Scholarly article on topic 'The processing of English regular inflections: Phonological cues to morphological structure'

The processing of English regular inflections: Phonological cues to morphological structure Academic research paper on "Psychology"

CC BY
0
0
Share paper
Academic journal
Cognition
OECD Field of science
Keywords
{"Speech processing" / Morphology / Phonology / "English inflection"}

Abstract of research paper on Psychology, author of scientific article — Brechtje Post, William D. Marslen-Wilson, Billi Randall, Lorraine K. Tyler

Abstract Previous studies suggest that different neural and functional mechanisms are involved in the analysis of irregular (caught) and regular (filled) past tense forms in English. In particular, the comprehension and production of regular forms is argued to require processes of morpho-phonological assembly and disassembly, analysing these forms into a stem plus an inflectional affix (e.g., {fill}+{-ed}), as opposed to irregular forms, which do not have an overt stem+affix structure and must be analysed as full forms [Marslen-Wilson, W. D., & Tyler, L. K. (1997). Dissociating types of mental computation. Nature, 387, 592–594; Marslen-Wilson, W. D., & Tyler, L. K. (1998). Rules, representations, and the English past tense. Trends in Cognitive Science, 2, 428–435]. On this account, any incoming string that shows the critical diagnostic properties of an inflected form – a final coronal consonant (/t/, /d/, /s/, /z/) that agrees in voicing with the preceding segment as in filled, mild, or nilled – will automatically trigger an attempt at segmentation. We report an auditory speeded judgment experiment which explored the contribution of these critical morpho-phonological properties (labelled as the English inflectional rhyme pattern) to the processing of English regular inflections. The results show that any stimulus that can be interpreted as ending in a regular inflection, whether it is a real inflection (filled–fill), a pseudo-inflection (mild–mile) or a phonologically matched nonword (nilled–nill), is responded to more slowly than an unambiguously monomorphemic stimulus pair (e.g., belt–bell). This morpho-phonological effect was independent of phonological effects of voicing and syllabicity. The findings are interpreted as evidence for a basic morpho-phonological parsing process that applies to all items with the criterial phonological properties.

Academic research paper on topic "The processing of English regular inflections: Phonological cues to morphological structure"

Contents lists available at ScienceDirect

Cognition

journal homepage: www.elsevier.com/locate/COGNIT

The processing of English regular inflections: Phonological cues to morphological structure

Brechtje Posta'*, William D. Marslen-Wilsonb, Billi Randallc, Lorraine K. Tylerc

aResearch Centre for English and Applied Linguistics, University of Cambridge, English Faculty Building, 9 West Road, Cambridge CB3 9DP, United Kingdom b MRC Cognition and Brain Sciences Unit, 15 Chaucer Road, Cambridge, United Kingdom

c Centre for Speech, Language and the Brain, Department of Experimental Psychology, University of Cambridge, Downing Street, Cambridge, United Kingdom

ARTICLE INFO

ABSTRACT

Article history: Received 18 July 2006 Revised 12 June 2008 Accepted 22 June 2008

Keywords: Speech processing Morphology Phonology English inflection

Previous studies suggest that different neural and functional mechanisms are involved in the analysis of irregular (caught) and regular (filled) past tense forms in English. In particular, the comprehension and production of regular forms is argued to require processes of morpho-phonological assembly and disassembly, analysing these forms into a stem plus an inflectional affix (e.g., {fill} + {-ed}), as opposed to irregular forms, which do not have an overt stem + affix structure and must be analysed as full forms [Marslen-Wilson, W. D., & Tyler, L. K. (1997). Dissociating types of mental computation. Nature, 387, 592-594; Mar-slen-Wilson, W. D., & Tyler, L. K. (1998). Rules, representations, and the English past tense. Trends in Cognitive Science, 2, 428-435]. On this account, any incoming string that shows the critical diagnostic properties of an inflected form - a final coronal consonant (/t/, /d/, /s/, /z/) that agrees in voicing with the preceding segment as in filled, mild, or nilled - will automatically trigger an attempt at segmentation. We report an auditory speeded judgment experiment which explored the contribution of these critical morpho-phonological properties (labelled as the English inflectional rhyme pattern) to the processing of English regular inflections. The results show that any stimulus that can be interpreted as ending in a regular inflection, whether it is a real inflection (filled—fill), a pseudo-inflection (mild-mile) or a phonologically matched nonword (nilled-nill), is responded to more slowly than an unambiguously monomorphemic stimulus pair (e.g., belt-bell). This morpho-phonological effect was independent of phonological effects of voicing and syllabicity. The findings are interpreted as evidence for a basic morpho-phonological parsing process that applies to all items with the criterial phonological properties.

© 2008 Elsevier B.V. All rights reserved.

1. Introduction

Recent research into the neural and functional architecture of the human language system has been strongly influenced by the contrast between the regular and irregular past tense in English, which is regarded as a critical test case for discriminating competing claims about the organisation of the language system (Joanisse & Seidenberg, 1999; Marslen-Wilson & Tyler, 1997; Marslen-Wilson & Tyler, 1998; Marslen-Wilson & Tyler, 2003; Pinker, 1991;

* Corresponding author. E-mail address: bmbp2@cam.ac.uk (B. Post).

0010-0277/$ - see front matter © 2008 Elsevier B.V. All rights reserved. doi:10.1016/j.cognition.2008.06.011

Pinker & Ullman, 2002; Plunkett & Marchman, 1993; Rumelhart & McClelland, 1986; Ullman 2004; Ullman et al., 1997). The central issue is whether the representation and processing of regular past tense forms, involving the combination of stems and affixes (e.g., play + /d/ = played), requires specialised cognitive and neural procedures which are not invoked by the unpredictable and idiosyncratic irregular forms (e.g., buy-bought, hit-hit, or creep-crept), where there is typically no overt combination of stem and affix.

The investigation of this issue has raised a set of more specific questions about the processing mechanisms involved in the cognitive analysis of regular and irregular

forms in language comprehension (and production). Based on a number of neuropsychological and neuro-imaging studies, Marslen-Wilson, Tyler and colleagues argue for a distinction between lexical access processes that involve morpho-phonological decomposition, and those based on more direct access to stored forms (Marslen-Wilson & Tyler, 1997; Marslen-Wilson & Tyler, 1998; Marslen-Wilson & Tyler, 2003; Marslen-Wilson & Tyler, 2007; Tyler, Mar-slen-Wilson, & Stamatakis, 2005b; Tyler, Randall, & Mar-slen-Wilson, 2002b; Tyler, Stamatakis, Post, Randall, & Marslen-Wilson, 2005a; Tyler et al., 2002a). This emphasis on the morphological decomposition of regular inflected forms clearly allies this account with the ''Words and Rules" (e.g., Clahsen, 1999; Pinker, 1999) and the procedural/declarative (e.g., Ullman, 2004) approaches, in distinction to non-decompositional, usually connectionist approaches, which deny the existence of separable stem and inflectional morphemes, and which argue instead that inflected forms are processed and represented as patterns of activation across pools of simple neuron-like processing units which share certain semantic, phonological and orthographic information (e.g., McClelland & Patterson, 2003; Rumelhart & McClelland, 1986). Unlike Pinker and colleagues, however, we do not assume that the presence or absence of grammatical morphemes implicates differences in the nature of mental computation. Our concern is with the functional architecture of the language processing system, not its underlying computational properties.

The Marslen-Wilson and Tyler morpho-phonological account was directly tested in a comprehension experiment with four nonfluent aphasics with a documented regular past tense deficit, using a speeded same-different judgment task (Tyler et al., 2002a). In this task, participants hear trials of matched pairs of spoken stimuli. Half of the trials are same pairs in which the first word is repeated (e.g., filled-filled), and half are different pairs with a minimally different first and second word (e.g., filled—fill). Participants decide as quickly and as accurately as possible whether the two members of a pair sound the same or different. The effects of phonological complexity were controlled by including monomorphemic words and nonwords that were matched for form on a one-to-one basis to the regular and irregular pairs (e.g., pseudo-regular jade—jay and nonword kade—kay for real regular played— play). In a further condition, controlling for possible segmentation effects, the first and the second word again differed in the removal of the final phoneme of the first member of the pair (e.g., claim—clay), but here the final phoneme was not a possible inflectional affix.

There were two important aspects to the results. The first was that the performance of patients with past tense deficits was most impaired for the real regular pairs, with significantly slower responses than to matched conditions (such as pseudo-regulars and nonword regulars). The second main result - and the stimulus for the research reported here - was that overall, the patients were substantially impaired for all conditions that contained a potential regular inflectional affix. Thus, although performance was poorest for the real regulars (mean reaction time of 1420 ms and error rate of 31%), it was also poor both for the pseudo-regulars (RT of 1252 ms and error rate

of 25%) and for the nonword regulars (RT of 1244 ms and error rate of 22%). Performance was much less impaired, and closer to normal levels of accuracy, for the control pairs such as claim-clay (RT of 1044 ms and error rate of 5%), where the first member of the pair did not end in a potential affix. A similar grouping of responses to real regulars, pseudo-regulars and nonword regulars, in distinction to morphologically simple pairs of the claim-clay type, was also observed in a subsequent neuro-imaging study, where unimpaired young adults performed the same-different task on the same types of material in an event-related fMRl study (Tyler et al., 2005a).

The hypothesis we explore here is that this commonality between the three critical regular past tense conditions (real, pseudo, nonword) reflects their common morpho-phonological properties - i.e., that they all share specific phonological features that are diagnostic of the presence of a potential inflectional suffix, and will therefore place specific demands on the neural and functional machinery underlying the perceptual processing of spoken words in English. These diagnostic phonological features, which we label the English inflectional rhyme pattern (IRP), have two components: the presence of a word-final coronal consonant (i.e., any sound articulated with the tip or blade of the tongue raised towards the teeth or the alveolar ridge, such as /d, t, s, z/), and the agreement in voice between the final coronal consonant and the segment that precedes it. Thus, the sequence passed /pa:st/ is a potential combination of a verbal stem pass /pa:s/ with an inflectional suffix, because it ends with the unvoiced coronal consonant /t/, and this agrees in voice with the preceding unvoiced segment /s/. The same applies to the pseudo-regular word fast, which is potentially analyzable as the (nonexistent) verb stem /fa:s/, plus the inflectional morpheme /t/. The same, furthermore, would also hold for the nonword sequence nast, also potentially analyzable as /na:s/ plus /t/.

The presence or absence of these diagnostic features (the IRP) will lead a spoken lexical input to interact differentially with the machinery of lexical access and linguistic interpretation (cf. Marslen-Wilson & Tyler, 2007). This machinery, broadly speaking, is concerned on the one hand with the mapping of phonological inputs onto stored lexical representations, and on the other with the extraction and interpretation of grammatical morphemes (and other cues to structure). The successful functioning of this system requires the appropriate segmentation of speech inputs into stems and different types of affix. When an input such as /pa:st/ is encountered, corresponding to the past tense form passed, the system needs both to access the semantic and syntactic properties associated with the stem {pass} and to extract the processing implications of the presence of the grammatical morpheme {-t}.

A critical claim about the functioning of this system -motivated in particular by the effects for nonword regulars in the earlier neuropsychological and neuro-imaging studies (Tyler et al., 2002a; Tyler et al., 2005a) - is that the attempt at segmentation into stem and affix is automatically triggered by any input that has the critical diagnostic properties. Whenever the system encounters a candidate string that ends with a coronal consonant that agrees in voice with its preceding segment, this will always have to be

evaluated both as a monomorphemic input and as a verbal or nominal stem with an accompanying inflectional affix. This seems to be forced by the pervasive ambiguity of possible inflected inputs. An input like /peist/ may be either the monomorphemic form paste or the past tense of pace; /treid/ could be the past tense of the stem tray, and so forth. The system cannot decide in advance which strings with the appropriate properties are inflected forms or not - even if, in a case like trade, the stem in question may never be used as a verb.

This predicts that listeners should make slower responses whenever they are asked to make same-different judgments where at least one member of the pair ends with the characteristic inflectional rhyme pattern, relative to control pairs which deviate from this pattern - for example, pairs like tent-ten (where the first member of the pair ends in a coronal which does not agree in voicing with the preceding segment) or clamp-clam and bark-bar (where the first word ends in labial or velar consonants). These are less complex to process because they can only be interpreted as full forms, and will not engage morpho-phonological segmentation mechanisms that generate additional lexical complexity and competition. The purpose of the research reported here, therefore, is to explore this specifically morpho-phonological hypothesis about the nature of the processing operations associated with the presence of regular inflectional morphology in English.

We need to consider this hypothesis, however, in the context of a widely held competing view, which argues that patients' difficulties with regular past tense inflections do not reflect the morphological or morpho-phonological properties of inflected words, but rather purely phonetic and phonological factors relating to the perceptual complexity of the forms in question. These are the single mechanism connectionist models, where there is no independent representation of morphology, either regular or irregular. The currently most prominent of these models has been developed by Joanisse and Seidenberg (1999), claiming that differential effects for regulars and irregulars can be modelled in a single undifferentiated network without reference to morphological features, by exploiting the statistical regularities in the variation of semantic and phonological overlap between stem and past tense form that distinguish regulars and irregulars. This approach makes the strong prediction that any apparently morpho-phono-logical effects for regular inflected forms are in fact primarily phonological in nature (cf. Bird, Lambon Ralph, Seidenberg, McClelland, & Patterson, 2003; McClelland & Patterson, 2003). Regulars are argued to be more difficult to process than irregulars because they have ''greater artic-ulatory complexity and perceptual subtlety" (McClelland & Patterson, 2003, p63). The poorer performance of nonflu-ent patients on regular forms is therefore primarily attributed to general problems in phonological processing, and not to any specifically morphological deficit (e.g., Bird et al., 2003; cf. Tyler et al., 2002b, and Marshall & Van der Lely, 2006, for a recent study with children).

Although notions of phonological complexity and perceptual salience are not fully defined in the relevant single mechanism publications, they can be taken to refer to characteristic phonological properties of regular inflected forms

- in particular the structure of the final consonant cluster, and the types of phoneme involved. Since the inflection is realised as /t/ or /d/ except when the final consonant is already a /t/ or /d/ (as in forms like greeted or sighted), this often leads to unusual final consonant clusters like /spt/ in clasped (cf. Marshall & van der Lely, 2006). At the same time, as noted previously, the inflectional morpheme always ends in a coronal, which may have a special status linguistically (Paradis & Prunet, 1991). Coronals are the most common phonemes cross-linguistically (Maddieson, 1985). Moreover, they may be perceptually less salient than labials (like /p/) and dorsals (like /k/; Hume, Johnson, Seo, &Tserdanelis, 1999, for stops in Korean and English), they tend to be replaced by velars or alveolars in speech errors (Stemberger, 1991), and they have been found to be more susceptible to regressive place assimilation (Jun, 1995, reported in Kochetov, 2004). There is no clear evidence, however, that coronality itself selectively affects the phonological processing of regular past tense forms.

In the current experiment, we will evaluate the processing implications of a number of phonological and morpho-phonological properties of regular inflected forms in English, taking care to control closely for phonological complexity, defined as the type of CV sequence that makes up a particular word form. The further organisation of the experiment is laid out in the following section.

1.1. The inflectional rhyme pattern: Coronality and voicing

The primary set of contrasts explores the significance of the different components of the diagnostic rhyme pattern that indicate the presence of a potential word-final inflectional morpheme. Taking a set of real regular past tense items (filled-fill, blessed-bless) as a form of baseline, these are compared with three sets of morphologically simple real words, as well as with matched sets of nonwords, as shown in Table 1.

The first comparison set of real words, as in previous experiments, is a set of pseudo-regular forms like mild-mile or crest-cress. These [+Coronal, +VoiceAgreement] materials share the inflectional rhyme pattern, and have an embedded real stem, but the full form is not itself an attested past tense form. We expect slower processing of these forms relative to uninflected forms that do not have the rhyme pattern, just as for the real regulars, because the presence of the rhyme pattern should trigger the same morpho-phonological segmentation and evaluation processes. These processes will activate, for example, the spurious lexical candidate cress, when the first word crest is heard, which should both increase processing load for crest, and potentially interfere with the same-different decision to the second word, cress.

The second comparison set, containing pairs like startstar or tent-ten, retains the coronal ending, but violates the second component of the inflectional rhyme pattern, since this final consonant does not agree in voice with the preceding consonant - if ten had a past tense, it would have to be tenned (/tend/, homophonous with tend). If the presence of a word-final coronal intrinsically causes perceptual difficulties, then responses should also be slower for these [+Coronal, -VoiceAgreement] materials. If it is

Table 1

Overview of experimental conditions

Real word conditions Nonword conditions

Regular past tense baseline 1 Regular past in /t/& /d/ [+Coronal, +Voice Agreement] filled - fill in /t/ and /d/ gubbed - gub

Coronality and voicing 2 3 4 Pseudo past in /t/ & /d/ [+Coronal, +Voice Agreement] mild -[+Coronal, -VoiceAgreement] in /t/ belt - bell [-Coronal, -VoiceAgreement] in /p/ & /k/ lamp - lamb mile also gubbed - gub in /t/ steet - stee in /p/ and /k/ wump - wum

Inflectional paradigm (Manner of articulation) 5 6 Present tense in /s/ & /z/ [+Coronal, +Voice Agreement] fails -Plural in /s/ & /z/ [+Coronal, +Voice Agreement] meals - meal fail in /s/ and /z/ pakes - pake also pakes - pake

Syllabicity 7 8 Regular past in /id/ folded - fold Progressive aspect /iQ/ fending - fend in /id/ milted - milt in/l|]/ sunching - sunch

the presence of the inflectional rhyme pattern that is critical, then these materials should be treated as monomor-phemic, and therefore less complex than either the regulars or the pseudo-regulars.

The third set of words moves further away from the inflectional rhyme pattern by replacing the final consonant with a noncoronal segment (either a labial /p/, as in clamp-clam, or a dorsal /k/, as in milk-mill). These [-Coronal, -VoiceAgreement] materials violate both components of the inflectional rhyme pattern, and should be treated as monomorphemic full forms in the same way as the [+Coro-nal, -VoiceAgreement] stimuli in the previous condition. However, if it is the presence of a coronal consonant that is critical, on a perceptual difficulty account, then performance should be improved here, relative to the [+Coronal] conditions.

ln order to neutralise the phonological complexity issue, all three of these sets of words will be closely matched to each other, and to the real regulars, in their CV structures, covering both word onsets and word offsets. If there are differences between conditions, these should not reflect differences in complexity. Note that this means that highly complex regular past tense forms, ending in a CCC sequence (as in clasped or asked), are not included here as part of the main experiment, since they cannot be matched across the other real word conditions.

The three sets of monomorphemic comparison conditions were accompanied by three matched sets of non-words - a pseudo-regular [+Coronal, +VoiceAgreement] set, as in minned-min /mlnd/ /mln/ or stessed-stess /stest/ /stes/, a coronal noninflectional [+Coronal, -VoiceAgreement] set, as in rint-rin /rint/ /rin/ or lart-lar /la:t/ /la:/ and a noncoronal [-Coronal, -VoiceAgreement] set, as in plamp-plam /ptemp/ /ptem/ or tulk-tul /tAlk/ /tAl/. Again, the morpho-phonological account predicts more complex processing for the pseudo-regular set than for the other two, on the assumption that the presence of the inflectional rhyme pattern will trigger segmentation attempts that generate a pseudo-stem, such as min or stess, potentially disrupting same-different judgements to the actual second member of the pair. On the perceptual difficulty account, the [-Coronal, -VoiceAgreement] stimuli should contrast with the first two sets which both contain wordfinal coronal consonants. All three sets are matched to each other, and to the real regulars, in their CV structure.

Two further contrasts complement this main set of materials, while also expanding the variety of materials

the participants are exposed to. The first of these, using the s inflection, expands the coverage of this research beyond the set of past tense inflections, allowing us to examine the generality of the claims being made here for the influence of the inflectional rhyme pattern.

The regular s inflection in English, used to mark noun plurals (as in cats-cat or yards-yard) and the third person present tense (as in lick-licks or begs-beg), observes the same diagnostic constraints as the past tense inflection. The inflection itself, whether realised as an /s/ or a /z/, is a coronal consonant, and it must agree in voice with the preceding segment - a form like sparse cannot be interpreted as the plural of spar. The difference with the regular past tense forms is the manner of articulation of the final consonant; /s/ and /z/ are fricatives, while /t/ and /d/ are stops. If the inflectional rhyme pattern has the same processing consequences here as for the past tense inflection, then we expect present tense and plural forms in /s/ and /z/ to pattern with the real and pseudo past tense forms in /t/ and /d/, rather than with the [+Coronal, -VoiceAgreement] and [-Coronal, -VoiceAgreement] sets, such as tent-ten and milk-mill, which do not conform to the inflectional rhyme pattern. The same should hold, for the same reasons, for nonword pairs like pakes-pake /peIks/ /peIk/ or dags-dag /dœgz/ /dœg/, relative to the parallel past tense nonword sets.

English regular inflection also includes two cases which do not obey the inflectional rhyme constraint, and which are both syllabic, as opposed to the cases we have considered above. These are the syllabic past tense allomorph /id/ (e.g., folded-fold), and the progressive aspect morpheme /il]/, which we include here primarily to give full coverage of English inflection, as well as increasing the variety of materials the participants are exposed to. Although the syllabic past tense allomorph /id/ ends in a coronal just like the nonsyllabic past tense allo-morphs /d/ and /t/, it does not show voicing agreement with the rhyme. Progressive /if]/ does not share any phonological features with the nonsyllabic allomorphs. There is not a clear prediction here for the same-different task. in the absence of the rhyme diagnostics, we would not expect to see the slow down in response times that we anticipate for the monosyllabic inflections. At the same time, it is possible that the syllabic nature of the inflection makes the same-different judgments highly perceptually salient, which may be reflected in faster response times.

Table 2

Mean duration of the second word for each condition

Condition Duration word 2 in ms

Real words Nonwords

Overall Same Diff. Overall Same Diff.

Real regular past filled-fill 741 765 717 733 752 713

Pseudo past in /t/ /d/ mild-mile 721 753 689

[-Cor, -Voice] in /p/ /k/ lamp-lamb 683 707 658 661 680 641

[+Cor, -Voice] in /t/ belt-bell 665 697 632 693 713 672

Present tense in /s/ /z/ fails-fail 725 796 654 703 761 645

Plural in /s/ and /z/ meals-meal 761 813 709

Regular past in /Id/ folded-fold 754 819 689 784 822 745

Progressive aspect fending-fend 810 851 768 846 879 812

2. Methods

2.1. Design

We examined the contribution of morphological and phonological factors to regular past tense processing by systematically manipulating, relative to a baseline set of past tense forms (condition 1 in Table 1 above): (i) coronal-ity and voicing (inflectional rhyme pattern present in condition 2; absent in conditions 3 and 4), (ii) inflectional paradigm (present tense and plural in conditions 5 and 6), and (iii) syllabicity (syllabic allomorphs in conditions 7 and 8). The items are listed in Appendix A.

2.2. Materials

There were 48 trials per condition, consisting of 24 different pairs with a minimally different first and second word (or nonword) which were also used as same pairs in which the first word was repeated as the second word. Eight of the 14 conditions were real word conditions and six were nonword conditions. The nonword items were mostly derived from the real words in the corresponding conditions so as to maximise their phonological similarity (sharing as many phonemes in onsets, nuclei and codas as possible). Except for items in real inflected conditions, the word pairs were semantically unrelated. All items were based on matched monosyllabic and monomorphemic stems, to avoid any confound between morpho-phonolog-ical decomposition of inflections and other morphological or phonological processes. All monosyllabic conditions were matched for the phonological complexity of the first

member of each test pair, with complexity defined here as CV structure. This covered both onsets and codas across all the relevant conditions. If, for example, the regular past tense set contained the form prayed, with the CV structure CCVC, then a form with the same CV structure would occur across the other conditions. For all same-different pairs, the second word of the pair would have the same CV structure but without the final segment. For the syllabic conditions, the CV structure of the stem of the first word in the pair was used for matching instead of the whole word form (e.g., greeted-greet was matched with prayed-pray).

Since the experiment contained monosyllabic and bisyl-labic items in different conditions, the duration of the second word varied significantly between conditions (see Table 2; [F(13,648) = 11.20, MSE = 137,899, p <.001]). A second factor that played a role in the duration of the second word was the voicing of the final consonant (Gimson, 1961). In English, vowels that precede a voiceless consonant are shorter than vowels which precede voiced material (e.g., plate is much shorter than played; e.g., Wiik (1965)). In our experiment, since the monomorphemic conditions in /t/, /k/ and /p/ all ended in voiceless consonants, while the morphologically complex forms and the pseudo past contained both voiced and voiceless codas (12/24 voiced codas for real past, present and plural, and 16/24 voiced codas for the pseudo past), the second word in the word pairs in the latter conditions are on average longer than in the former (see Table 2). The conditions could not be matched for voicing, because the set of pseudo past forms is too small to be limited to the subset with voiceless codas. The alternative of voicing half of the items

Table 3

Descriptive lexical statistics by condition: median lemma frequencies per million, summed word form frequencies per million, and familiarity and imageability ratings (word 2 medians reported for the different pairs)

Condition Word 1 Word 2

Lemma noun verb Wf sum Fam Imag Lemma noun verb Wf sum Fam Imag

Real regular past filled-fill 0 37 14 500 415 0 37 13 504 441

Pseudo past in /t/ /d/ mild-mile 5 0 14 539 504 5 0 14 506 505

[-Cor, -Voice] in /p/ /k/ lamp-lamb 3 3 3 435 463 5 3 10 469 467

[+Cor, -Voice] in /t/ belt-bell 10 1 9 481 449 9 1 10 517 447

Present tense in /s/ /z/ fails-fail 0 14 1 469 400 0 14 1 484 400

Plural in /s/ and /z/ meals-meal 37 0 13 510 549 37 0 10 510 569

Regular past in /Id/ folded-fold 0 7 2 491 398 0 7 5 491 408

Progressive aspect fending-fend 0 22 6 429 384 0 22 7 468 398

in the nonpast condition in /t/ was not a possibility, since this would have led to voicing agreement in the rhyme, turning them into pseudo past or real past items (e.g., voicing final /t/ in trait gives trade). This means that both voicing and duration of the second word have to be taken into account in the reaction time analyses.

We controlled as far as possible for form class ambiguity. Items used in verbal contexts were generally more frequent as verbs, even if a noun form existed and items used in nominal contexts were more frequent as nouns, even when a homonymic verb existed. In addition, subject to the above constraints, we matched the conditions as closely as possible for lemma frequency, word form frequency, familiarity and imageability (see Table 3; based on the CELEX lexical database (Baayen, Piepenbrock, & Gul-ikers, 1995), the MRC Psycholinguistic data base (Wilson, 1988), and our own data base of locally collected information from paper-and-pencil rating tasks with minimally 15 participants). One-way analyses of variance, carried out for each "nuisance" variable with the factor condition (8 levels), nevertheless showed a significant main effect for each variable, with the exception of familiarity of the second word in the word pair.1

The stimuli were recorded on DAT tape in a sound-attenuated booth, and digitised at 22.05 kHz for further processing. The first item in each pair was spoken by a male speaker, and the second by a female speaker to ensure that the judgments were not made on the basis of the low-level acoustic or phonetic properties of the test pairs. Both speakers were native speakers of British English.

2.3. Procedure

The pairs were presented in a single-version experiment in pseudo-random order in four sessions of 184 stimuli, each introduced by 4 dummy stimuli, and with equal numbers of pairs from each condition in each session (108 real words and 72 nonwords). Numbers of same and different stimuli were also balanced (90 of each per session), and at least one session intervened between repetitions of an item as a same or a different stimulus. The experiment began with a practise session (24 items), and each session was followed by a short break. The experimental software package DMDX (Forster & Forster, 2003) was used for the presentation of the stimuli. Participants were tested in quiet conditions, using a two-button response box, and wearing headphones. They would first hear the inflected form or its equivalent, spoken by the male voice, and after a 100 ms delay, they would hear the second item, spoken by the female voice. The intertrial interval was 850 ms, and the time-out was set at 3 s. The participants were asked to press the but-

1 Word 1: lemma frequency noun F(7,371) = 6.83, p <.001, lemma frequency verb F(7,371) = 4.01, p <.001, word form frequency F(7,371) = 2.74, p <.01, marginal for familiarity F(7,371) = 2.00, p = .054, imageability F(7,371) = 6.70, p <.001; word 2: lemma frequency noun F(7,371) = 6.84, p <.001], lemma frequency verb F(7,371) = 3.20, p <.01, word form frequency F(7,371) = .92, p = .49, imageability F(7,371) = 6.54, p < .001, familiarity F(7,371) = .59, p = .77.

ton labelled same when the two items sounded the same, and different when they sounded different. The experiment took about 45 min.

2.4. Participants

We tested 20 participants, 5 men and 15 women, from the subject pool of the Centre for Speech, Language and the Brain. They were all native speakers of English, aged between 18 and 25, and had no known hearing deficits. The participants were paid a small fee.

3. Results

3.1. Reaction times

One item was removed from the analysis because of a very high error rate (85%), and nine items were lost because of programming error (1.25% of items).2 The RT data, which were only collected for correct responses, were then inverse transformed to reduce the effects of outliers (Ratcliff, 1993). in the subject analysis (F1), the data were averaged over items, and in the item analysis (F2), over subjects. The mean reaction times and error proportions are shown for each condition in Table 4.

As a first examination of the data, we conducted an overall analysis of variance to see if there were differences between conditions and judgment types. There were two repeated measures in the subject analysis (condition with 14 levels and judgment type with two levels: same and different). in the item analysis, there was one repeated measure (judgment type) and one independent measure (condition). The results show a significant effect of condition [F1(13,247) = 53.42, p <.001; F2(13,312) = 8.41, p <.001], and a marginal effect of judgment type in the items analysis only [F1(1,19) = 1.09, p = .310; F2(1,312) = 3.41, p = .07]. in addition to the main effects there was a significant interaction between condition and judgment type [F1(13,247) = 6.51, p <.001; F2(13,312) = 2.06, p <.05].3 This is because same items are not always responded to faster than different items, even though overall, same items are responded to 20 ms more slowly than different items (see Table 4 and Appendix B; this interaction is explored further in the regression analyses reported below).

The overall effect of condition justifies further analyses to explore the principal factors of interest - coronality and voicing, inflectional paradigm (reflected by manner of articulation) and syllabicity. At the same time, we needed to evaluate whether any of the "nuisance" variables af-

2 The item with the 85% error score was an outlier: only 5 other items had error rates above 25% (1 at 35% and 4 at 25%). The missing items were two regular past different items, one regular past same item, one present tense different item, one noncoronal /k/ /p/ different item, two different items from the nonwords in /t/ and /d/, one different item from the nonwords in /k/ and /p/, and one different item from the nonwords in /Id/.

3 The same pattern of results obtained when we included word 2 duration as a covariate in an analysis of variance with between-subject factors condition (14 levels) and judgment type (2 levels) [condition: F(13,633) = 15.81, p <.001, judgment type: F(1,633) = 60.05, p <.001, condition x judgment type: F(13,633) = 2.51, p <.01, word 2 duration: F(1,633) = 324.07, p <.001].

Table 4

Harmonic mean reaction times and error rates for same and different judgments: (1) grand mean, (2) same items, and (3) different items, by condition

Condition Harmonic mean RT (ms) Error proportion (%)

Overall Same Diff. Overall Same Diff.

Regular past /t/ /d/ filled-fill 949 922 979 4.0 4.1 3.9

Pseudo past /t/ /d/ mild-mile 932 895 973 5.5 4.2 6.9

[-Cor, -VoiAgree] /k/ /p/ lamp-lamb 821 831 812 1.6 2.3 0.9

[+Cor, -VoiAgree] /t/ belt-bell 806 813 799 2.5 1.9 3.1

Present /s/ /z/ fails-fail 909 905 914 2.9 4.0 1.7

Plural /s/ /z/ meals-meal 927 912 942 3.8 3.5 4.0

Regular past /Id/ folded-fold 819 832 806 2.0 2.5 1.5

Progressive fending-fend 876 887 866 3.1 2.5 3.8

Nonword /d/ /t/ gubbed-gub 908 886 932 3.1 1.5 4.8

Nonword /t/ (not past) steet-stee 853 858 848 3.2 4.6 1.9

Nonword /p/ /k/ wump-wum 816 817 815 3.2 4.2 2.2

Nonword /s/ /z/ pakes-pake 900 872 929 3.6 3.1 4.2

Nonword /Id/ milted-milt 860 866 853 3.8 5.0 2.6

Nonword sunching-sunch 879 879 879 2.2 3.3 1.0

fected reaction time, especially since not all of these could be fully matched across conditions (see Tables 2 and 3 above). Correlation analyses showed that none of these variables, except for the duration of the second word [r(662) = -0.49, p < .001], did in fact correlate with RT.4

We also tested for a possible confound, raised by a reviewer, with the diphone frequency of the final phonemes in the word pairs. This might have affected reaction times if participants simply relied on how dissimilar the final phonemes were to make their judgements. Responses would be quickest for same items, where the words of the pair are identical. For different items, however, less frequent combinations of consonants might be harder to process than more frequent ones, resulting in relatively slower reaction times. Calculations of the logged frequencies of the final diphones of word 1 and word 2 in the stimulus pairs as well as their ratio (using CELEX wordform frequencies, Baayen et al., 1995), showed that these did differ significantly between conditions. Correlational analyses, however, showed that diphone frequency did not correlate with RT in the experiment: word 1 r(379) = -.05, p = .31, word 2 r(379) = -.08, p = .13; ratio diphone frequency word 1/word 2 r(379) = -.08, p = .14.

In the main analysis reported here, using regression techniques, we included the duration of the second word as a continuous regressor, in addition to the binary regres-sors which represented the factors of interest. These were (1) morphological status: items that ended in a real inflectional affix versus items that were monomorphemic (including the pseudo past tense forms), (2) word type: real words versus nonwords, (3) rhyme pattern: items that are compatible with an inflection (including pseudo past forms) versus all other items, (4) place of articulation: items that end in a coronal consonant versus all other items, (5) voice: items that end in a voiceless consonant

4 For word 1: lemma frequency noun r(379) = -.04, p = .47, lemma

frequency verb r(379) = -.07, p = .15, word form frequency r(379) = -.06, p = .21, familiarity r(379) = -.01, p = .43, imageability r(379) = -.08, p = .13; for word 2: lemma frequency noun r(379) = -.01, p = .93, lemma frequency verb r(379) = -.08, p = .13, word form frequency r(379) = -.01, p = .87, familiarity r(379) = -.03, p = .62, imageability r(379)= -.08, p = .10. The correlation is negative for inverse transformed reaction times.

versus items that end in a voiced consonant (including bisyllabic past and progressive forms), (6) syllabicity: mono- versus bisyllabic forms, and (7) manner of articulation: forms that end in a stop (including bisyllabic past tense forms) versus all other forms. We also included all interactions with judgment type (same or different item), because of the presence of the judgement type/condition interaction in the overall analysis of variance. The most important question here was whether the rhyme pattern effect varied as a function of the judgement required. The results are summarised in Table 5.

The duration of the second word, the rhyme pattern, syllabicity and judgment type all contributed significantly to response times in our experiment but morphological status, voice, word type, and place and manner of articula-

Table 5

Regression analysis of reaction time including all interactions with judgment type: standardized coefficients with t values and significance levels

St. Coeff. Beta t Sig.

Level 1: regressors

(Constant) 58.89 p < .001

Duration word 2 -0.63 -18.78 p < .001

Morphological structure 0.00 -0.09 p = .92

Word type 0.02 0.54 p = .59

Rhyme pattern -0.35 -6.79 p < .001

Place (coronal or not) 0.03 0.71 p = .48

Voice 0.07 1.53 p = .13

Syllabicity 0.41 9.64 p < .001

Manner (stop or not) -0.05 -1.21 p = .23

Judgment type -0.71 -3.36 p < .001

Level 2: Interactions with judgment type

Duration word 2 x judgment type 0.06 1.23 p = .22

Morphological structure x judgment -0.03 -0.75 p = .45

Word type x judgment type -0.24 -4.33 p < .001

Rhyme pattern x judgment type -0.02 -0.43 p = .67

Place (coronal or not) x judgment 0.10 2.23 p < .05

Voice x judgment type 0.04 0.92 p = .36

Syllabicity x judgment type 0.04 1.04 p = .30

Manner (stop or not) x judgment 0.61 2.98 p < .001

ra i.i *

■a E

1000 975 950 925 900

£ ui <n «

'p' 'k'

■g 0 ■g 0 ■g

c 'o :> c 'o :> c

Z o o + o o

850 825 800

Fig. 1. Harmonic estimated marginal mean reaction times for the conditions testing coronality and the pattern of the rhyme (different items only).

<u g E

c ra 2E

1025 1000 975 950 925 900 875 850 825 800

'p' 'k'

Fig. 2. Harmonic estimated marginal mean reaction times for the voiceless items in the conditions testing coronality and rhyme pattern (different items only).

tion do not. The results also show (Table 5, level 2) that judgment type does not interact significantly with rhyme pattern (for estimated marginal means for same and different items in each condition see Appendix B). Although a number of other factors did interact with judgement type, none of these seem to bear significantly on the main questions at issue here.5

For ease of visualisation, Fig. 1 shows the effect of the rhyme pattern in terms of the relevant experimental conditions. The mean reaction times in the figure have been corrected for the effect of the duration of the second word by calculating estimated marginal means in an analysis of variance with the fixed factor condition (14 levels) and the duration of the second word as a covariate [condition: F(13,313) = 12.66, p < .001; duration word 2: F(1,313) = 109.53, p< .001] (see Appendix B for same items).

The cluster on the left-hand side represents real, pseudo and nonword regular past forms which have the rhyme pattern, and they clearly have longer reaction times than the other clusters which represent uninflected forms that do not have the rhyme pattern (all p < .01 in Bonferroni post hoc comparisons, except for nonword /t/ /d/ versus nonword /t/ p > .10). This effect occurs irrespective of coronality, morphological status and word type.

Since voicing affected reaction times in the regression analysis (although only for different items), and, as mentioned in Section 2 above, the conditions were not matched for voicing, the question arises how this affects the mor-pho-phonological effect of the rhyme pattern. Fig. 2 shows the mean values for the same conditions when all voiced items have been excluded, estimated in an analysis of var-

5 The significant interaction between judgment type and word type can be attributed to smaller differences between same and different items for nonwords than for real words; the interaction with place of articulation reflects equally fast responses to same and different items for the /k/ and /p/ conditions compared to differences between the judgement types in the other conditions, and similarly, the interaction with manner reflects a fairly consistent difference between same and different responses for all items in /s/ and /z/, but varied differences between judgment types for the other conditions.

iance with the fixed factor condition (10 levels) and the duration of the second word as a covariate [condition: F(9,147) = 12.42, p <.001; duration word 2: F(1,147) = 46.36, p < .001]6 (see Appendix B).

The figure shows that voicing has an effect, with increased response times in real, pseudo and nonword past items when voiced items are removed from these conditions. As a result, the differences between the conditions with and without the rhyme pattern are even larger than those estimated in Fig. 1, corresponding to a robust mor-pho-phonological effect (all p < .001 in Bonferroni post hoc comparisons). This enhanced effect indicates that the lack of matching on voicing across conditions in fact biased the results against our morpho-phonological hypothesis, rather than creating a confound.

Fig. 3 illustrates that the effect of the rhyme pattern is not restricted to real and pseudo past tense forms, but also applies to the s inflection forms. The means in this figure were estimated in the same covariate analysis used for Fig. 1 (see Appendix B). Since manner of articulation did not contribute to reaction times, we can conclude that the morpho-phonological effect of the rhyme pattern appears to extend to all monosyllabic inflections in the experiment (all p < .01 in Bonferroni post hoc comparisons, except for nonword /t/ versus nonword /s/ and /z/, and nonword /t/ versus nonword /t/ and /d/, which were not significant).

Finally, Fig. 4 illustrates the purely phonological effect of syllabicity on response times when word 2 duration is taken into account (again estimated in the covariate analysis used for Fig. 1; see Appendix B). Syllabic inflections are responded to faster than nonsyllabic ones (all p < .05 in Bonferroni post hoc comparisons, except for pseudo past versus progressive p > .10, and nonword /t/ and /d/ versus nonword /Id/ p > .10).

6 Since the values were estimated in a different analysis, they are

different from those in Fig. 1 for all conditions, even when no items were removed from the condition (e.g., [+Cor, -VoiceAgreement] /t/). Note that there are unequal numbers of items in the conditions in Fig. 2.

■Я 975

<u 925

E | 900

1Л С

'5 875 re

g 850 825

't' 'p' 'k'

Ф ■g Ф ■g

'n * 'о is с

:> с >

n Z о Z

и + о

Fig. 3. Harmonic estimated marginal means for the conditions testing manner of articulation (inflectional paradigm), coronality and rhyme pattern (different items only).

S i.i *

<u g E

E гё

С я | E

Fig. 4. Harmonic estimated marginal mean reaction times for the conditions testing syllabicity (different items only).

3.2. Errors

Although the error rates were very low, the same set of analyses was carried out on the arcsine transformed proportion correct responses. The subjects and items analyses showed a significant effect of condition [F1(13,247) = 5.06, MSE = 0.251, p <.001; F2(13,312) = 1.82, MSE = 0.144, p <.05], and an interaction between judgment type and condition [F1(13,247) = 2.99, MSE = 0.151, p <.001; F2(13,312) = 2.34, MSE = 0.186, p < .01], but no main effect of judgment type [F1(1,19) = 2.32, MSE = 0.289, p = .14; F2(1,312) = 1.42, MSE = 0.113, p = .23].

None of the nuisance variables (listed in Tables 2 and 3 above) correlated with error proportion.7 Therefore, we only included our factors of interest in the regression analysis, together with judgment type: (1) morphological status,

(2) rhyme pattern, (3) syllabicity, (4) voicing, (5) manner of articulation, (6) place of articulation, and (7) word type. Rhyme pattern (b = -0.141, p < .05) and judgment type (b = 0.226, p < .001) contributed to the error scores, and there were interactions between judgment type and rhyme pattern (b = -0.299, p < .001), judgment type and word type (b = -0.130, p <.05), judgment type and morphological structure (b = 0.159, p < .05), and judgment type and voicing (b = 0.194, p = .001; nonsignificant regressors: morphological structure p = .56, syllabicity p = .33, voicing p = .14, manner p = .88, place p = .50, word type p = .70).

The major effect of the rhyme pattern indicates that all potentially inflected forms, such as filled-fill, fails-fail and mild-mile, are more likely to be heard as same items when they are actually different, compared to unambiguously monomorphemic items like saint-sane and bank-bang. This is consistent with the view that the morpho-phono-logical properties of the incoming speech sounds automatically trigger an attempt at segmentation into a stem and an affix, but only when the rhyme pattern signals that it is compatible with an inflection.

Finally, we established that although reaction times did correlate with error proportion [r(379) = 0.31, p <.001), there was no speed-accuracy trade-off, since faster reaction times tended to coincide with lower error rates.

4. Discussion

A clear pattern of results emerges from this experiment, showing that the processing of regular English inflections is influenced by morpho-phonological as well as phonological factors. We hypothesised that the neural and functional mechanisms involved in the processing of spoken words in English are differentially engaged by monomor-phemic forms and by real and pseudo-inflected forms which show the critical diagnostic properties of morphologically complex inflectional forms. These diagnostic properties (the IRP) are voicing agreement in the syllable rhyme in combination with a coronal place of articulation for the final consonant of the (pseudo-)inflected form (e.g., filled, fails, meals or mild, but not saint or bank). When the IRP is present, and since the perceptual system cannot decide on the basis of acoustic-phonetic information alone whether a form does in fact bear an inflection, the presence of the IRP should trigger automatically an attempt at segmentation into a stem and an affix. This would not be triggered by unambiguously monomorphemic forms. We therefore predicted a morpho-phonological effect reflected in a difference between items that showed the diagnostic rhyme pattern and those that did not, regardless of their actual morphological status.

Our findings support this hypothesis. In a same-different judgment task, we found elevated judgment times for potentially inflected items compared to items that could not be interpreted as inflected forms.8 The difference be-

7 Word 2 duration p = .24; for word 1: lemma frequency noun p = .59, lemma frequency verb p = .69, word form frequency p = .18, familiarity p = .67, imageability p = .19; for word 2: lemma frequency noun p = .97, lemma frequency verb p = .54, word form frequency p = .36, familiarity p = .85, imageability p = .56.

8 There is ample evidence that inflectional morphemes are not only extracted in the same-different task used here, but in many other processing conditions which do not encourage participants to focus on the endings of the items (e.g., Lehtonen, Vorobyev, Hugdahl, Tuokkola, & Laine 2006; Tyler et al. 2002a; Wurm, 1997).

tween the potentially inflected and the uninflected forms in /t/ and /d/ shows that coronality only has an effect when it combines with voicing agreement in the rhyme, but not when it occurs on its own. The findings for the pseudo-and nonword inflections confirm that the rhyme pattern itself is the critical feature, rather than the actual morphological status of the item, and that segmentation must be automatic, since reaction times are comparable for any item that shows the diagnostic pattern, including meaningless strings of sounds and real words that are not actually inflected. This also includes present and plural inflections which also combine coronality and voicing agreement in their rhymes, but do not have the same place of articulation as the regular past tense (i.e., fricative /s/ and /z/).

In addition to the morpho-phonological effect, we found effects for two purely phonological factors that are independent of the morphological status of the word, for which we had made no predictions. The results show that when the past tense morpheme is realised as a syllable, participants are faster to make same-different judgments than when it is realised as a single segment. This effect of syllabicity extends to the progressive inflection /iq/, where reaction times were comparable to the syllabic past, and to the matched nonword conditions, which behaved just like the real word syllabic items.

This across-the-board syllabic effect is likely to reflect the greater perceptual salience of syllabic inflections (or pseudo-inflections), even before the added syllable is heard. Polysyllabic forms provide additional early phonological cues to the presence or absence of extra material at the end of the word, giving a strong indication of whether both words in the stimulus pair are the same or different. A strong early cue is the shortening of the first syllable - for example in melted relative to melt (Klatt, 1976; Lehiste, 1972) - which has been shown to affect lexical access processes in spoken word recognition (Davis, Marslen-Wilson, & Gaskell, 2002). Other early cues could involve changes in metrical structure and in syllabification.

The second phonological factor was voicing of the final segment. Overall, items which end in a voiced consonant, such as raised - raise, are responded to more quickly than items that end in a voiceless consonant like raced - race. Listeners pick up on cues to the presence or absence of a final consonant more quickly in the voiced items, possibly because cues to voicing become available early in the speech signal (e.g., Hawkins & Nguyen, 2004; Peterson & Lehiste, 1960). Such cues are exploited online to identify a word before the consonant in question is fully articulated (Warren & Marslen-Wilson, 1987; Warren & Marslen-Wil-son, 1988), although it is not clear why the same cues could not be used to determine that a voiceless item was being heard.

The voicing effect is independent of the morpho-phono-logical effect, although it introduced a bias against our hypothesis in the data. As explained in Section 2 above, the conditions did not have equal numbers of items with voiced and voiceless final consonants. Since the (pseu-do)inflected items had both voiced and voiceless endings, while uninflected items only had voiceless final consonants, the faster response times for the voiced items in the (pseudo)inflected conditions speeded up responses rel-

ative to the uninflected conditions. As a result, when voicing was eliminated as a factor, the morpho-phonological effect became even stronger, further increasing the difference between monomorphemic and morphologically complex items.

5. Conclusion

The segmentation of regular and pseudo-regular forms places specific demands on the neural and functional mechanisms involved in speech processing. We propose that the neuro-cognitive machinery underlying perceptual processing of spoken words in English involves both lexical access and automatic interpretative processing of the grammatical properties of the incoming words, where the extraction and interpretation of grammatical morphemes is triggered by morpho-phonological cues. This view is best accommodated in an account in which segmentation of incoming speech sounds into stems and affixes is distinguished from direct mapping between phonological forms and lexical representations.

This decompositional account of the processing of regular inflections in English (Marslen-Wilson & Tyler, 2007), invoking a core left fronto-temporal neural substrate, would explain the results of the neuropsychological study discussed earlier (Tyler et al., 2002), in which patients not only had processing difficulties with regular past tense forms, but showed diminished performance across the board for items that show the critical morpho-phonologi-cal properties of inflected forms. It would also accommodate findings from priming experiments with patients with a regular past tense deficit, where regularly inflected forms did not prime either morphologically related forms (jumped - jump; Tyler, de Mornay-Davies, et al., 2002a) or semantically related forms (jumped - leap; Longworth, Marslen-Wilson, Randall, & Tyler, 2005). If the patient's morpho-phonologically triggered processing of these forms is disrupted, this would impair access to stored lexical representations of underlying stems. The morpho-pho-nological account also provides an alternative interpretation of some of the findings of the neuropsycho-logical study of Bird et al. (2003), who included two conditions that contrasted in voicing agreement (an - and versus an - ant) in a similar manner to the pseudo-inflected and [+Coronal, -Voice] conditions of the present experiment. Patients' performance was worse for the former, which exhibited the critical morpho-phonological properties of inflections, than for the latter, which did not.

In fact, if the pattern of the rhyme is as strong a cue as the present findings suggest, then the processing of irregular past tense forms that have the IRP, like slept, should initially resemble that of regulars like stepped, since the perceptual system is blind to the morphological composition of the incoming signal. The change in vowel quality in irregular past forms would still provide an early cue to the presence of an irregular rather than a real regular past tense form, but the morpho-phonologically triggered attempt at decomposing the incoming string might still slow down response times relative to ''irregular irregulars" like took. This type of morpho-phonologically triggered processing may have contributed to Joanisse and Seidenberg's

recent fMRI findings, using a verb production task, which showed different activation effects in both left and right frontal regions for regulars than for irregulars, but similar activation patterns to regulars for a subset of pseudo-regular irregulars of the slept type (Joanisse & Seidenberg, 1999).

This is not to say that there is no contribution from phonology in the processing of inflected forms. On the contrary, our findings for voicing and syllabicity confirm that phonological factors do play an important role in the ability to process inflected words in English, complementing previous findings about factors such as phonotactic probabilities, various semantic properties, and stem and word form frequency, which have also been shown to impinge on inflectional processing (e.g., Baayen & Moscoso del Prado Martin, 2005; Hare, Ford, & Marslen-Wilson, 1999; Hay, 2001; Sereno & Jongman, 1997; Tabak, Schreuder, & Baayen, 2005; Vitevich & Luce, 1999; Vitevich & Luce, 2005).

In the Words and Rules and declarative/procedural frameworks, such effects have led to the proposal that regular inflections may also be lexically stored and accessed as full forms if they are highly frequent (e.g., Pinker & Ullman,

2002). By contrast, our proposal stresses the automaticity of the segmentation process, which is triggered for any potentially inflected form. In this respect, the account advocated here is more reminiscent of Taft and Forster's claims for an early, obligatory process of prefix stripping in visual word recognition (Taft & Forster, 1975) - though since the focus of the two accounts is very different, it is hard to compare them directly.

A direct comparison with the Words and Rules and declarative/procedural accounts is also difficult, since their predictions about relative processing times for morphologically simplex and complex words tend to focus on production (see e.g., Pinker and Ullman (2002)). They claim that the retrieval of an irregular form will block the formation of a regularly inflected form in production. If this blocking process transfers to word recognition, it would predict that only real inflected forms undergo decompositional processing. If so, pseudo-inflected forms (and irregulars with the rhyme pattern) should group with monomorphemic forms, but this is not corroborated by our findings.

In terms of connectionist approaches to speech comprehension, the current results have both specific and general implications. First, the results here are inconsistent with the perceptual difficulty account proposed by McClelland, Patterson and associates (e.g., McClelland & Patterson,

2003) to explain apparent selective impairments in processes involving regular inflectional morphology. This account attributes patients' difficulties with such forms to the additional ''phonological complexity and perceptual subtlety" of words ending in coronal obstruents such as [t] and [d]. In the current research we see no evidence for greater processing demands associated with the presence of coronality or obstruency per se. Increases in response time and error rate are only seen when these phonemic elements occur in the wider context of the inflectional rhyme pattern, indicating that it is the broader functional context that is relevant here. This broader context seems to have specific morpho-phonological and mor-pho-syntactic properties, which are both aspects of lexical

processing systems that are assigned epiphenomenal status in current connectionist thinking.

These difficulties for the McClelland and Patterson account do not, of course, mean that connectionist accounts cannot be proposed which do incorporate the notion of the IRP, so that the statistical regularities associated with its distribution in English are used to influence network performance. In so far as these regularities have a signalling function that is specifically morphological in nature (as indicated by current results), then such augmented connectionist models would also need some way of capturing these higher-order properties. One way of doing so might be via the morphological sub-regularities that are seen to emerge in some recent models (Davis, van Caster-en, & Marslen-Wilson, 2003; Plaut & Gonnerman, 2000; Rueckl & Raveh, 1999). It is less clear how the behaviour of the pseudo-regular pairs could be captured, since semantically they are like the noninflected word pairs, but nonetheless group with the real inflected pairs in the present experiment (though see Plaut and Gonnerman (2000)).

More generally, however, even if a connectionist learning model is implemented that is able to capture the pattern of IRP effects observed here (and in earlier research), this does not mean that we are also obliged to take on board the claims of such a model for the functional architecture of the language processing system. Claims for a single system architecture are not consistent with the extensive evidence that has now accumulated for a more complex and differentiated neuro-biological substrate for human language, especially where morpho-syntactic functions are concerned (for reviews see Marslen-Wilson (2007); Marslen-Wilson and Tyler (2007); Tyler and Mar-slen-Wilson (2008)). There is indeed room - if not necessity - for statistical learning models as part of an explanatory neuro-cognitive theory, but only if appropriately related to a processing architecture that is neuro-bio-logically plausible.

The contribution of this paper, in summary, is that it takes our understanding of the language processing system to a different level of specificity, forcing us to be more explicit about the conditions in which we expect certain phonological and morpho-phonological segmentation processes to occur, and how they may impinge on each other. Clearly, a strictly modular interpretation of a dual mechanism account would not have the flexibility to accommodate the multiple interacting factors that play a role during the retrieval of word forms, where some factors would appear to affect all forms regardless of their morphological status, but others uniquely surface when potentially inflected forms are encountered. Rather, as Bybee and McClelland (2005, p399) point out, the processing of these items must use general and specific information simultaneously, without presupposing that they are mutually exclusive. Whatever its underlying computational properties, the processing system will have to be able to handle morpho-phonologically cued information about morphological complexity and more direct mapping between incoming form and lexical representation, while allowing for phonological, semantic and lexical factors to take effect.

Such conflicting pressures on the processing system could perhaps be insightfully considered in an optimality theoretic framework, as suggested by Burzio (2002), but as this study shows, the account will have to allow for reference to the morphological properties of words (e.g., in stratal optimality theory; Bermudez-Otero, 1999; Kipar-sky, 2000). The probabilistic nature of various factors in inflectional processing emphasised by Baayen and Moscos-o del Prado Martin (2005), for instance, could also be represented if the conflicting constraints are assumed to be probabilistic in nature (e.g., Boersma and Hayes (2001)). However, regardless of the framework chosen, our study shows that the account will have to be informed by a more

Appendix A. List of items

systematic examination of the scope of morpho-phonolog-ical effects in various inflectional paradigms in typological-ly different languages.

Acknowledgements

We would like to thank Harald Baayen and two anonymous reviewers for their helpful comments and suggestions. This work was supported by UK Medical Research Council funding to WMW (U.1055.04.002.00001.01) and LKT (G0500842), as well as a Small Research Grant from the British Academy to LKT.

Real regular past tense in /t/ and /d/ Nonwords in /t/ and /d/

Word 1 Word 2 Allomorph CV struct Word 1 Word 2 Ending CV struct

1 prayed/preyed pray/prey d CCVC cayed cay d CVC

2 plied ply d CCVC gared gare d CVC

3 laid lay d CVC geed ghee d CVC

4 thawed thaw/thor d CVC miered mier d CVC

5 paid pay d CVC kied kie d CVC

6 bowed (au) bow (au) d CVC kushed kush t CVCC

7 wooed woo d CVC larned larn d CVCC

8 filled fill d CVCC meast miece t CVCC

9 rubbed rub d CVCC minned min d CVCC

10 warned warn/worn d CVCC parft parf t CVCC

11 joined join d CVCC foped fope t CVCC

12 raised raise d CVCC dapped dap t CVCC

13 blessed bless t CCVCC hessed hess t CVCC

14 clapped clap t CCVCC rawled rawl d CVCC

15 thrashed thrash t CCVCC soined soin d CVCC

16 wiped wipe t CVCC wobbed wob d CVCC

17 soaked soak t CVCC gubbed gub d CVCC

18 seeped seep t CVCC keered keer d CVCC

19 coped cope t CVCC boaked boak t CVCC

20 gaped gape t CVCC bloud blou d CCVC

21 pushed push t CVCC chayed chay d CCVC

22 washed wash t CVCC starced starce t CCVCC

23 ceased cease t CVCC stessed stess t CCVCC

24 pierced pierce t CVCC jiped jipe t CVCC

Noncoronal in /p/ and /k/ Nonwords in /p/ and /k/

Word 1 Word 2 Ending CV struct Word 1 Word 2 Ending CV struct

1 peak/peek pea/pee k CVC bjupe bew p CCVC

2 bark bar/baa k CVC clope clow p CCVC

3 perk purr k CVC grulp grull p CCVCC

4 mark mar/ma k CVC qump qum p CCVCC

5 whelk well k CVCC plamp plam p CCVCC

6 bank bang k CVCC fupe foo p CVC

7 milk mill k CVCC rark rar k CVC

8 hulk hull k CVCC gurk gur k CVC

9 silk sill k CVCC sark sar k CVC

10 wink wing k CVCC durk durr k CVC

11 rink ring/wring k CVCC pelk pell k CVCC

12 kink king k CVCC tulk tull (hull) k CVCC

13 troop/troupe true p CCVC fink fing k CVCC

14 slope slow/sloe p CCVC mank mang k CVCC

Noncoronal in /p/ and /k/ Nonwords in /p/ and /k/

Word 1 Word 2 Ending CV struct Word 1 Word 2 Ending CV struct

15 clamp clam p CCVCC bink bing k CVCC

16 plump plum/plumb p CCVCC lilk lill k CVCC

17 cramp cram p CCVCC hink hing k CVCC

18 soap sew/so p CVC shulk shull k CVCC

19 hemp hem p CVCC famp fam p CVCC

20 lamp lamb p CVCC wump wum p CVCC

21 damp dam/damn p CVCC lemp lem p CVCC

22 hump hum p CVCC simp sim p CVCC

23 rump rum p CVCC mamp mam p CVCC

24 gulp gull p CVCC bamp bam p CVCC

Nonpseudo-regulars in /t/ Nonwords in /t/

Word 1 Word 2 Ending CV struct Word 1 Word 2 Ending CV struct

1 saint sane t CVCC twight twy t CCVC

2 cult cull t CVCC steet stee t CCVC

3 tint tin t CVCC prunt prun t CCVCC

4 dent den t CVCC prilt prill t CCVCC

5 bolt bowl t CVCC crilt crill t CCVCC

6 punt pun t CVCC nart nar t CVC

7 rent wren t CVCC lart lar t CVC

8 hilt hill t CVCC shaint shane t CVC

9 guilt gill t CVCC shayt shay t CVC

10 dint din t CVCC deet dee t CVC

11 tilt till t CVCC haint hain t CVCC

12 pint pine t CVCC sult sull t CVCC

13 belt bell t CVCC rint rin t CVCC

14 tent ten t CVCC nent nen t CVCC

15 light lie t CVC ghent ghen t CVCC

16 cart car t CVC nint nin t CVCC

17 neat knee t CVC sant san t CVCC

18 port pour/paw/poor t CVC kint kine t CVCC

19 bait/bate bay t CVC rult rull t CVCC

20 stilt still t CCVCC yilt yill t CVCC

21 quilt quill t CCVCC ghelt ghell t CVCC

22 stunt stun t CCVCC shent shen t CVCC

23 slight sly t CCVC shilt shill t CVCC

24 start star t CCVC lunt lun t CVCC

Pseudo-regulars in /t/ and /d/

Word 1 Word 2 Ending CV struct

1 ford for d CVC

2 wand wan d CVCC

3 wide why d CVC

4 field feel d CVCC

5 tweed twee d CCVC

6 cord core d CVC

7 beard beer d CVC

8 weird weir d CVC

9 gold goal d CVCC

10 mild mile d CVCC

11 proud prow d CCVC

12 wind win d CVCC

13 cold coal d CVCC

(continued on next page)

Pseudo-regulars in /t/ and /d/

Word 1 Word 2 Ending CV struct

14 bald ball/bawl d CVCC

15 fund fun d CVCC

16 mould mole d CVCC

17 rift riff t CVCC

18 graft graph t CCVCC

19 chest chess t CVCC

20 bust bus t CVCC

21 fast farce t CVCC

22 crest cress t CCVCC

23 deft deaf t CVCC

24 tuft tough t CVCC

Present in /s/ and /z/ Nonwords in /s/ and /z/

Word 1 Word 2 Allomorph CV struct Word 1 Word 2 Ending CV struct

1 gloats gloat s CCVCC groys groy z CCVC

2 sniffs sniff s CCVCC graws graw z CCVC

3 flips flip s CCVCC plocks plock s CCVCC

4 licks lick s CVCC throcks throck s CCVCC

5 mocks mock s CVCC glips glip s CCVCC

6 tucks tuck s CVCC fies fie z CVC

7 picks pick s CVCC koes ko z CVC

8 bakes bake s CVCC sares sare z CVC

9 pokes poke s CVCC tays tay z CVC

10 wraps/raps wrap/rap s CVCC tause tau z CVC

11 reaps reap s CVCC pives pive z CVCC

12 rips rip s CVCC baves bave z CVCC

13 stirs stir z CCVC cuvs cuv z CVCC

14 stows stow z CCVC dags dag z CVCC

15 dies/dyes die/dye z CVC dapes dape z CVCC

16 mows mow z CVC bips bip s CVCC

17 gnaws gnaw/nor z CVC pakes pake s CVCC

18 shows show z CVC soats soat s CVCC

19 coos/coups coo z CVC bots bot s CVCC

20 lives live z CVCC gucks guck s CVCC

21 fails fail z CVCC bicks bick s CVCC

22 begs beg z CVCC dakes dake s CVCC

23 saves save z CVCC hicks hick s CVCC

24 nags nag z CVCC lats lat s CVCC

Plural in /s/ and /z/

Word 1 Word 2 Ending CV struct

1 threats threat s CCVCC

2 grapes grape s CCVCC

3 snouts snout s CCVCC

4 lips lip s CVCC

5 debts debt s CVCC

6 cats cat s CVCC

7 shirts shirt s CVCC

8 lakes lake s CVCC

9 goats goat s CVCC

10 hips hip s CVCC

11 shots shot s CVCC

Appendix A (continued)

Plural in /s/ and /z/

Word 1 Word 2 Ending CV struct

12 paths path s CVCC

13 dunes dune z CCVC

14 globes globe z CCVC

15 cows cow z CVC

16 psalms psalm z CVC

17 doors door z CVC

18 thorns thorn z CVC

19 whims whim z CVC

20 birds bird z CVCC

21 cones cone z CVCC

22 barns barn z CVCC

23 meals meal z CVCC

24 yards yard z CVCC

Real regular past in /Id/ Nonword /Id/

Word 1 Word 2 Allomorph CV struct Word 1 Word 2 Ending CV struct

1 greeted greet Id CCVC(VC) scaded scade Id CCVC

2 fretted fret Id CCVC(VC) snaiting snait Id CCVC

3 flaunted flaunt Id CCVCC(VC) clended clend Id CCVCC

4 jilted jilt Id CCVCC(VC) flanted flant Id CCVCC

5 scolded scold Id CCVCC(VC) groasted groast Id CCVCC

6 knitted knit/nit Id CVC(VC) ketted ket Id CVC

7 waded wade/weighed Id CVC(VC) reeted reet Id CVC

8 waited/weighted wait/weight Id CVC(VC) yitted yit Id CVC

9 faded fade Id CVC(VC) bodded bod Id CVC

10 nodded nod Id CVC(VC) saded sade Id CVC

11 lasted last Id CVCC(VC) maisted maist Id CVCC

12 melted melt Id CVCC(VC) woasted woast Id CVCC

13 welded weld Id CVCC(VC) waunted waunt Id CVCC

14 wanted want/wont Id CVCC(VC) pilded pild Id CVCC

15 basted baste/based Id CVCC(VC) bulted bult Id CVCC

16 mended mend Id CVCC(VC) relted relt Id CVCC

17 wilted wilt Id CVCC(VC) wended wend Id CVCC

18 wielded wield Id CVCC(VC) sested sest Id CVCC

19 gilded gild/guild Id CVCC(VC) walded wald Id CVCC

20 wafted waft Id CVCC(VC) munted munt Id CVCC

21 pelted pelt Id CVCC(VC) banted bant Id CVCC

22 ranted rant Id CVCC(VC) bafted baft Id CVCC

23 boasted boast Id CVCC(VC) gielded gield Id CVCC

24 shunted shunt Id CVCC(VC) milted milt Id CVCC

Progressive aspect /irj/ Nonword /iq/

Word 1 Word 2 Allomorph CV struct Word 1 Word 2 Ending CV struct

1 drowning drown /IQ/ CCVC(VC) cleading clead ing CCVC

2 pleading plead /11]/ CCVC(VC) cleeming cleem ing CCVC

3 cleansing cleanse /IQ/ CCVCC(VC) ploaxing ploax ing CCVCC

4 clanking clank /IQ/ CCVCC(VC) slending slend ing CCVCC

5 slinking slink /IQ/ CCVCC(VC) frinking frink ing CCVCC

6 hurting hurt /IQ/ CVC(VC) bowning bown ing CVC

7 seeming seem/seam /IQ/ CVC(VC) dading dade ing CVC

8 lurking lurk /IQ/ CVC(VC) surking surk ing CVC

9 teasing tease /IQ/ CVC(VC) measing mease ing CVC

10 leaning lean /IQ/ CVC(VC) werting wert ing CVC

(continued on next page)

Progressive aspect /iq/ Nonword /iq/

Word 1 Word 2 Allomorph CV struct Word 1 Word 2 Ending CV struct

11 yanking yank /II]/ CVCC(VC) fanking fank ing CVCC

12 solving solve /IQ/ CVCC(VC) folving folve ing CVCC

13 thanking thank /II]/ CVCC(VC) panking pank ing CVCC

14 founding found /II]/ CVCC(VC) lounding lound ing CVCC

15 resting/wresting rest/wrest /IQ/ CVCC(VC) tarsting tarst ing CVCC

16 tending tend /II]/ CVCC(VC) drending drend ing CVCC

1l bending bend /IQ/ CVCC(VC) sunching sunch ing CVCC

18 coaxing coax/cokes /II]/ CVCC(VC) rouncing rounce ing CVCC

19 sounding sound /II]/ CVCC(VC) fasking fask ing CVCC

20 yielding yield /IQ/ CVCC(VC) fifting fift ing CVCC

21 sifting sift /II]/ CVCC(VC) selving selve ing CVCC

22 fending fend /IQ/ CVCC(VC) loasting loast ing CVCC

23 bouncing bounce /IQ/ CVCC(VC) touncing tounce ing CVCC

24 folding fold /IQ/ CVCC(VC) yixing yix ing CVCC

Appendix B. Harmonic estimated marginal mean reaction times for condition estimated in two separate analyses of variance for different and same items with fixed factor condition (14 levels) and covariate word 2 duration

Condition Mean (ms) different itemsa Mean (ms) same itemsb

Regular past /t/ /d/ filled-fill 96l 926

Pseudo past /t/ /d/ mild-mile 977 907

[-Cor, -VoiAgree] /k, p/ lamp-lamb 827 869

[+Cor, -VoiAgree] /t/ belt-bell 823 856

Present /s/ /z/ fails-fail 935 889

Plural /s/ /z/ meals-meal 935 884

Regular past /Id/ folded-fold 809 805

Progressive /iq/ fending-fend 835 837

Nonword /d/ /t/ gubbed-gub 922 899

Nonword /t/ (not past) steet-stee 858 896

Nonword /p/ /k/ wump-wum 838 872

Nonword /s/ /z/ pakes-pake 956 879

Nonword /Id/ milted-milt 832 836

Nonword /iq/ sunching-sunch 829 815

a Condition: F(13,313) = 12.66, p < .001; word 2 duration: F(1,313) = 109.53, p < .001. Covariates appearing in the model are evaluated at the following values: duration word 2 = 696.14.

b Condition: F(13,319) = 5.25, p < .001; word 2 duration: F(1,319) = 239.66, p < 001. Covariates appearing in the model are evaluated at the following values: duration word 2 = 772.03.

References

Baayen, R. H., & Moscoso del Prado Martín, F. (2005). Semantic density and past-tense formation in three Germanic languages. Language, 81(3), 666-698.

Baayen, R. H., Piepenbrock, R., & Gulikers, L. (1995). The CELEX Lexical Database [CD-ROM]. Philadelphia, PA: Linguistic Data Consortium, University of Pennsylvania..

Bermúdez-Otero, R. (1999). Constraint interaction in language change. Opacity and globality in phonological change. Unpublished doctoral dissertation, University of Manchester/Universidad de Santiago de Compostela.

Bird, H., Lambon Ralph, M. A., Seidenberg, M. S., McClelland, J. L., & Patterson, K. (2003). Deficits in phonology and past-tense morphology: What's the connection? Journal of Memory and Language, 48(3), 502-526.

Boersma, P., & Hayes, B. (2001). Empirical tests of the gradual learning algorithm. Linguistic Inquiry, 32, 45-86.

Burzio, L. (2002). Missing players: Phonology and the past-tense debate. Lingua, 112, 157-199.

Bybee, J., & McClelland, J. L. (2005). Alternatives to the combinatorial paradigm of linguistic theory based on domain general principles of human cognition. The Linguistic Review, 22, 381-410.

Clahsen, H. (1999). Lexical entries and rules of language: A multi-disciplinary study of German inflection. Behavioral and Brain Sciences, 22(6), 991-1013.

Davis, M. H., Marslen-Wilson, W. D., & Gaskell, M. G. (2002). Leading up the lexical garden-path: Segmentation and ambiguity in spoken word recognition. Journal of Experimental Psychology: Human Perception and Performance, 28, 218-244.

Davis, M. H., van Casteren, M., & Marslen-Wilson, W. D. (2003). Frequency effects in processing inflected Dutch nouns: A distributed connectionist account. In R. H. Baayen & R. Schreuder (Eds.), Morphological structure in language processing (pp. 427-462). Berlin: Mouton.

Forster, K. I., & Forster, J. C. (2003). DMDX: A windows display program with millisecond accuracy. Behaviour Research Methods, Instruments, and Computers, 35,116-124.

Gimson, A. C. (1961). An introduction to the pronunciation of English. London: Edward Arnold.

Hare, M., Ford, M., & Marslen-Wilson, W. D. (1999). Ambiguity and frequency effects in regular verb inflection. In J. Bybee & P. Hopper (Eds.), Frequency and the emergence of linguistic structure (pp. 181-200). Amsterdam: John Benjamins.

Hawkins, S., & Nguyen, N. (2004). Influence of syllable-coda voicing on the acoustic properties of syllable-onset /l/ in English. Journal of Phonetics, 32, 199-231.

Hay, J. (2001). Lexical frequency in morphology. Is everything relative? Linguistics, 39, 1041-1070.

Hume, E., Johnson, K., Seo, M., & Tserdanelis, G. (1999). A cross-linguistic study of stop place perception. Proceedings of the International Congress of the Phonetic Sciences, 2069-2072.

Joanisse, M. F., & Seidenberg, M. S. (1999). Impairments in verb morphology after brain injury: A connectionist model. Proceedings of the National Academy of Sciences, 96, 7592-7597.

Jun, J. (1995). Perceptual and articulatory factors in place assimilation: An Optimality Theoretic approach. Doctoral dissertation, University of California, Los Angeles.

Kiparsky, P. (2000). Opacity and cyclicity. The Linguistic Review, 17, 351-365.

Klatt, D. (1976). Linguistic uses of segmental duration in English: Acoustic and perceptual evidence. Journal of the Acoustical Society of America, 59, 1208-1221.

Kochetov, A. (2004). Perception of place and secondary articulation contrasts in different syllable positions: Language-particular and language-independent asymmetries. Language and Speech, 47(4), 351-382.

Lehiste, I. (1972). The timing of utterances and linguistic boundaries. Journal of the Acoustical Society of America, 51 , 2018-2024.

Lehtonen, M., Vorobyev, V. A., Hugdahl, K., Tuokkola, T., & Laine, M. (2006). Neural correlates of morphological decomposition in a morphologically rich language: An fMRI study. Brain and Language, 98, 182-193.

Longworth, C. E., Marslen-Wilson, W. D., Randall, B., & Tyler, L. K. (2005). Getting to the meaning of the regular past tense: Evidence from neuropsychology. Journal of Cognitive Neuroscience, 17(7), 1087-1097.

Maddieson, I. (1985). Patterns of sounds. Cambridge: Cambridge University Press.

Marshall, C., & van der Lely, H. (2006). A challenge to current models of past tense inflection: The impact of phonotactics. Cognition, 100, 302-320.

Marslen-Wilson, W. D. (2007). Morphological processes in language comprehension. In G. Gaskell (Ed.), Oxford handbook of psycholinguistics (pp. 175-193). Oxford: OUP.

Marslen-Wilson, W. D., & Tyler, L. K. (1997). Dissociating types of mental computation. Nature, 387, 592-594.

Marslen-Wilson, W. D., & Tyler, L. K. (1998). Rules, representations, and the English past tense. Trends in Cognitive Science, 2, 428-435.

Marslen-Wilson, W. D., & Tyler, L. K. (2003). Capturing underlying differentiation in the human language system. Trends in Cognitive Sciences, 7(2), 62-63.

Marslen-Wilson, W. D., & Tyler, L. K. (2007). Morphology, language, and the brain: A decompositional substrate for language comprehension. Philosophical Transactions of the Royal Society B: Biological Sciences, 362, 823-836.

McClelland, J., & Patterson, K. (2003). Differentiation and integration in human language. Reply to Marslen-Wilson and Tyler. Trends in Cognitive Sciences, 7(2), 63-64.

Paradis, C., & Prunet, J. F. (Eds.). (1991). The special status of coronals: Internal and external evidence (phonology and phonetics 2). San Diego: Academic Press.

Peterson, G. E., & Lehiste, L. (1960). Duration of syllable nuclei in English. Journal of the Acoustical Society of America, 32, 693-703.

Pinker, S. (1991). Rules of language. Science, 253, 530-535.

Pinker, S. (1999). Words and rules: The ingredients of language. New York: HarperCollins.

Pinker, S., & Ullman, M. T. (2002). The past and future of the past tense. Trends in Cognitive Sciences, 6(11), 456-463.

Plaut, D. C., & Gonnerman, L. M. (2000). Are non-semantic morphological effects incompatible with a distributed connectionist approach to lexical processing? Language and Cognitive Processes, 15, 445-485.

Plunkett, K., & Marchman, V. (1993). From rote learning to system building: Acquiring verb morphology in children and connectionist nets. Cognition, 48, 21-69.

Ratcliff, R. (1993). Methods for dealing with reaction time outliers. Psychological Bulletin, 114, 510-532.

Rueckl, J. G., & Raveh, M. (1999). The influence of morphological regularities on the dynamics of a connectionist network. Brain and Language, 68,110-117.

Rumelhart, D. E., & McClelland, J. L. (1986). On learning the past tenses of English verbs. In J. L. McClelland, D. E. Rumelhart, & the PDP Research Group (Eds.), Parallel distributed processing: Explorations in the microstructures of cognition (Vol. 2, pp. 216-271). Cambridge: Bradford/MIT Press.

Sereno, J., & Jongman, A. (1997). Processing of English inflectional morphology. Memory and Cognition, 25, 425-437.

Stemberger, J. P. (1991). Apparent anti-frequency effects in language production: The Addition Bias and phonological underspecification. Journal of Memory and Language, 30, 161-185.

Tabak, W., Schreuder, R., & Baayen, R. H. (2005). Lexical statistics and lexical processing: Semantic density, information complexty, sex, and irregularity in Dutch. In S. Kepser & M. Reis (Eds.), Linguistic evidence: Empirical, theoretical, and computational perspectives (pp. 529-555). The Hague: Mouton.

Taft, M., & Forster, K. I. (1975). Lexical storage and retrieval of prefixed words. Journal of Verbal Learning and Verbal Behavior, 14, 638-647.

Tyler, L. K., de Mornay Davies, P., Anokhina, R., Longworth, C., Randall, B., & Marslen-Wilson, W. D. (2002a). Dissociations in processing past tense morphology: Neuropathology and behavioural studies. Journal of Cognitive Neuroscience, 14(1), 79-95.

Tyler, L. K., Randall, B., & Marslen-Wilson, W. D. (2002b). Phonology and neuropsychology of the English past tense. Neuropsychologia, 40, 1154-1166.

Tyler, L. K., Stamatakis, E. A., Post, B., Randall, B., & Marslen-Wilson, W. D. (2005a). Temporal and frontal systems in speech comprehension: An fMRI study of past tense processing. Neuropsychologia, 43(13), 1963-1974.

Tyler, L. K., Marslen-Wilson, W. D., & Stamatakis, E. A. (2005b). Differentiating lexical form, meaning and structure in the neural language system. Proceedings of the National Academy of Sciences, 102(23), 8375-8380.

Tyler, L. K., & Marslen-Wilson, W. D. (2008). Fronto-temporal brain systems supporting spoken language comprehension. Philosophical Transactions of the Royal Society of London B: Biological Sciences, 363, 1037-1054. Theme Issue 'The perception of speech: From sound to meaning'.

Ullman, M. T. (2004). Contributions of memory circuits to language: The declarative/procedural model. Cognition, 92, 231-270.

Ullman, M. T., Corkin, S., Coppola, M., Hickok, G., Growdon, J. H., Koroshetz, W. J., et al (1997). A neural dissociation within language: Evidence that the mental dictionary is part of declarative memory, and that grammatical rules are processed by the procedural system. Journal of Cognitive Neuroscience, 9, 266-276.

Vitevich, M. S., & Luce, P. A. (1999). Probabilistic phonotactics and spoken word recognition. Journal of Memory and Language, 40, 374-408.

Vitevich, M. S., & Luce, P. A. (2005). Increases in phonotactic probability facilitate spoken nonword repetition. Journal of Memory and Language, 52, 193-204.

Warren, P., & Marslen-Wilson, W. D. (1987). Continuous uptake of acoustic cues in spoken word-recognition. Perception &Psychophysics, 41, 262-275.

Warren, P., & Marslen-Wilson, W. D. (1988). Cues to lexical choice: Discriminating place and voice. Perception & Psychophysics, 43, 21-30.

Wiik, K. (1965). Finnish and English vowels. Turku.

Wilson, M. D. (1988). The MRC Psycholinguistic Database: Machine readable dictionary, version 2. Behavioural Research Methods, Instruments and Computers, 20(1), 6-11.

Wurm, L. H. (1997). Auditory processing of prefixed English words is both continuous and decompositional. Journal of Memory and Language, 37, 438-461.