Journal of Research in Reading

In languages with lexical stress, in which stress position can vary and contributes to lexical identity, stress assignment is a necessary component of successful reading aloud. Presumably, each lexical entry includes stress information, otherwise it could not be pronounced correctly. But this does not imply that stress assignment in reading must be based on the lexically stored information only. Indeed, studies have shown that additional sources of stress information are taken into account, observed most clearly in pseudoword reading. For example, orthographic stress information in the form of a special diacritic affects stress assignment when available, such as in the Greek language (Protopapas, 2006; Protopapas, Gerakaki, & Alexandri, 2006, 2007; Protopapas & Gerakaki, 2009). Dominant or default patterns also affect stress assignment in languages such as Italian and Greek (Colombo, 1992; Burani, Paizi, & Sulpizio, 2014; Sulpizio & Colombo, 2013; Protopapas, 2006; for reviews, see Protopapas, 2016; Sulpizio, Burani, & Colombo, 2015).

Indirect lexical effects on stress assignment in reading have also been documented, both from individual words and from word groups. Specifically, in Greek, pseudowords preferentially receive the stress patterns of similar words when closely resembling them (i.e., when they share most of their letters), across ages (Protopapas et al., 2006, 2007; the effect being strongest in the early elementary grades; Protopapas & Gerakaki, 2009). In Italian, pseudowords preferentially receive the majority stress patterns of word groups that are similar (i.e., written with the same letters) in their final two syllables (termed “stress neighbourhoods”; e.g., Burani & Arduino, 2004; Burani et al., 2014; Colombo, 1992; see Sulpizio et al., 2015, for review). Further removed from specific words, significant statistical associations between stress positions and orthographic patterns, particularly word beginnings and (especially) endings, have been documented in several languages (Arciuli, Monaghan, & Seva, 2010; Monaghan, Arciuli, & Seva, 2016; Ševa, Monaghan, & Arciuli, 2009), making up an additional, probabilistic source of stress assignment information. These associations appear to be informative in stress assignment processes, at least in English (Arciuli & Cupples, 2006; Arciuli et al., 2010; Wade‐Woolley & Heggie, 2015). Indeed, word beginnings and endings associated with particular stress patterns have been used to model stress assignment in reading multi‐syllabic English words in both connectionist and rule‐based computational implementations (Arciuli et al., 2010; Rastle & Coltheart, 2000).

However, in most of these languages, there are no specific cues to stress (i.e., no diacritic) and no general default pattern. The question then arises: would these weak statistical associations affect actual stress assignment behaviour when reading in a language with strong and specific sources of stress information? Although one might reasonably expect that reliable and specific sources would dominate stress assignment behaviour and swamp minor statistical associations, there are also reasons to expect otherwise, as Greek readers have been found to not rely entirely on the stress diacritic despite its reliable presence and excellent cue validity (Protopapas, 2016). This question is of major theoretical importance because it concerns the long‐debated notion of probabilistic versus rule‐based cognitive systems. If sensitivity to reliable but relatively minor statistical associations can be demonstrated in the face of well‐established alternatives, then the appeal of positing absolute rules (and thereby also defining “exceptions” to these rules be “irregular”) may be diminished.

Greek offers the opportunity to address this question in conjunction with a related morphological hypothesis coming from linguistics. Specifically, on the one hand, there are documented weak but significant statistical associations between initial and final letter groups and stress positions (Monaghan et al., 2016). On the other hand, there is a family of derivational suffixes that, according to linguistic analysis, carry their own stress pattern (being either accented or unaccentable) and impose it on the derived words they form, overriding any stress patterns that may be associated with the roots they attach to (Revithiadou, 1999). For example, the suffix ‐άδα /'a.ða/ combines with nouns or adjectives to form nouns, such as πoρτoκάλι /por.to.'ka.li/ “orange” → πoρτoκαλάδα /por.to.ka.'la.ða / “orange juice.” Similar suffixes exist in English and have been found to modulate stress assignment preference in pseudowords (Wade‐Woolley & Heggie, 2015). Note that accenting is an absolute, formal property of suffixes, which cannot be diluted or trumped by other properties or constraints, and is not subject to probabilistic influences from other lexical items or related processes.

These two hypotheses, that is, statistical associations and morphological accenting, both suggest that word‐final letter sequences can control, or at least contribute to, stress assignment in reading in Greek, as in other languages. However, they can also lead to competing predictions, taking into account their formal properties. Specifically, if critical letter sequences are recognised as accenting morphemes and if this morphological source of stress assignment information is typically exploited when reading, then we should observe full compliance with the suffix stress pattern, even in pseudowords, as long as the pseudowords are appropriately suffixed. This amounts to a morphological accenting hypothesis for stress assignment. A crucial property of this hypothesis is its absoluteness, as formal linguistic rules are posited to apply strictly (not probabilistically) whenever their relevant domain is matched. The suggestion that reading aloud words (and even pseudowords) is based on morphemic parsing is consistent with the findings of Wade‐Woolley and Heggie (2015), who found that English pseudowords suffixed with accenting morphemes are preferred when stressed in accordance with the suffix. It is also consistent with the model of Rastle and Coltheart (2000), which is based on the identification of specific prefixes and suffixes in English that are associated with stress patterns via hard‐coded rules. It is further supported by evidence in Italian (Burani, Arduino, & Marcolini, 2006; Burani, Marcolini, De Luca, & Zoccolotti, 2008), which is relatively similar to Greek in orthographic transparency and morphological complexity. Therefore, if words (and pseudowords) are routinely parsed into roots and suffixes in the course of processing for reading aloud (cf. morpho‐orthographic segmentation; see, e.g., Beyersmann et al., 2016, and references therein), it is reasonable to hypothesise that the stress properties of roots and suffixes are available to the (adult) skilled reader and may be operative in directing stress assignment.

In contrast, if word‐final letter sequences are only probabilistically associated with stress patterns, irrespective of their status as morphemes or other linguistic units, and if this statistical information is taken into account, then we should observe stress assignment consistent with the pattern typically associated with the letter sequence – but only to the extent that it dominates alternative patterns. That is, if words exist that end in the same letter sequences but have different stress patterns, then pseudowords ending in these letter sequences would be assigned stress less systematically than pseudowords ending in letter sequences that only appear in words with a single stress pattern. In other words, the presence of stress competitors (and their type and token frequency) would be expected to affect the position and reliability of stress assignment. This can be termed the probabilistic stress assignment hypothesis, consistent with related findings in English (Arciuli & Cupples, 2006; Arciuli et al., 2010). It is closely related to stress consistency effects observed in Italian stress neighbourhoods because those effects are also thought to arise from knowledge of the distribution of stress patterns over words with common word‐final segment sequences (e.g., Sulpizio, Arduino, Paizi, & Burani, 2013; Sulpizio & Colombo, 2013). However, our probabilistic stress assignment hypothesis differs from stress neighbourhood effects, as they are traditionally defined, in that it is not a priori restricted to a prespecified domain such as the last two syllables of the words. Rather, letter sequences of any size may be potential cues to the extent that they are systematically associated with a particular stress pattern. The cumulative effect of multiple, size‐varying, and potentially overlapping cues would then be observable in stress assignment behaviour on unfamiliar pseudowords, in which there is no fixed lexical knowledge to dominate the probabilistic cues and mask their influence.

Greek is a language with a relatively transparent orthography (estimated at 95% at the grapheme–phoneme level for the reading direction; Protopapas & Vlahou, 2009), in which stress is obligatorily marked with a diacritic. Previous studies in Greek have documented effects of three sources of stress assignment information: lexical (i.e., knowledge of the word), orthographic (i.e., diacritic) and default (i.e., the penultimate syllable; Protopapas, 2016). Of these, the lexical source dominates responses to pseudowords already from Grade 2, while the diacritic gains strength gradually as reading expertise develops. The default pattern results in a weak stress assignment tendency, evident only in the absence of other sources. To examine the relative strength of morphological or probabilistic stress assignment information in skilled Greek readers, in the present study we presented pseudowords with accenting suffixes without a diacritic or with a diacritic congruent or incongruent with the suffix. The hypothesised morphological and probabilistic source would be most clearly evident in the absence of a diacritic, whereas its relative strength, compared with the diacritic, can be determined by stress assignment preferences in the incongruent condition. Operation of the default source would be evident in a preference for penultimate‐syllable stress.

Method

Results

Stress assignment

Table 3 shows the proportion of stress assignment in the three diacritic conditions as a percentage of total responses, broken down by suffix stress position, including all responses (top) or (segmentally) correct responses only (bottom; percentages not adding up to 100% due to mispronunciations). Incorrect responses are only shown for completeness and were not analysed further. Across suffixes, taking only correct responses into account, unanimous‐stress items were stressed consistent with the suffix 75.7%, 94.0% and 18.3% in the absent, congruent and incongruent diacritic condition, respectively, whereas the corresponding percentages for majority‐stress items were 63.4%, 93.0% and 12.1%. The distribution (participant means) of stress assignment responses consistent with the suffix in each diacritic condition, averaged over suffix stress positions, is displayed graphically in Figure 1 (left).

Table 3. Proportion of stress assignment (%) in the three diacritic conditions as a function of suffix stress position.

Table 3a. All responses
	Absent				Congruent				Incongruent
Suffix stress	NA	1	2	3	NA	1	2	3	NA	1	2	3
	Unanimous stress items
1	3.0	71.7	7.8	17.5	6.4	91.7	0.8	1.1	5.0	17.5	54.6	23.0
2	2.9	7.1	77.6	12.4	1.3	0.8	95.3	2.6	4.5	46.1	19.5	30.0
3	5.7	0.8	16.6	77.0	3.7	0.0	1.6	94.7	6.6	40.5	34.7	18.1
	Majority stress items
1	1.0	71.3	2.4	25.4	2.9	94.3	0.5	2.4	1.9	6.2	0.0	91.9
2	2.6	1.1	57.9	38.4	3.2	0.0	91.1	5.8	2.6	10.5	18.9	67.9
3	1.8	0.0	45.6	52.6	1.8	0.0	3.5	94.7	7.0	0.0	82.5	10.5

Table 3b. Correct (i.e., segmentally accurate) responses only
	Absent				Congruent				Incongruent
Suffix stress		1	2	3		1	2	3		1	2	3
	Unanimous stress items
1		61.8	7.5	16.1		80.9	0.8	1.1		13.0	51.2	19.7
2		6.1	72.1	10.8		0.5	86.8	2.4		40.0	15.8	26.3
3		0.8	13.1	69.4		0.0	1.2	84.0		36.6	31.6	13.6
	Majority stress items
1		67.5	2.4	23.4		89.0	0.0	2.4		5.3	0.0	86.6
2		0.5	52.6	34.7		0.0	83.2	4.2		10.5	17.4	62.6
3		0.0	43.9	52.6		0.0	1.8	89.5		0.0	82.5	10.5

• Note: Stress positions in syllables, counted from the end: 1 = final, 2 = penultimate, 3 = antepenultimate. NA = stress not assigned.

To examine the effect of suffix preference and stress position on stress assignment behaviour, the proportion of stress assignment in responses consistent with the suffix (strsuf; a binomial variable) was analysed in models including suffix stress position (strpos, with three levels: final, penultimate and antepenultimate; simple‐coded) and stress preference (strpref, with two levels: unanimous and majority; dummy‐coded) as fixed factors, along with random intercepts for participants and items and random slopes of the two fixed factors and their interaction per participant. In R notation, this was coded as

strsuf ~ strpos * strpref + (strpos * strpref | subject) + (1 | item).

In the no diacritic condition, there was a significant intercept (β = 1.87, z = 10.36, p < .001), indicating stress assignment consistent with the suffix rather than equally distributed. For unanimous‐stress pairs, there was no effect of suffix stress position (final vs penultimate: β = −0.44, z = −0.93, p = .351; antepenultimate vs penultimate: β = 0.26, z = 0.60, p = .548). However, there was an effect of stress preference (β = −1.19, z = −3.42, p < .001), indicating that majority‐stress items produced fewer suffix‐consistent stress assignments than unanimous‐stress items. The interaction between stress position and stress preference was not significant (final vs penultimate: β = 1.20, z = 1.69, p = .092; antepenultimate vs penultimate: β = −0.51, z = −0.55, p = .583).

In the congruent diacritic condition, again there was a significant intercept (β = 19.44, z = 5.02, p < .001), indicating stress assignment consistent with the suffix and with the diacritic. There was no significant effect of stress position (final vs penultimate: β = −1.60, z = −0.17, p = .863; antepenultimate vs penultimate: β = 1.97, z = 0.22, p = .825) or stress preference (β = 12.59, z = 1.11, p = .267) and no interaction between the two (final vs penultimate: β = −3.59, z = −0.25, p = .803; antepenultimate vs penultimate: β = 4.93, z = 0.16, p = .872).

In the incongruent diacritic condition, there was a significant negative intercept (β = −2.86, z = −8.49, p < .001), indicating stress assignment predominantly inconsistent with the suffix. There was no significant effect of stress position (final vs penultimate: β = −0.06, z = −0.09, p = .925; antepenultimate vs penultimate: β = −0.69, z = −1.37, p = .172) for unanimous‐stress items. There was a significant effect of stress preference (β = −1.51, z = −3.07, p = .002) and a significant interaction of stress preference with stress position (final vs penultimate: β = −4.45, z = −4.01, p < .001; antepenultimate vs penultimate: β = 0.18, z = 0.15, p = .881).

To examine whether this pattern of results in the incongruent diacritic condition is better accounted for by the position of the diacritic than by the suffix, we fit the same model after replacing the dependent variable with a new binomial variable expressing stress assignments consistent with the diacritic. There was now a significant positive intercept (β = 2.43, z = 7.09, p < .001), indicating stress assignment predominantly consistent with the diacritic. Again, there was no significant effect of stress position (final vs penultimate: β = −0.43, z = −0.81, p = .416; antepenultimate vs penultimate: β = 0.005, z = 0.01, p = .992) for unanimous‐stress items. There was a significant effect of stress preference (β = 1.86, z = 3.87, p < .001) and a significant interaction of stress preference with stress position (final vs penultimate: β = 4.99, z = 4.65, p < .001; antepenultimate vs penultimate: β = 0.48, z = 0.43, p = .665), due to majority‐stress items with a penultimate‐stress suffix, which were assigned stress consistent with the diacritic somewhat less reliably. This effect is shown in Figure 1 (right).

Discussion

The results showed that word‐final letter sequences significantly affected stress assignment to pseudowords in the absence of a diacritic. In this condition, the mean proportion of responses consistent with the suffix was 76% for unanimous‐stress items and 64% for majority‐stress items. This difference was statistically significant. The substantial proportion of responses inconsistent with the suffix and the difference between unanimous and majority stress items suggest that stress assignment is probabilistically associated with item‐final letter sequences and not absolutely based on morphological accenting.

Our results for Greek are consistent with studies in English, in which probabilistic associations have been documented between word‐final letter sequences and stress patterns and, moreover, behavioural evidence indicates that these associations are operative in stress assignment by children and adult readers (Arciuli & Cupples, 2006; Arciuli et al., 2010; Monaghan et al., 2016; Wade‐Woolley & Heggie, 2015), subject to a developmental progression that can be accounted for by a statistical (connectionist) model learning to assign stress by exposure to a word corpus (Arciuli et al., 2010). Although the Greek stress system is very different from the English one in many ways, including phonologically, phonetically and orthographically (see, e.g., Protopapas, Panagaki, Andrikopoulou, Gutiérrez Palma, & Arvaniti, 2016), it seems that stress assignment behaviour in both languages is consistent with cumulative effects of probabilistic cues, based on lexical experience and knowledge, rather than absolute stress assignment rules associated with specific morphemes or other units. In particular, our findings do not support the notion that so‐called accenting morphemes are strongly associated with specific stress patterns in the sense of strictly determining stress assignment behaviour when reading aloud novel items (pseudowords). Therefore – and in agreement with Arciuli et al. (2010) – we submit that the available empirical evidence is not consistent with a view of stress assignment based on the operation of absolute, suffix‐specific rules, arising either from linguistic or from lexical analysis.

Sensitivity to multiple probabilistic cues is hardly a novel idea in language learning and psycholinguistics (Chater & Manning, 2006; Christiansen, 2013). The more general notion of statistical learning has been tied to individual differences in language skill (e.g., Arciuli & von Koss Torkildsen, 2012; Kidd, & Arciuli, 2016), language representations in the brain (e.g., Arciuli, McMahon, & de Zubicaray, 2012), and literacy (e.g., Arciuli & Simpson, 2012; Gabay, Thiessen, & Holt, 2015; Spencer, Kaschak, Jones, & Lonigan, 2015). Although many studies of statistical learning have mainly relied on sequential learning paradigms involving tracking of transitional probabilities, there are also other paradigms involving multiple probabilistic cues that have been linked to individual differences (and difficulties) in language and literacy (see discussion in Krishnan, Watkins, & Bishop, 2016). In particular, there is strong evidence for statistical learning of spelling patterns across ages and languages (e.g., Deacon, Conrad, & Pacton, 2008; Samara & Caravolas, 2014), and there is evidence that children are sensitive to spelling patterns even before they can spell phonologically (Pollo, Kessler, & Treiman, 2009). In this context, the suggestion that stress assignment in reading operates on the basis of multiple probabilistic constraints fits within a productive and well‐supported general theoretical framework for language processing.

In contrast, our findings do not seem to be amenable to modelling approaches akin to the dual route formulation of Rastle and Coltheart (2000), who hard‐coded absolute rules specific to particular prefixes and suffixes into their model for reading English, thereby defining “regularly” and “irregularly” stressed items by reference to those rules. Notably, the rules in the model of Rastle and Coltheart (2000) were based on an analysis of word types, that is, on observed probabilistic associations between stress patterns and word beginnings and endings. Moreover, the rules were applied to model behavioural data, again probabilistically associating stress patterns with particular items across participants. To posit a level of theoretically absolute rules, intervening between two levels of probabilistic observations, seems not only unparsimonious but also empirically unjustified. Specifically, the performance of the model incorporating stress assignment rules was 89.7% correct. This seems impressive, but it should be judged against the baseline of 83.2% correct that would be achieved simply by stressing every item on the first syllable, due to the great asymmetry in English stress patterns. Thus it seems that the case for rule‐based stress assignment appears weak in English as well as cross‐linguistically.

However, our findings cannot be interpreted as entirely precluding a role for morphological processing in Greek. In particular, these findings cannot be taken as evidence against the accenting properties of morphemes as proposed by linguistic analysis. Such properties may be used to account for the observed distribution of stress patterns over derived words in a language, evident in word production, and need not participate in processing mechanisms such as reading aloud. The processing demands of reading aloud (which involves visual word recognition) may be sufficiently distinct from those of spontaneous word production, or of productive derivation, so that each may depend on partially different representations, at different levels of abstraction and with different formal and statistical properties (cf. Protopapas et al., 2016).

The moderation of suffix‐consistent stress assignment by the existence of stress competitors, which was observed in our data, may conceivably reflect the operation of two conflicting processes: One process tends to assign stress based on the accenting properties of the suffix while another process may create opposition on the basis of the statistical association between orthographic and phonological (stress) patterns in the lexicon. At present, we do not have any specific evidence for the operation of the former (i.e., the morphological) process in Greek reading aloud, which remains to be investigated in future studies. However, such a hybrid approach is theoretically unparsimonious, because the main rationale behind positing a formal rule‐based operation is the productivity as observed in highly systematic associations in the data. If the actual empirical situation suggests that patterns are only probabilistically associated, then the rule‐based approach loses much of its driving force, as it does not seem to do any theoretically useful work. Therefore, even though our data cannot preclude the possibility of a morphological operation working against, or in combination with, a probabilistic process, it seems theoretically preferable to side with Arciuli et al. (2010) in rejecting this possibility in favour of statistical processing alone.

Our findings suggest a potential reinterpretation of the well‐established neighbourhood consistency effects typically observed in studies of reading Italian (Sulpizio et al., 2015). Although the notion of stress neighbourhoods seems well motivated and empirically supported, it has not been specifically investigated whether the operative domain must be precisely limited to the last two syllables of the words, excluding the onset of the penult, or whether it may be more flexibly extended or abridged. It is conceivable that smaller or larger word parts may be taken into account in the processing of written words by Italian readers, to reflect the grain size of associations between stress patterns and letter groups as they happen to have formed over the course of evolution of the language, which may not always align with the boundary between the onset and the nucleus of the penultimate syllable. Indeed, the analysis of Monaghan et al. (2016) suggests that letter groups of various sizes are associated with stress patterns in Italian, not necessarily limited to the stress neighbourhood domain. Note that freedom from the two‐syllable restriction does not imply another fixed domain of a different size. Rather, letter groups of arbitrary sizes may conceivably exert a simultaneous probabilistic influence to the extent they are associated with stress patterns. The cumulative influence of all reliably associated units will then be observable as systematic stress assignment behaviour. This proposal is akin to the proposal of learning orthographic patterns and graphophonemic mappings of multiple lengths simultaneously, as a result of systematicities occurring at different grain sizes, leading to sight word reading as well as to sublexical orthographic and decoding knowledge (Ehri, 2005, 2014; see also Apel, 2009, 2011).

Along these lines, a computational model of reading aloud Italian, including an explicit representation of stress patterns but no representation of stress neighbourhoods or morphemes, has successfully simulated both stress consistency and morphological constituency effects (Perry, Ziegler, & Zorzi, 2014). In the model, these effects presumably arose from interactions between a connectionist network performing graphophonemic mappings and a representation of arbitrary lexical mappings from orthography to phonology. This finding confirms that explicit representation of special processing units, although theoretically attractive, is not necessary when probabilistic information is effectively coded indirectly in large‐scale representations of lexical and sublexical mappings. Thus, the notion of stress neighbourhood, useful as it has been proven to be, may amount to a higher level statistical association insofar as the orthography‐phonology covariation involving stress is typically (but not necessarily always) captured at this grain size. Further empirical investigation, both corpus‐based and behavioural, will be required to resolve this issue.

Our data suggest that probabilistic letter cues to stress assignment in Greek are relatively weak, insofar as they were trumped by the diacritic: when unanimous stress items were presented with a suffix‐incongruent diacritic, the mean proportion of responses consistent with the suffix was only 18%; for majority stress items, the mean proportion was 12% (in the face of great individual variability, seen in Figure 1, left). However, probabilistic letter cues evidently override the influence from the default pattern, as there were no significant differences in stress assignment proportions consistent with the suffix between items with suffixes stressed on different syllables. If there were a substantial default effect, we should have observed more assignments consistent with the suffix when stressed on the penultimate than when stressed on other syllables, because the penultimate would be supported by both the suffix and the default. There was, however, an effect of suffix stress position for majority‐stress items presented with an incongruent diacritic. Specifically, an incongruent diacritic was slightly less effective when the suffix was stressed on the penultimate, consistent with a stronger “pull” towards the penultimate. Therefore, when the probabilistic letter cue is weak, due to the existence of stress competitors, the conflict from an incongruent diacritic apparently permits a weak influence of the default pattern to surface.

In a study of stress assignment in pseudoword reading by adult skilled readers, Protopapas et al. (2007, p. 708) reported the relative strength of stress information sources, when uncontested by other source, as follows: default 60%, lexical 80% and diacritic 90%. Note that this underestimates the typical force of the lexical source because the estimates were derived from pseudowords, which, even though resembling specific words well enough to elicit their activation, were clearly mismatching the words, evidenced by their segmentally correct pronunciation. Therefore, in typical word reading, the lexicon is probably the dominant source of stress information. Taking into account the findings of the present study, the order of importance of stress assignment information sources for Greek seems to be lexical > diacritic > final letters > default.

Our conclusions and suggestions may be tempered by the fact that we could only examine a limited set of suffixes, which was unbalanced with respect to the critical distinction between unanimous and majority items. This was an unavoidable consequence of applying criteria meant to exclude suffixes that would be unlikely to produce the sought‐after stress assignment effect. That is, suffixes appearing in only a few word types or tokens and suffixes with too many stress competitors might have unfairly disadvantaged the morphological accenting hypothesis. In this sense, we can be relatively confident that probabilistic association is more likely than absolute morphological accenting, because the deck was not stacked in the direction of this outcome. However, it remains possible that the small, and largely forced, set of suffixes examined, especially the majority items, may have resulted in a finding that somehow depends on (some of the) particular suffixes rather than on a general psycholinguistic processing principle. We believe that this limitation is not very severe, as our conclusion is consistent with the interpretation of empirical psycholinguistic observations in other studies across languages, even if it is at somewhat odds with certain linguistic analyses.

In conclusion, in the present study, we have documented an additional source of stress assignment information in reading aloud Greek. Specifically, probabilistic word‐final information associated with accenting derivational suffixes was found to affect reading behaviour, especially in the absence of lexical and diacritic information. Although statistical associations between word endings and stress patterns have been previously documented in other languages, it is not known in most cases whether this kind of information is operative in actual reading processes. Our finding enriches our understanding of stress assignment and suggests that the reading system can be sensitive to any relevant information, weighing each potential source to achieve optimal performance.