cues. These results confirm that orthographic weight is an independent cue to stress assignment. In Experiment 3 we sought to determine whether this was also the case for vowel length.
Experiment 3: Prefixation and Vowel Length With Orthographic Weight Controlled
Experiment 3 reexamined the effects of prefixation and vowel length of the second syllable on stress assignment during nonword reading aloud, after controlling for the effect of orthographic weight of the second syllable. In this experiment, all nonword stimuli contained the same number of letters in their second syllable. We predicted that readers would be more likely to assign second-syllable stress to prefixed nonwords compared with nonprefixed nonwords, and to nonwords with a long vowel in their second syllable compared with nonwords with a short vowel in their second syllable.
Twenty newly recruited undergraduate students from Royal Holloway, University of London, were paid £5 to participate in the study. Participants were monolingual native speakers of Southern British English and reported no visual, reading, or language difficulties.
Stimuli and design
Stimuli comprised 80 disyllabic nonwords, ranging in length from six to seven letters. All nonwords had the same number of letters (and phonemes) in their second syllable and were varied factorially on (a) the inclusion of a prefix and (b) the length of the second vowel. Nonwords with a short vowel in the second syllable (prefixed short vowel and non-prefixed short vowel) were the same as those used in Experiment 2 in the prefixed heavy syllable condition (e.g., prelell) and the non-prefixed heavy syllable condition (e.g., pralell), respectively. Nonwords with a long vowel (prefixed long vowel and non-prefixed long vowel) were the same as those used in Experiment 1 in the prefixed long vowel condition (e.g., preleal) and the non-prefixed long vowel condition (e.g., praleal), respectively. Because of the way nonwords were constructed in Experiments 1 and 2, items were already pairwise matched between the different conditions of Experiment 3.
Nonwords across the four conditions were group-wise matched on Coltheart’s N: prefixed short vowel (M = 0, SD = 0); prefixed long vowel (M = 0, SD = 0); non-prefixed short vowel (M = 0.05, SD = 0.22); non-prefixed long vowel (M = 0.05, SD = 0.22); F(1, 76) < 2.00, p > .05. None of the nonwords had orthographic neighbors that take second-syllable stress.
Following the same rationale as in Experiments 1 and 2, nonwords were divided into four blocks of 20 items. All experimental conditions were equally represented but no pairwise-matched nonwords appeared in the same block. Nonwords in each block were presented in a random order, while the presentation order of each block was counterbalanced across participants. Following a practice session of five trials, each participant was presented the four blocks. The nonword stimuli used in Experiment 3 are listed in Appendix C.
Apparatus and procedure
The apparatus and procedure were the same as in Experiments 1 and 2.
Disyllabic responses were classified as having received stress on the first or the second syllable, while hesitations and responses with anything other than two syllables were marked as erroneous and discarded (2.4% of all responses). Table 3 displays the proportion of second-syllable stress responses across the four conditions.
Human and Model Mean Proportions of Second-Syllable Stress as a Function of Prefixation and Second-Vowel Length, With the Orthographic Weight of the Second Syllable Controlled, in Experiment 3 (Mean Standard Error for the Human Data in Parentheses)
As for Experiments 1 and 2, stress assignment data were analyzed using generalized linear mixed-effects model. Our model included stressed syllable (first vs. second) as the dependent variable, prefixation (non-prefixed vs. prefixed), second-vowel length (short vs. long) and their interaction as fixed effects, and participants and items as crossed random effects. OLD20, χ2(1) = 15.71, p < .001, contributed significantly to the model’s goodness of fit and was thus included in the final model as a covariate.
Results revealed a main effect of prefixation, χ2(1) = 11.90, p < .001, with prefixed nonwords receiving more second-syllable stress than non-prefixed nonwords. As predicted by the fitted logistic regression model, the probability of second-syllable stress for non-prefixed items was 30%, whereas the corresponding probability for prefixed items was 52%. There was also a main effect of second-vowel length, χ2(1) = 6.63, p = .010, with nonwords containing a long second vowel receiving more second-syllable stress than those with a short vowel in the second syllable. The predicted probability of second-syllable stress for items with a short second vowel was 33%, whereas the corresponding probability for items with a long second vowel was 48%. The interaction between prefixation and vowel length was not significant, χ2(1) = 0.57, p = .449.
Correlational analyses examined stress consistency between the subsets of nonwords that appeared both in Experiments 1 and 3 (i.e., the nonwords in the prefixed and non-prefixed long vowel conditions of Experiments 1 and 3) and between the subsets of nonwords that appeared both in Experiments 2 and 3 (i.e., the nonwords in the prefixed and non-prefixed heavy syllable conditions of Experiment 2 and the nonwords in the prefixed and non-prefixed short vowel conditions of Experiment 3). Results revealed a high degree of item-based reliability (r = .73, p < .001 and r = .78, p < .001, respectively) on the assignment of stress for these nonwords across the three experiments.
The results from Experiment 3 replicated the findings from Experiments 1 and 2 by revealing an effect of prefixation on the assignment of stress. Readers were more likely to assign second-syllable stress to prefixed nonwords than non-prefixed nonwords. The results from the present experiment also established an effect of vowel length on stress assignment. Readers were more likely to assign second-syllable stress to nonwords with a long vowel in the second syllable than nonwords with a short second vowel. The effect of vowel length was independent of the effect of prefixation, as evidenced by the lack of a significant interaction between these two sublexical cues. Furthermore, in the current experiment, the effect of vowel length remained significant even after controlling for the orthographic weight of the second syllable, which was found to influence second-syllable stress assignment in Experiment 2. This finding provides clear evidence that the effect of vowel length is independent of the number of letters present in a syllable, and that vowel length constitutes a sublexical cue to stress in its own right.
Using the nonword stimuli presented in Experiments 1, 2, and 3, we sought to assess the performance of the RC00 algorithm (Rastle & Coltheart, 2000), the CDP++ model (Perry et al., 2010), and the SMA09 network (Ševa et al., 2009). First, in order to obtain a general index of the models’ stress assignment performance in relation to that of the human readers, we assessed the success of each model in capturing the modal stress assigned to each of the nonwords presented in Experiments 1, 2, and 3 across experiments and participants. Second, in order to determine whether the rule-based or the connectionist approach best captures the specific sensitivities that human readers show to the sublexical stress cues under investigation, we compared the human stress data obtained separately from Experiments 1, 2, and 3 with simulation results from these three accounts of disyllabic reading in English. The identification of a prefix by the RC00 algorithm unambiguously results in second-syllable stress. Hence we predicted that this model would be more likely to assign second-syllable stress to prefixed nonwords compared with non-prefixed nonwords. However, this algorithm contains no explicit rules regarding the sublexical cues of orthographic weight and vowel length, and so we did not expect it to be sensitive to these cues. The CDP++ model and the SMA09 network adopt a statistical-learning approach that does not allow the explicit formulation of hypotheses in respect of the sublexical cues under investigation. Hence, we made no predictions about the performance of these models.
Model Performance in Capturing Modal Human Stress Assignment
We calculated the modal stress produced by human participants for each of the 120 items that were presented across Experiments 1, 2, and 3. Participants assigned first-syllable stress to 75% of the items and second-syllable stress to 25% of the items, mirroring the distribution of stress in English disyllables (Baayen et al., 1995). To test the models’ performance against the modal human stress, we performed a binary classification analysis. Adopting the same approach as Ševa et al. (2009), we assessed each model’s sensitivity in assigning first and second-syllable stress by using the d′ measure, which is calculated by taking into account both the model’s correct classifications as well as its misclassifications. We further obtained a measure of response bias (c), indicating whether a model was biased toward first- or second-syllable stress assignment. An increase in the absolute value of d′ and c coefficients would indicate an increase in the model’s stress assignment sensitivity and response bias, respectively, with a response bias toward first-syllable stress being denoted by a negative value. The distribution of stress pattern assigned by human participants and the stress pattern predicted by the models are reported in Table 4.
Contingency Tables Showing the Distribution of Stress Pattern Assigned by Human Readers and Predicted by the RC00 Algorithm, the CDP++ Model, the SMA09 Network, and the Bayesian Account to the Items Presented in Experiments 1, 2, and 3
The RC00 algorithm predicted the modal human stress for 60% of the items that were assigned first-syllable stress and for 77% of the items that were assigned second-syllable stress (d′ = 0.98), and revealed a response bias toward second-syllable stress (c = .24). The CDP++ model correctly stressed 88% and 53% of the items that received first-syllable and second-syllable stress, respectively, by the majority of human participants (d′ = 1.25), and showed a response bias toward first-syllable stress (c = −.54). Finally, the SMA09 network predicted first and second-syllable stress for 64% and 53% of the items (d′ = 0.45), respectively, in accordance with the modal human stress, and showed a small response bias toward first-syllable stress (c = −.14). This analysis thus suggests that the CDP++ model performed slightly better than the other two models in predicting human stress assignment.
Model Sensitivity to Sublexical Cues
To ascertain whether the models were sensitive to the same sublexical cues as human readers, the simulation data for each model were submitted to logistic regression analyses. These analyses assessed the probability of second-syllable stress (a binary variable) occurring as a function of the binary variables of prefixation and second-vowel length (Experiments 1 and 3), and the binary variables of prefixation and orthographic weight (Experiment 2). The results from the logistic regression analyses are reported in Table 5.3
1. Historical landmarks
Gill (1619). Cap. XXV-XXVI.
Steele (1775) distinguished three levels of stress.
Mr. William Archer, after a long list of seemingly arbitrary accentuations in the English language (America To-Day, p. 193), goes on to say: `But the larger our list of examples, the more capricious does our accentuation seem, the more evidently subject to mere accidents of fashion. There is scarcely a trace of consistent or rational principle in the matter.' It will be the object of the following pages to show that there are principles, and that the `capriciousness' is merely the natural consequences of the fact that there is not one single principle, but several principles working sometimes against each other. (p. 160)Kingdon (1958) distinguished a) "Romanic-type compounds" (derived Latinate words) b) "Greek-type compounds" (like lexical compounds) c) "English-type compounds" (lexical compounds).
Chomsky, Halle & Lukoff (1956), (summary account in Chomsky and Miller 1963):
These rules ... are ordered, and apply in a cycle, first to the smallest constituents (that is, lexical morphemes), then to the next larger ones, and so on, until the largest domain of phonetic processes is reached .... essentiallly the same rules apply both inside and outside the word. Thus ... a single cycle of transformational rules ... by repeated application, determines the phonetic structure of a complex form.a) A substantive rule that assigns stress in initial position in nouns (also stems) under very general circumstances. [Germanic stress].
b) A nuclear stress rule that makes the last main stress dominant, thus weakening all other stresses in the construction.
c) The vowel reduction rule.
d) A rule of stress adjustment that weakens all nonmainstresses in a word by one.
2. Chomsky & Halle (1968), especially chs. 2-3. The fulfillment of their concerted effort to determine a complete set of rules for English phonology, dominated by stress assignment and its consequences.
2.1. Stress placement is sensitive to [syllable] weight
p. 29 weak cluster - "a string consisting of a simple vocalic nucleus followed by no more than one consonant".
strong cluster - "a string consisting of either a vocalic nucleus followed by two or more consonants or a complex vocalic nucleus followed by any number of consonants." (i.e. light vs. heavy rimes).
2.2. Stress rules are sensitive to lexical categories
(21) Main Stress Rule
V -> [1 stress] / X - C0]NAV
Rule (21) assigns primary stress to the final vowel of the word, e.g. eváde, supréme, exíst, absúrd, all of which end in a strong cluster. If a verb or adjective has a final weak cluster, the stress is placed on the penultimate syllable e.g. rélish, cóvet, devélop, stólid, cómmon, clandéstine.
2.3. Sensitivity to the Latinate/Germanic distinction
a) Germanic affixes don't affect stress placement, e.g. éarth, éarthly, unéarthly, unéarthliness.
b) Latinate suffixes may affect stress placement, e.g. témpest, tempéstuous, tempestuósity.
c) Stress may shift onto Latinate prefixes, e.g. invést vs. ínverse, càtatónic vs. catastrophe.
d) Stress never shifts onto Germanic prefixes, e.g. òver[cóok] - no forms like ovéric [oUvrIk]
Segmentally homophonous Latinate and Germanic suffixes with different stress behaviours:
2.4. Sensitivity to morpheme boundaries
pérson+al, not persónal SW+al
theátric+al not theatrícal WSW+al
Phonological operations restricted to specific morphological domains / environments:
|im-||-press||-ion||over- [im-||-press||-ion] -able|
2.5. Stress subordination
p. 64 "The rules that determine stress contours are, for the most part, rules that assign primary stress in certain positions, at the same time weakening the stresses in all other positions".
2.5. Alternating stress rule (p.78) (for secondary stresses)
V -> [1 stress] / - C0 V C0 V1 C0]
hurricAn => hurricA1n [Main stress rule]
=> hu1rricA1n [Alternating stress rule]
=> hu1rricA2n [Stress subordination]
3. Liberman and Prince (1977)
"certain features of prosodic systems like that of English, in particular the phenomenon of `stress subordination', are not to be referred primarily to the properties of individual segments (or syllables), but rather reflect a hierarchical rhythmic structuring that organizes the syllables, words, and syntactic phrases of a sentence."
"only a stressed syllable may be the strong element of a metrical foot".
Digression on Iambic Reversal / Rhythm Rule / Stress Retraction
SPE and Metrical Phonology both capture the preservation of relative prominence under embedding. But there are counter-cases,
e.g. thirtéen vs. thìrteen mén, àchromátic vs. áchromàtic lêns
Liberman and Prince: "we need an account of linguistic rhythm in terms of which the appropriate stress configurations are marked as `clashing', thus producing a pressure for change."
The Metrical Grid:
For more on the Rhythm rule, see Kager and Visch (1988), and papers by Grabe and Warren, Vogel et al. and Shattuck-Hufnagel in
Connell and Arvaniti (1995). According to these later studies, iambic reversal is not a phonological movement rule at all, bur arises from the interaction of lexical stress and phrase-final accent:
4. Kiparsky (1979): stress assignment is cyclic.
Liberman and Prince (1977) derive the stress of "sensationality" incorrectly:
|2||3||1||- according to Liberman and Prince's stress-marking algorithm|
"Since no cyclic rules in [Liberman & Prince 1977] are sensitive to metrical structure, one could equivalently stipulate that metrical structure is assigned only on the last cycle."
If assignment of metrical structure is cyclic (i.e. derivation respects structure built on earlier cycles) the (correct) derivation will be:
"metrical structure assigned in earlier cycles is kept insofar as it is not redrawn by the reapplication of [foot construction]."
5. Hayes (1982): extrametricality
i) The final syllable is extrametrical in nouns and suffixed adjectives, e.g. serendipi<ty>, sensa<tion>, perso<nal>. With this proviso, antepenultimate stress is eliminated. Word final syllables are stressed if heavy, otherwise stress is penultimate (modulo extrametricality).
ii) The final consonant is extrametrical in underived verbs and adjectives, e.g. soli<d>, supre<me>.
6. English stress parameters
As work in metrical theory progressed and was extended to many languages, our conception of English stress assignment became embedded in, and was constrained by, a parametric view of options for metrical structure (Halle and Vergnaud 1987, Booij 1983), taking a lead from Chomskyan syntax.
1) Principles: Words consist of feet and feet consist of syllables.
2) Parameter: In English the rightmost foot is strongest (domain of main stress) e(ráse), i(ráte), mu(tátion), (ècu)(méni)<cal>, (ànti)(dìse)(stàblish)<men>(tári)<an>. Cf. Russian: leftmost - (úa)(sàm).
3) Parameter: Bounded feet are maximally binary; ternary feet are dealt with by extrametricality, and can only ocur at the edges of words (or cycles). Thus: (Hàma)(mèlid)(ánthe)<mum>. Cf. unbounded feet in Khalkha Mongolian (xötElbr) `leadership', French (originalité).
4) Parameter: In words with an odd number of syllables (excepting extrametrical ones), left-over syllables occur at the beginning; e.g. a(génda), To(péka), a(ríse). In Maranungku they occur at the end e.g. (lángka)(ràte)tì.
5) Parameter: In English, non-tonic binary feet immediately precede the tonic foot. Problem: (àbra)ca(dábra).
6) Parameter: The leftmost syllable in a foot is strongest e.g. (mán), (mánner), (mána)<ger>. Cf. Weri (kù)(lipú), (ulù)(amít), (á)(kunè)(tepál), in which the rightmost syllable is strongest.
7) Parameter: Feet are quantity-sensitive i.e. a heavy syllable must be the head of a foot. (Does not exclude possibility that light syllables could be syllable heads too) e.g. (fùnda)(méntal), (spíral) but in(ért).
8) Paramter: Secondary stresses are placed from right to left on the head of every foot other than the tonic foot. (Stress assignment is iterative. Cf. Spanish, Polish.) e.g. (hàma)(mèli)(dánthe)<mum>.
7. The Abracadabra problem.
(àbra)ca(dábra), Kalamazoo, Luxipalilla, Hardecanute, okefenokee, Nebuchadnezzar, paraphernalia, Kilimanjaro: these words appear to require either a medial extrametrical syllable or a ternary foot.
Hammond's solution: secondary feet are built left to right. Following Hammond's suggestion, see also Halle and Kenstowicz (1991) section 7, McCarthy and Prince (1993) section 3.
8. Phrasal stress
Roca and Johnson follow Chomsky and Halle (1968) in contrasting e.g. lexical bláckbìrd vs. phrasal blàck bírd. But we cannot in general state that phrasal stress is right-headed. Noun-noun sequences are usually left-headed (e.g. dóg-hoùse), but there are exceptions (e.g. lòbster ragóut).
Booij, G.E. (1983) Principles and parameters in prosodic phonology. Linguistics 21, 249-280.
Chomsky, N.& M. Halle (1968) The Sound Pattern of English. New York: Harper and Row. Reprinted in 1991 by MIT Press.
Chomsky, N., M. Halle and F. Lukoff (1956) On Accent and Juncture in English. In M. Halle, H.G. Lunt, H. McLean and C.H. van Schooneveld, eds. For Roman Jakobson: Essays on the occasion of his sixtieth birthday. The Hague: Mouton & Co. 65-80.
Chomsky, N.& G.A. Miller (1963) Introduction to the Formal Analysis of Natural Languages. In R.D. Luce, R.R. Bush and E. Galanter, eds. Handbook of Mathematical Psychology Volume II, New.York: John Wiley. 269-321.
Connell, B. and A. Arvaniti (1995) Phonology and Phonetic Evidence: Papers in Laboratory Phonology IV. Cambridge University Press.
Gil, A. (1619) Logonomia Anglica. [Scolar Press facsimile reprint, 1967]. Or see Alexander Gill's Logonomia Anglica (1619), Part II: Biographical and Bibliographical Introductions. Notes by Bror Danielson and Arvid Gabrielson. Translation by Robin C. Alston. Stockholm Studies in English XXVII. Stockholm: Almqvist and Wiksell.
Halle, M.& M. Kenstowicz (1991) The Free Element Condition and Cyclic versus Noncyclic Stress. Linguistic Inquiry 22(1), 457-501.
Halle, M. and J.-R. Vergnaud (1987) An Essay on Stress. MIT Press.
Hayes, B (1981) A Metrical Theory of Stress Rules. IULC.
Jesperson, O. (1909/1954) A Modern English Grammar on Historical Principles. Part I: Sounds and Spellings. London: George Allen and Unwin.
Kager, R. & E. Visch (1988), Metrical constituency and rhythmic adjustment. Phonology 5.1, 21-71.
Kingdon, R. (1958) The Groundwork of English Stress. London: Longman.
Kiparsky, P. (1979) Metrical Structure Assignment is Cyclic. Linguistic Inquiry 10(3), 421-441.
Liberman, M. and A. Prince (1977) On Stress and Linguistic Rhythm. Linguistic Inquiry 8(2), 249-336.
McCarthy, J. and A. Prince (1993) Generalized Alignment. Yearbook of Morphology 1993. 79-153.
Steele, J. (1975) An Essay towards Establishing the Melody and Measure of Speech. [Scolar Press Facsimile edition, 1969].
Determine the stress of each syllable and the quality of each vowel in the following examples. (Refer to a pronouncing dictionary if you are not a native speaker of English.) Parse the words into prefixes, stems and suffixes. In each case, account for the alternations of stress and vowel quality in the prefix.