do not necessarily reflect the views of UKDiss.com.
What links snow to dough: Investigating the processing mechanisms for spoken complex words
The properties of the English past tense enable us to probe both the nature of the mental computation of language as well as the nature of lexical representation. A plethora of behavioural research has attempted to determine the presence of independent morphological processing as well as compare and contrast the way that regular and irregular verbs are processed and represented. However, the dominant methodology, namely morphological priming remains subject to the objections of joint interaction theorists, postulating that morphological effects can be reduced to the joint interactions of semantics and phonology. Utilising rhyme priming in a lexical decision task, our study removed semantic confounds while controlling for phonological factors. Ultimately providing evidence for independent morphological processing and the discociability of regular and irregular verbs free of the potential confound of semantic and phonological interaction. Suggesting that our data is best explained by underlying neurocognitive distinctions between the verb sub-types accounted for by a dual-mechanism model.
TABLE OF CONTENTS
In language the connection between sound and meaning is ultimately arbitrary, with the sound alone not encoding any lexical information. However, morphology, the way in which words are formed and structured, captures consistencies in language and thus imposes a level of systemisation onto these sound to meaning mappings. These consistencies (morphemes) have a meaning of their own that can be applied to other words. For example, the suffix -ed when found with in conjunction with a verb stem indicates that the action implied by the stem, be it to jump to climb or to study, occurred in the past. This ability to combine and convert words using morphology is critical for language productivity, especially in regards to syntax as small numbers of inflectional elements are recombined to help express infinite number of thoughts, opinions, concepts and actions.
Complex words, resulting from combining morphemes, raise questions not only about the processing and representational ramifications of morphology but also about lexical representation in entirety. Theoretically if morphological informational was processed, cognitively encoded and formed part of the structure of our lexical representations it could be used as a principle of lexical organisation and processing. In contrast to this morphemic approach it is possible that morphology does not form part of our lexical representations and items are stored as whole-words. Despite extensive research exactly how morphology is mapped onto our neurocognitive reality remains unclear. In particular research has focused on inflectional morphology as the properties of the English past tense allow us to directly contrast these two approaches to lexical representation. Although both regular and irregular past tense verbs carry comparable syntactic and semantic information they differ in regards to morphology. Past tense regular verbs (e.g. snowed, travelled, slipped), formed by combining a phonologically unchanged stem with the inflectional suffix –ed, are morphophono-logically overt as they can be easily decomposed back into a stem + affix format. While past tense irregular verbs (e.g. rode, sank, sung) do not have a predictable relationship between their stem and inflected form, so they cannot be decomposed in the same way into a stem + affix structure and thus are morphophono-logically hidden.
In addition to these differences in morphology, English past tense also provides a contrast between the rule-like pattern of regular verbs and an idiosyncratic irregular form. Excluding the approximately 160 irregular verbs, the formation of past tense verbs is rule-like as it involves the predictable addition of –ed to the verb stem, with the exact pronunciation (/d/, /t/, or /Id/) dictated by the phonological form of the stem.
These properties of the English past tense enable us to probe both the nature of the mental computation of language as well as the nature of lexical representation. Whether these differences in linguistic properties reflect differences in lexical representations is a subject of much debate, with major theories of lexical representation holding conflicting views.
Single mechanism models utilise a connectionist theory (Rumelhart and McClelland, 1986) holding that both regular and irregular verbs are processed by a single mechanism based on phonology and semantics. This theory suggests that morphology simply reflects the statistical correspondence between form and meaning, thus stems and inflectional morphemes are not separable and that inflected forms are learned and represented as whole forms but with overlap with other forms based on phonological and semantic similarities.
Competing views postulate that different neurocognitive mechanisms underpin the processing and representation of regular and irregular verbs. These dual mechanism theories agree that there are different processes at work, but conjecture over the finer details has led to slight variations on the theory. However, almost all agree that irregular forms must be represented as indecomposable whole forms as well as learned and processed using a connectionist mechanism.
Pinker, Ullman and colleagues (date) describe a rule-based system utilising symbolic computations to dictate the process of regular inflection. In this system the suffix –ed is automatically added to regularly inflected verbs, while irregular verbs utilise an associative memory system which encodes their past tense forms as wholes. In regards to the nature of lexical representation this model posits that regular verbs have one entry in the lexicon for the stem, while irregular verbs have two, though the past tense entry is stored separately in a pattern-association network away from its stem. As this theory makes a categorical distinction between rule-generated verbs and the irregular exceptions it would predict that there are strong dissociations between the two processes.
Marslen-Wilson and Tyler (1998) theorise that the dissociations between the two verb sub-types are not due to regulars being formed by a rule, but instead for access in comprehension they require phonological disassembly while irregulars, as they lack overt morphophono-logical structure, are accessed through a full-form route. But like the previous model agrees that irregulars unlike regular verbs do not share a lexical representation with their stems.
However, compelling neuropsychological evidence for two separable neural systems continued to mount. With numerous studies (X) demonstrating that stroke patients can have selective deficits in either irregular or regular inflection. The led to the revised single mechanism model (Joanisse and Seidenberg, 1999) where although the underlying mechanism remains connectionist there are phonological and semantic subsystems. With irregular verbs being more dependent on the meaning and thus the semantic subsystem, while regulars are more dependent on phonology and thus utilise the phonological subsystem. Thus this model attributes the neuropsychological dissociations as damage to either subsystem, while remaining a single mechanism model without Independent Morphological Processing (IMP).
Therefore to resolve this controversy a plethora of behavioural research has attempted to clarify whether IMP occurs as well as compare and contrast the way that regular and irregular verbs are processed and represented, and when contrasts are found if they can be attributed to a single-mechanism model or if they reflect underlying neurocognitive distinctions explained by a dual-mechanism model.
In regards to dissociations between the processing of regular and irregular verbs it has been generally found that regular verbs are facilitated when preceded by a morphologically related prime (Rastle et al, 2000). While irregular verbs are less consistent, sometimes producing weaker or non-existing priming effects (Stanners, Neiser, Hernon, and Hall, 1979); Napps ,1989; Kempley and Morton, 1982) but not always with many studies reporting priming effects with both regular and irregular verbs (Davis, Schoknecht, and Carter ,1987; Marslen-Wilson and Tyler, 1997; Tyler et al., 2002). Similarly the dominant methodologies for investigating IMP, such as lexical decision tasks and morpheme frequency effects (Bertram at al, 2000, Clahsen et al, 2005), are mixed. Although morphological priming has shown facilitation (Marslen-Wilson, 2007) it is difficult to dissociate the effects of morphology from the effects of phonology and semantics, thus these results remain inconclusive.
Previous methodologies aiming to elucidate the theory of IMP failed to exclude a possible interplay between morphology, semantics and phonology. As any regular past tense item (jump –ed) is phonologically, morphologically and semantically related to its stem (jump). Therefore any apparent morphological effects could be reduced to interactions between semantics and phonology, and thus remain in keeping with a single mechanism model.
Attempts to behaviourally distinguish between these alternative explanations have utilised long-distance priming, where the prime is separated from the target by a number of intervening items (Marslen-Wilson, 2007; Kouider and Dupoux, 2009). Facilitation has been reported for morphological priming at longer distances than for phonological and semantic priming, leading to the conclusion that this extended facilitation was due to the independence of morphological effects from phonology and semantics (Kouider and Dupoux, 2009). This has been criticised (Gonnerman, Seidenberg and Andersen, 2007) as the effects seen could be due to the joint interaction between semantic and phonological similarity, and that once again morphology is not processed independently.
The joint interaction theory postulates that any observed morphological effect relies on the co-presence of shared semantics and phonology. Therefore if one of these elements was removed, while controlling the other, the joint interaction theory could be rejected in favour of evidence for IMP.
A recent novel methodology removed this joint interaction confound by utilising rhyme priming (Bacovcin et al, 2017). Rhyme priming investigates facilitation effects caused by a rhyme, which in phonological terms is defined as a shared vowel and coda consonants. Rhyme priming has previously been used to explore phonological processing, demonstrating that the facilitation seen is due to the combined but ultimately separate effects of rhyme and phonological relatedness (Slowlaczek, 2000), with a shared rhyme resulting in strong facilitation while phonological relatedness only produces a weak priming effect. Adaption of this methodology has allowed a direct investigation of morphological processing, while avoiding the potential confound of joint interaction by eliminating semantic similarity and providing controls for phonological relatedness.
This study provided evidence that stem access during the processing of a morphologically complex word is morphological in nature. Facilitation was seen with a prime (dough) that rhymed with a complex word target (snowed) despite the pair being semantically and relatively phonetically unrelated. Therefore the facilitation seen reflects morphological access of the stem (snow) due to its rhyming relationship with the prime (dough) and cannot be due to an interaction of semantics and phonology.
We aim to replicate the findings of Bacovcin et al (2017) and extend their methodology to encompass irregular verbs. To test whether irregular verbs undergo IMP we will utilize a prime (pink) that rhymes with a complex past tense irregular verb target (sank). This is essentially equivalent to the (dough snowed) regular verb example except that irregular verbs are morphophono-logically hidden and thus cannot be comparably segmented. Demonstrating whether the irregular verb stem is accessed during complex word processing will provide evidence for the dissociability of the two verb sub-types. This in conjunction with potentially replicating findings regarding IMP will provide compelling behavioural evidence for a dual-mechanism model and against a single-mechanism one.
The original experiment included four conditions, two aimed to study the main effect while two provided phonological controls. The first condition, the bare stem, used a regular verb stem as the target with a rhyming prime (dough snow). This condition further replicated results from rhyme priming and acts as a point of comparison for condition two. Condition two, the past tense, consisted of the decisive morphologically complex regular verb form as the target while the prime rhymed with its stem (dough snowed).
The third and fourth conditions were controls for particular phonological confounds. The third, past tense rhyme control, aimed to show that any facilitation seen in the second condition was not due to the partial rhyme that exists between the prime and the target. This partial rhyme results from the fact that the both prime and target share a syllable nucleus. This control utilises targets that were non-morphologically complex words, words that cannot be decomposed into a stem and an affix, which shared a syllable nucleus with the prime. Alternatively the target can be thought of as a word that rhymes with the regular past tense target from condition two, e.g. (dough code, as dough and code share a syllable nucleus exemplified by the fact that code and snowed rhyme). Any facilitation seen in condition three would be due to the partial phonological overlap caused by the shared syllable nucleus, meaning that the facilitation seen in condition two could be explained by partial rhyme and not by morphological access. Condition three also aims
Condition four addresses the potential confound of word embedding. Due to the incremental nature of auditory speech processing, any facilitation seen may be due to the priming effect caused by the phonologically embedded word, and not by the morphological relationship of the target to the stem (E.g. dough snowed, any facilitation seen could be due to the phonological embedding of snow and thus the facilitation is picking up on the relationship between dough and snow, not dough and snowed). This control utilises targets that phonologically embed other words (grove embeds grow), but the embedded words are not morphologically related to the target (grow is not morphologically related to grove), while the prime rhymes with the embedded word (dough grove). This is therefore reflective of the embedding seen in condition two, thus if facilitation is seen in condition four it is evidence that any facilitation seen in condition two is due to phonological embedding resulting in stem access and not due to IMP.
Conditions five and six are the novel conditions aiming to see whether this rhyme priming methodology can be applied to irregular verbs. Condition five, essentially repeats condition one but with an irregular verb stem as the target (pink sink) and has the similar aims of replicating rhyme priming results as well as acting as a comparison for condition six. Condition six, the key experimental condition, utilises an irregular verb form as the target while the prime rhymes with its stem (pink sank). Facilitation seen here would provide evidence for IMP of irregular verbs while if not, it would provide evidence for the disscociability of regular and irregular verb processing and representation.
Fifty-six individuals (34 male, 22 female) were recruited through social media advertisements. All participants were native English speakers with no hearing or language difficulties. Prior to the experiment the Psychology Department Ethics Committee granted ethical approval. Informed consent was obtained prior to the test and consent to use the collected data was obtained at the end of the experiment. On completion the participants were remunerated 5 pounds.
The materials consisted of six conditions:
- Regular verb base stem
- Regular verb past tense
- Past tense rhyme control
- Embedded control
- Irregular verb base stems
- Irregular verb past tense
Each condition contains 20 prime-target pairs. See table (1).
||Rhymes with target||Regular verb stem||Dough Snow|
||Rhymes with targets stem||Past Tense Regular verb||Dough Snowed|
||Shares a syllable nucleus with target||Morphologically simple word||Dough Code|
||Rhymes with the embedded word in the target||Phonologically embeds another word||Dough Grove|
||Rhymes with target||Irregular Verb stem||Pink Sink|
||Rhymes with stem of target||Past tense irregular verb||Pink Sank|
Table 1: Examples and Explanations of the relationships between Prime and Target Experimental Stimuli
Using Match software (van Casteren and Davis, 2007), the target stimuli were matched as close as possible on word frequency (occurrence per million words), number of syllables and number of phonemes using data from the N-watch programme (Davis, 2005). See Table 2.
|Frequency||20.46 (21.13)||20.89 (24.17)||21.42 (21.12)||21.24
|Number of Syllables||1.10
|Number of Phonemes||3.35
Table 2: Mean stimuli properties of targets across conditions, standard deviations are included in brackets
Conditions 1-6 did not have significantly different number of syllables [F (5,114) = 1.637, p =.156]. Full matching on word frequency across all six conditions was not possible as English irregular verbs are much more frequently used than regular verbs. However, a one-way ANOVA showed that there was no significant difference between conditions 1-4 on frequency [F (3,76) = 0.007, p =.999], or 5-6 on frequency [p = .829]. The phoneme addition required in regular inflection meant that there was also a small difference in the number of phonemes across conditions [F(5,114) = 4.357, p=.001], however this will have little bearing on our experiment as we ensured that our conditions match on spoken word length.
Our primes were also matched in a similar way with a second set of one-way ANOVAs demonstrating that there was no significant difference between conditions 1-6 on frequency [F (5,114) = 0.029, p =1.00], number of phonemes [F (5,114) = 1.393, p = .232], and as all our primes were monosyllabic there was also no significant difference in syllable number.
To generate a point of comparison and therefore detect the presence of priming, each target was also paired with an unrelated prime. This way any facilitatory effect could be calculated by comparing reaction times when the target was preceded by an unrelated prime verses by a related prime.
Our unrelated primes were also matched in a similar way with a third set of one-way ANOVAs demonstrating that there was no significant difference between conditions 1-6 on frequency [F (5,114) = 1.157, p = .335], number of syllables [F (5,114) = .000, p = 1.00] or number of phonemes [F (5,114) = .814, p = .542].
To mask the aims of the conditions additional 80 filler words were added. A further 120 non-words were generated using the ARC non-word database in order to create the lexical decision task. Care was taken to ensure that some of these fillers and non-words were morphologically complex.
Lastly an additional 16 practice items were created, giving a total of 576 items. A complete list of items can be located in the Appendix.
A native speaker recorded all the stimuli in a soundproof booth. The audio files were later processed using Audacity. The amplitudes of the files were equalised to the root square mean using MatLab, thus ensuring that any amplitude differences caused by variations in the speakers voice or proximity to the recording device will not affect word processing by the listener.
The lengths of the recordings were analysed using a one-way ANOVA, demonstrating there was no main effect of the recorded target length between the different conditions [F (5, 114) =1.709, p=.138] (see Table 3).
|Average Recording length (ms)||748
|774 (149)||741 (125)||726 (138)||678 (83)|
Table 3: Target stimuli recording lengths across conditions, standard deviations are included in brackets
There was also no significant difference between the lengths of both related [F (5,114) = .452, p = .811] and unrelated primes [F (5,114) = 0.005, p = 1.00] across conditions. A repeated measures ANOVA was done to ensure that there was also no difference between our two prime types [F (1,38) = 2.144, p = .151].
To minimise target repetition, two lists were created so that each participant would only hear the target items in each condition once, paired either with a related or unrelated prime. To do so, The target stimuli in each condition were split into two sets, A and B. In list 1, set A targets were paired with related primes while set B targets were paired with unrelated primes. This was reversed in list 2. Both list versions contained 60 related prime-target pairs, 60 unrelated prime-target pairs. Each list contained all of the non-word and filler items, meaning that each list contained 440 items. An example breakdown for a single list is represented in Table 4.
|PRIME TYPE ||RELATED||UNRELATED||TOTAL|
|TARGET CONDITION |
|TOTAL||60 pairs||60 pairs||120 pairs|
Table 4: Example breakdown of the number of target prime pairs included in one list.
Each list was split into 4 blocks of 55 pairs each, while ensuring that the six experimental conditions were equally divided between these blocks. Each block was then further randomized with two practice prime target pairs placed at the beginning of each block.
Participants were randomly assigned to list 1 or list 2, and were similarly assigned a randomized block order. The differing block order, created using a Latin square design, was used to minimize potential fatigue confounds. Stimuli were presented over headphones at an appropriate listening level. Participants were asked to listen to each sound in entirety then at the end decide if they had heard a word or a non-word, pressing the keys “l“ and “a” respectively. In left-handed participants the keyboard was rotated 180 degrees to ensure that the individuals dominant hand was always making the decision for a word. Reaction times were recorded from the onset of the auditory file to ensure that even premature presses were logged. Between each block there was a subject controlled break, and in total the task took approximately 35 minutes.
Any incorrect responses were automatically excluded from the analysis. Two participants were removed, one due to equipment malfunction resulting in missing data; another due to low accuracy (15% error rate) in comparison to other participants (5%). The average error rate for each item was examined, with items with a greater than 20% error rate being excluded from further analysis (what how many). Reaction times over 2500ms were automatically excluded, as were reaction times under 500ms.
|PRIME TYPE ||Related
||528 (73)||593 (92)
||618 (68)||634 (93)|
Table 5: Summary of Reaction Times with standard deviations in parenthesises.
The data (see table 5 and figure 1) was analysed using a Mixed Design Analysis of Variance (ANOVA) with two within-subject variables; condition (six levels) and priming (two levels), and list as the between subject factor. As we must not assume that all items within a condition behave in the same way (Clark, 1973), it is common in psycholinguistics to analyse the data both by subject (F1) and by item (F2).
Due to a significant sphericity values across our analyses, Greenhouse-Geisser corrections were utilised throughout, unless stated otherwise.
There was no main effect of list in both the subject [F1 (1, 50) = .284, p = .596] EFFECT SIZE and item analysis [F2 (1, 14) = 0.440, p = .518]. There was a main effect of condition in the by subject analysis [F1 (1.340, 67.014) = 3.855, p = .042] while this was only marginally significant in the by item analysis [F2 (1.826, 25.566) = 2.968, p = .073]. The main effect of priming was also significant in the subject analysis [F1 (1, 50) = 4.796, p = .033] while also remaining marginally significant in the by items analysis [F2 (1, 14) = 4.031, p = .064]. However, crucially and disappointingly no interaction between condition and priming was observed in either our subject [F (1.232, 61.621) = .1.735, p = .193] or item analysis [F (1.531, 21.434) = .624, p = .504].
The weaker effects consistently seen in the by items analyses indicate that there is a source of noise in the item analysis that is not present in the subject analysis, such that items from the same condition may not be producing consistent effects. One possible source of this could be the confounding effect of frequency, as the frequency of targets in condition 5 and 6 was significantly different from conditions 1 – 4. To test this hypothesis, a second Mixed Design Analysis of Variance (ANOVA) was completed on the item data, removing frequency effects by including frequency as a covariant.
This second ANOVA also found no main effect of list [F2 (1, 2) = 1.072, p = .409]. Furthermore there was neither a main effect of condition [F2 (1.465, 55.655) = .493, p = .555] nor a main effect of priming [F2 (1, 2) = 2.361, p = .264]. These priming results are not unexpected, as per Bacovcin et al (2017) experiment priming was not expected in the control conditions (3 and 4) thus this would influence whether the main effect of priming is seen when looking across the conditions in entirety.
This analysis however revealed that, once the effect of frequency were taken into account, there was an interaction between prime and condition [F2 (1.938, 3.877) = 7.149, p = .050], indicating that the amount of priming differed between conditions.
To further investigate the interaction between prime and condition, and since there has consistently been no effect of list, a tertiary analysis was completed. As each target was preceded by both a related and unrelated prime a subtraction of these two mean target values (unrelated – related) was completed with the difference representing the facilitatory effect of the related prime on target response. Using this data (see figure 2) a one-way ANOVA across the six conditions was completed, with post hoc pairwise comparisons enabling examination of priming between conditions.
Figure 2: Box plot showing the effect of condition on the amount of priming (ms)
* effect of priming is significant (p < 0.01)
As expected conditions 1 (regular bare stem), 2 (regular past tense) and 5 (irregular bare stem) showed significant facilitation, while conditions 3 (partial rhyme control), 4 (embedded control), 6 (irregular past tense) did not. This pattern was corroborated by the results from the pairwise comparisons. Condition 1 and 2 were not significantly different from each other p= .782, while both conditions 1 and 2 were significantly different from the control conditions, 3 and 4, where priming did not take place (see table 6 for significance values). A similar pattern was seen for condition 5, which did not show significant difference to conditions 1 and 2 where a significant priming effect also took place, but which was significantly different from conditions 3 and 4. The results for the crucial experimental condition, condition 6, differ from the other results: although no significant priming effect was seen, no significant differences were found when compared to the other conditions, including its main point of comparison, condition 5 (p=.102). This pattern indicates that a small amount of facilitation may have occurred, although a significant level of priming was not recorded.
Table 6 Pairwise comparisons between conditions. Non-significant results are marked in red, while significant results are marked in green
To ensure full evaluation of potential sources of variance, both a by subject analysis (F1) and by item (F2) analysis was conducted on the prime reaction time data, using the same Mixed Design Analysis of Variance (ANOVA) setup with condition (six levels) and priming (two levels: related and unrelated) as within-subject variables, list as a between-subject variable, and frequency data as a covariate. The analysis revealed no main effect of list [F1 (1, 38) = 2.618, p = .114]); ([F2 (1, 1) = 2.723, p = .347], condition [F1 (2.409, 91.542) = .639, p = .558]); ([F2 (1, 1) = .270, p = .695], or priming [F1 (1, 38) = 1.550, p = .221]); ([F2 (1,1) = .881, p = .520]. There were also no significant interactions between these three factors in either analysis. Given these results no further investigation was completed.
Error analysis was completed using the same structure as the two previous Mixed Design Analysis of Variance (ANOVAs).
|LIST ||List 1||List 2||Average|
|PRIME TYPE ||Related
||5 (1)||5 (2)||2 (1)||3 (2)||3 (3)||4 (6)|
||6 (2)||8 (2)||7 (2)||5 (2)||6 (5)||7 (6)|
||6 (2)||4 (1)||4 (2)||2 (1)||5 (5)||5 (5)|
||4 (1)||2 (1)||2 (1)||3 (1)||3 (3)||3 (3)|
||4 (1)||0 (1)||0 (1)||3 (1)||2 (3)||2 (2)|
||4 (1)||3 (1)||3 (1)||2 (1)||3 (4)||3 (3)|
Table 7: Summary of errors (%), standard deviations are included in brackets
There was no main effect of list in both the subject [F1 (1, 38) = .301, p = .586] and item analysis [F2 (1, 2) = .375, p = .602]. There was no main effect of condition in the by subject analysis [F1 (3.301, 125.431) = .725, p = .551] or in the by item analysis [F2 (1.530, 3.060) = 1.559, p = .257]. The main effect of priming was also not significant in the subject [F1 (1, 38) = .147, p = .703] or the item analysis [F2 (1,2) = 0.31, p = .877]. There were also no significant interactions between these three factors. Given these results no further investigation was completed.
This study aimed to provide evidence that distinguished between the major theories of lexical processing and representation. By using rhyme priming in conjunction with a lexical decision task we investigated the presence of IMP during the lexical access of regular and irregular verbs. Importantly this methodology avoided the semantic and phonological confounds incurred by previous priming methodologies, thus providing evidence free from the objections of joint interaction theorists.
In regards to regular verbs we aimed to replicate the results of a previous study (Bacovcin, 2017), thus lending more weight to the theory of IMP. While the addition of irregular verb conditions enabled comparison between the way regular and irregular verbs are processed and represented, a potentially crucial test between different theories.
Our expectations, given recent imaging evidence (Bozic et al, 2010), was that our results would support a dual-mechanism theory, showing a pattern of facilitation indicative of IMP with our regular verb condition (2), while there would be dissociation from this pattern with our irregular condition (6) indicating that morphological processing does not play a role in the processing of irregular verbs, thus supporting a dual-mechanism model.
Independent Morphological Processing
Our results mirror those of Bacovcin and colleagues. In condition 1, participants responded faster to the regular bare stem target (snow), when preceded by a related prime (dough) in comparison to when preceded by an unrelated prime (conch) (p<0.01), thus this result reflects the facilitatory effect of the shared vowel and coda consonants that constitute a rhyme (Slowiaczek et al, 1987). Critically these effects were repeated in condition two, with participants also responding faster to the regular past tense target (snowed) when preceded by a prime that rhymed with its bare stem (dough) in comparison to an unrelated prime (p<0.01). Conditions 1 and 2 showed equivalent levels of priming (p=.782), therefore indicating that the base stem (snow) is accessed during the processing of a complex regular verb (snowed).
Importantly neither control showed a significant priming effect, therefore suggesting that neither partial rhyme (p =.771) nor word embedding (p=.734) are responsible for observed facilitatory effect in condition 2. These results are corroborated by the pairwise comparisons illustrating significant difference between both conditions 1 and 2 and 3 and 4 (see table 6).
Therefore as this methodology utilises primes that are semantically unrelated to the targets and included controls to quantify the possible, but ultimately non-existent, confounding effect of phonology, it provided further support for the theory of IMP which is unaffected by possibility that morphological effects could be attributed to connections between form and meaning. However, it is important to note that although this finding does not support models where morphology is epiphenomenal, such as the single mechanism connectionist model (Rumelhart and McClelland, 1986), it does not enable distinction between other models which evoke IMP, as although these models agree that IMP takes place there are subtle differences that this finding does not enable a distinction to be made. For example some theories postulate that complex words are decomposed into their constituent morphemes (Marslen-Wilson and Tyler, 1998) while others such as the parallel dual-route model (Schreuder and Baayen, 1995) postulate that decomposition does not occur but that the bare stems are still related though association to the complex word.
By investigating IMP our study was able to create a point of comparison for the way that regular and irregular verbs are processed and represented. In condition 5, participants responded faster to the irregular bare stem target (sink), when preceded by a related prime (pink) in comparison to when preceded by an unrelated prime (spot) (p<0.01), thus repeating the phenomenon of rhyme priming also shown in condition 1. Condition 5 also enabled a key point of comparison to condition 6, where an irregular past tense target (sank) was preceded by a prime that either rhymed with the bare stem (pink) or was unrelated (spot). Although a significant effect of priming was not see (p =.381), it did not differ significantly from the priming seen in Condition 5 (p=.102). Indeed condition 5 did not differ significantly both from conditions where had significant priming results, namely condition 1 (p=.95), 2 (p=.56) and 5 (p=.102), nor those where priming was also not seen, conditions 3 (p=.308) and 4 (p =.627). However, as there was marginal significance for both conditions 1 and 2, although the lack of priming is not as unambiguous as in the control conditions there is not sufficient evidence to claim that IMP also occurs in irregular verbs. Therefore our data indicates dissociation in the way regular and irregular verbs are processed. However, the question remains as to whether this contrast could be attributed to a single-mechanism model or if it reflects underlying neurocognitive distinctions explained by a dual-mechanism model.
It seems evident that dissociations between the two verb subclasses would exclude single mechanism accounts, however, this is not the case as it has been argued (Seidenberg and Hoeffner, 1998) that there are non-mechanistic differences between the two verb-subclasses, such as their frequency, acquisition or pattern in the corpora. They argue that the resulting behavioural evidence reflects these distributional differences opposed to genuine differences in the underlying neurocognitive machinery.
The dual mechanisms models almost all agree that irregular forms must be represented as indecomposable whole forms, and learned and processed using a connectionist learning mechanism, while regular verbs are processed by another mechanism, and thus would predict dissociations between the two verb sub-classes. However, our results do not dissociate between the different dual-mechanism approaches with any certainty, it could be argued that the slight ambiguity seen is not in keeping with a Pinker’s model, which would predict strong disscociations between the two classes. While our results offer more support to less categorical accounts such as the parallel dual-route model (Baayen et al, 1997) which postulates that lexical access has two parallel routes, one which processes on the basis of full form while the other processes on the basis of the words constituent morphemes, with the route depending on the frequency, predictability and transparency, not it’s grammatical class therefore this model does not predict a categorical distinction between verb types.
Limitations and Improvements
The English past tense has been extensively utilised due to the fact that it provides words that carry comparable syntactic and semantic information while differing in regards to morphology. However, it is not the only language that contains irregular verbs, thus this experimental methodology could be repeated in languages such as German and French which also contain both regular and irregular verbs.
We must also consider whether the categorical approach taken to regular and irregular verbs is accurate at the neurocognitive level. As although we have taken irregular to mean that conjugation does not consist of the addition of the suffix –ed irregular verbs are not completely idiosyncratic, and there are some more subtle forms of regularity in the conjugation of irregular verbs. These can include a change in consonant at the end of the word (e.g. sleep slept) or in the form of a vowel change in the middle of the word (run ran, swim swum, sing sung) there are also conjugations, which make use of both of these processes (think thought). With these regularities in mind the only true irregular verbs are those that are completely suppletive (go went). There is evidence that these linguistic equivalences affect processing (Kielar, 2008), and therefore provide evidence against the categorical distinction taken in our research. Therefore one could argue that the ambiguity of Condition 6 was due to the inclusion of some irregular verbs that behave more like regulars at the neural level due to some of these aforementioned semi-uniform conjugation patterns. Future research could tease apart this effect by utilising the same methodology but separating out the proposed three types of irregular verbs. This could potentially provide behavioural evidence that dissociates between the more categorical dual mechanism models (Pinker, Ullman) and those that can account for these effects (Marslen-Wilson Baayen et al, 1997; ).
Although this experiment utilised English inflectional morphology to replicate results surrounding IMP as it enabled a comparison for our expansion into irregular verbs, English inflectional morphology is considerable more simplistic than other many other languages. Therefore repeating this rhyme priming methodology in other languages and with other types of inflection would provide compelling cross-linguistic evidence for models of lexical representation and processing.
- Kouider and Dupoux
(Schreuder and Baayen, 1995)
Rastle et al, 2000
Stanners, Neiser, Hernon, and Hall, 1979
Kempley and Morton, 1982
(Davis, Schoknecht, and Carter ,1987
Marslen-Wilson and Tyler, 1997
Tyler et al., 2002
Bertram at al, 2000
Clahsen et al, 2005
Gonnerman, Seidenberg and Andersen, 2007
Seidenberg and Hoeffner, 1998)
Allen, M., & Badecker, W. (2002). Inflectional regularity: Probing the nature of lexical representation in a cross-modal priming task. Journal of Memory and Language, 46(4), 705–722. https://doi.org/10.1006/jmla.2001.2831
Bacovcin, H. A., Goodwin Davies, A., Wilder, R. J., & Embick, D. (2017). Auditory morphological processing: Evidence from phonological priming. Cognition, 164, 102–106. https://doi.org/10.1016/j.cognition.2017.03.011
Bozic, M., Tyler, L. K., Ives, D. T., Randall, B., & Marslen-Wilson, W. D. (2010). Bihemispheric foundations for human speech comprehension. Proceedings of the National Academy of Sciences, 107(40), 17439–17444. https://doi.org/10.1073/pnas.1000531107
Clark, H. (1973). The Language-as-Fixed-Effect Fallacy: A Critique of Language Statistics in Psychological Research. Journal of Verbal Learning and Verbal Behavior, 12, 335–359.
Cohen-Shikora, E. R., & Balota, D. A. (2013). Past tense route priming. Cognition, 126(3), 397–404. https://doi.org/10.1016/j.cognition.2012.11.009
Davis, C. J. (2005). N-Watch: A program for deriving neighbourhood size and other pyscholinguistic statistics. Behaviour Research Methods, 37(1), 65–70.
Kielar, A., Joanisse, M. F., & Hare, M. L. (2008). Priming English past tense verbs: Rules or statistics? Journal of Memory and Language, 58(2), 327–346. https://doi.org/10.1016/j.jml.2007.10.002
Lieberman, E., Michel, J. B., Jackson, J., Tang, T., & Nowak, M. A. (2007). Quantifying the evolutionary dynamics of language. Nature, 449(7163), 713–716. https://doi.org/10.1038/nature06137
Lignos, C. (2013). Modeling Words in the Mind. University of Pennsylvania.
Marslen-Wilson, W. D., & Tyler, L. K. (1997). Dissociating types of mental computation. Nature, 387(6633), 592–594. https://doi.org/10.1038/42456
Marslen-Wilson, W., & Tyler, L. K. (1998). Rules, representations, and the English past tense. Trends in Cognitive Sciences, 2(11), 428–435. https://doi.org/10.1016/S1364-6613(98)01239-X
Post, B., Marslen-Wilson, W. D., Randall, B., & Tyler, L. K. (2008). The processing of English regular inflections: Phonological cues to morphological structure. Cognition, 109(1), 1–17. https://doi.org/10.1016/j.cognition.2008.06.011
Slowiaczek, L. M., McQueen, J. M., Soltano, E. G., & Lynch, M. (2000). Phonological Representations in Prelexical Speech Processing: Evidence from Form-Based Priming. Journal of Memory and Language, 43(3), 530–560. https://doi.org/10.1006/jmla.2000.2710
Tyler, L. K., Stamatakis, E. A., Post, B., Randall, B., & Marslen-Wilson, W. (2005). Temporal and frontal systems in speech comprehension: An fMRI study of past tense processing. Neuropsychologia, 43(13), 1963–1974. https://doi.org/10.1016/j.neuropsychologia.2005.03.008
Van Casteren, M., & Davis, M. H. (2007). Match: A program to assist in matching the conditions of factorial experiments. Behavior Research Methods, 39(4), 973–978. https://doi.org/10.3758/BF03192992
Rastle, K., Harrington, J., & Coltheart, M. (2002). 358,534 nonwords: The ARC Nonword Database. Quarterly Journal of Experimental Psychology, 55A, 1339-1362.
Appendix A – Test Stimuli
|Condition||Target||Related Prime||Unrelated Prime|
|Condition||Target||Related Prime||Unrelated Prime|
|Condition||Target||Related Prime||Unrelated Prime|
|Condition||Target||Related Prime||Unrelated Prime|
|Condition||Target||Related Prime||Unrelated Prime|
Appendix B – Control Stimuli
 A morpheme is the smallest unit of language that still carries meaning