Exploring the nature of morphological regularity: an fMRI study on Russian

ABSTRACT This paper explores the nature of the differences in the processing of morphologically regular and irregular forms in the brain. Verbs cannot be simply divided into regular and irregular in Russian – there are many inflextional classes that differ in defaultness, type frequency, and productivity. In the present functional magnetic resonance imaging study, we chose three verb classes that allow teasing these factors apart and asked 24 subjects to select verb forms agreeing with different pronouns. We combined measures for local brain activity and generalised psychophysiological interactions. We revealed that regularity effects are primarily driven by defaultness associated with more effective and automated processing in the left-lateralised fronto-temporal combinatorial brain network rather than by productivity or type frequency.


Introduction
Although it is well known that the left-lateralised frontotemporal brain system is responsible for processing complex language forms, many basic questions are still open.One of them revolves around the notion of morphological regularity.This notion has been especially important for the debate between so-called dual route (DR) and single route (SR) models of inflection and has been discussed in many other studies.
A number of neurophysiological experiments specifically focused on the distinction between regular and irregular forms.However, as we show below, so far their results have been inconclusive.Moreover, the notion itself allows for different interpretations, which further complicates the picture.
In this paper, we report an fMRI experiment on the Russian language that sheds new light on this ongoing debate.Russian verbs form a complex system of inflextional classes that allows teasing apart several characteristics associated with morphological regularity in different models, including defaultness, productivity and inflextional class, or type frequency.The paper has the following structure.Section 1.1 briefly presents different approaches to morphological regularity and neuroimaging experiments focusing on this notion.Section 1.2 covers the relevant aspects of Russian verb morphology and a previous fMRI study (Kireev et al., 2015;Slioussar et al., 2014) dedicated to it.Section 1.3 introduces the present study.
The present study replicated some of the key results from (Kireev et al., 2015;Slioussar et al., 2014) using a different method: selecting a verb form that matches a previously presented pronoun rather than producing forms.Like (Kireev et al., 2015;Slioussar et al., 2014) and unlike the majority of other studies, we combined different methods to analyse fMRI data: not only the classic subtractive analysis revealing the localisation and the direction of the change in the levels of functional activity, but also the analysis of psychophysiological interactions (PPI) revealing context-dependent changes in distant connectivity among brain regions.These methods were instrumental to detect different effects and to choose between alternative interpretations.
We revealed a morphological regularity effect associated with the inflextional class defaultness.As we show in section 1.1, the property of defaultness is introduced only in DR modelsother approaches rely on different characteristics like type frequency and productivity.Our study demonstrates that the notion of defaultness is psycholinguistically relevant, sheds new light on the nature of morphological regularity and provides a novel piece of evidence in favour of DR models.

Different approaches to morphological regularity and previous fMRI studies focusing on this notion
For experimental studies of inflextional morphology, DR and SR models played an important role.The DR approach assumes that wherever there are multiple ways to derive a particular form (a past tense form, a plural etc.), there is one inflextional class in which forms are produced and processed using symbolic rules and the other classes whose processing relies on associative memory (e.g.Clahsen, 1999;Marslen-Wilson & Tyler, 1997;Orsolini & Marslen-Wilson, 1997;Pinker, 1991Pinker, , 1999;;Pinker & Prince, 1988;Pinker & Ullman, 2002;Ullman, 2004).The former class is expected to show the properties of the morphological default: to have the lowest number of phonological restrictions and to be most readily generalisable.The latter classes may be productive (i.e. the relevant inflextional pattern may be applied to novel words) or highly frequent, but only the former is called regular.Thus, under this approach categorical distinctions are expected between regular and irregular form processing, and regularity is associated with defaultness, rather than with productivity and type frequency (although they may also play a secondary role for irregular form processing).
According to the SR approach, all forms are computed by a single system that contains no symbolic rules (e.g.MacWhinney & Leinbach, 1991;McClelland & Patterson, 2002;Plunkett & Marchman, 1993;Ragnasdóttir et al., 1999;Rumelhart & McClelland, 1986).In this case, some effects of morphological regularity can also be expected, but they are predicted to be gradual and to correlate with such properties as type frequency and productivity.In some other models, e.g. in Yang's (2002) model relying on multiple rules of different status, all properties mentioned above are predicted to be relevant.
Behavioural studies focusing on the problem of morphological regularity have analysed a variety of languages using different methods.However, in most cases, their results could be interpreted both within the DR and the SR approaches.Many authors have placed high hopes on neuroimaging studies that could reveal distinctive patterns of neuronal activity associated with regular and irregular form processing.But they have faced a number of problems.
Secondly, some studies had similar results, while in the other cases, the opposite patterns were observed.Moreover, the interpretation of these datawhat could cause the differences and, crucially, also how to explain similar findingsremained a matter of controversy.Many neuroimaging studies arguing for the DR approach, as well as a number of studies that do not focus on the regular vs. irregular distinction (e.g.Bozic et al., 2010Bozic et al., , 2013;;Marslen-Wilson and Tyler, 2007;Szlachta et al., 2012), assumed that rule-based processing is supported by the fronto-temporal network of the left hemisphere, particularly by Broca's area.However, only three fMRI studiestwo production experiments (Dhond et al., 2003;Oh et al., 2011) and one same-different judgment task (Tyler et al., 2005) found more activation in Broca's area for regular forms.
In the other fMRI studies comparing regular vs. irregular forms, Broca's area was activated more by irregulars (Beretta et al., 2003;de Diego-Balaguer et al., 2006;Desai et al., 2006;Sahin et al., 2006).Some authors argued that this can be taken to support the DR approach and can be explained by conflict monitoring between the regular rule and irregular form or by inhibition of regular rule application (e.g.Sahin et al., 2006).The others assumed that this finding is more in line with the SR approach and reflects greater processing load posed by irregulars, which rely on less frequent inflection patterns than regular forms and therefore have greater attentional and response selection demands (Desai et al., 2006).
Another perspective on the notion of morphological regularity can be found in many descriptive grammars of highly inflected languages (e.g.Serianni, 1988;Zaliznyak, 1977) and was adopted in a recent fMRI study by De Martino et al. (2020).De Martino et al. analysed the inflection of Italian verbs.They can be divided into three main inflextional classes: the 1st, 2nd and 3rd, with infinitives ending in -are, -ere or -ire, respectively (-re is the suffix of the infinitive, and -a-, -e-and -i-are the socalled thematic vowels that are inserted between the stem and different inflextional suffixes).All classes are relatively large, but the 1st class is much more frequent than the other two: according to Thornton et al. (1997), it contains 71% of all verbs.The 1st and 3rd classes are productive.
However, rather than labelling the 1st class or the two productive classes regular, De Martino et al. use this term for the verbs that are inflected like the majority of words in their inflextional class.For example, the 1st class is very homogeneous: it contains only four verbs that have some idiosyncrasies in their inflextional paradigms.The opposite is true for the 2nd class, while the 3rd class is in the middle.De Martino et al. conclude that all classes contain some regular verbs (for which all forms can be fully predicted based on their inflextional class) and some irregular verbs, although their distribution is different.In their fMRI study, they aim to tease apart the role of the inflextional class and regularity in this particular sense. 1  As we show in section 1.2, Russian also has many inflextional verb classes, so De Martino et al.'s (2020) study is very relevant for our work.We will compare our findings to their results in the discussion section.So far, let us only note that inflextional classes and regularity in De Martino et al.'s sense can be regarded as morphological rules of different status.On a larger scale, an Italian speaker must know which thematic vowel is combined with a given verb stem.On a smaller scale, the speaker must also know which suffix to use in a particular form (for example, most past participles have -to suffix, but other suffixes are also possible), whether there are any alternations in the stem etc.

Russian verb inflection and previous studies dedicated to it
There are several approaches to dividing Russian verbs into inflextional classes.According to the one developed in Jakobson (1948), Townsend (1975) and Davidson et al. (1996), Russian has eleven verb classes and several socalled anomalous verbs.Ten classes are identified by their suffixes, while the eleventh class has a zero suffix, and is subdivided into subclasses depending on the quality of the root-final consonant (Jakobson and Townsend counted them as 13 separate classes).
All verbs have two stems: the present/future tense stem and the past tense stem (also used in the infinitive form).Depending on the class, the correlation between them may include truncations or additions of the final consonant or vowel, stress shifts, suffix alternations, alternations of stem vowels and stem-final consonants.The verb class also determines which set of endings is used in the present and future tense (1st and 2nd conjugation types).
In general, there is no single highly frequent productive pattern that can be applied to any stem irrespective of its phonological characteristics.Five out of eleven verb classes are productive, but differ in type frequency.According to Slioussar et al. (2014), the Grammatical Dictionary of the Russian Language (Zaliznyak, 1977)  Behavioural studies testing SR and DR approaches on Russian looked at adult native speakers, L1 and L2 learners and subjects with various neurological and developmental deficits (e.g.Chernigovskaya et al., 2007;Gor, 2003Gor, , 2010;;Gor et al., 2009;Gor & Chernigovskaya, 2001, 2003, 2005;Gor & Jackson, 2013;Svistunova, 2008).Infinitives or past tense forms from real and nonce verbs were given to the participants who were asked to generate present tense forms.Adults often applied the most frequent productive AJ class pattern to nonce verbs irrespective of their morphological properties. 3Other highly frequent productive patterns were not overgeneralised in a similar way.The authors concluded that the AJ class pattern has the psycholinguistic status of the default in Russian.
However, while this finding can be taken to support the DR approach, other results could not be predicted by it.For example, children were found to overgeneralise several inflextional patterns in the course of acquisition.As a result, the group of authors working on Russian concluded that Yang's (2002) model relying on multiple rules might be better suited to account for the data, and Gor (2003) developed a similar model for Russian.
The role of regularity was also tested in an fMRI study (Kireev et al., 2015;Slioussar et al., 2014) that we will review in more detail.The materials included verbs from the most frequent, productive, default AJ class and from the least frequent non-productive classes (for the sake of convenience, we will further call them irregular).The authors reasoned that if any differences between these two groups were found, these classes could be compared to other verbs in subsequent studies.Unlike in most other fMRI studies, the materials also included nonce verbs modelled after real stimuli, as well as real and nonce nouns.Participants were asked to silently read stimuli (verbs in the infinitive form and nouns in the nominative singular) and to produce aloud the 1st person singular present tense form or the nominative plural form.
The results of the subtractive analysis were reported in (Slioussar et al., 2014).They found that functional activity within the fronto-parietal network was influenced by regularity and lexicality factors: it was greater for irregular verbs than for AJ class verbs and for nonce verbs than for real ones.Moreover, Slioussar et al. demonstrated that the effects of regularity and lexicality were very similar.In other words, the pattern they observed for real stimuli was similar to the one found in the majority of other fMRI studies focusing on morphological regularity: Broca's area was activated more by irregulars.However, because Slioussar et al.'s study also included nonce stimuli, they had additional arguments to choose between alternative explanations and concluded that the effect was due to processing difficulty, as Desai et al. (2006) suggested, and not to (ir)regularity as such.Kireev et al. (2015) conducted a ROIwhole brain voxel-wise analysis of context dependent changes in functional connectivity (a PPI analysis) of the same data.Firstly, they found that functional connectivity between the left inferior frontal gyrus (LIFG) and bilaterally distributed clusters in the superior temporal gyri (STG) was significantly greater in AJ class real verb trials than in irregular ones.Secondly, they observed a significant positive covariance between the number of mistakes in irregular real verb trials and the increase in functional connectivity between LIFG and the right anterior cingulate cortex (ACC) in these trails as compared to AJ class ones.Kireev et al. concluded that they could pinpoint a genuine regularity effect and dissociate it from processing difficulty effects.

The present study
In a previous fMRI study of Russian verb morphology (Kireev et al., 2015;Slioussar et al., 2014), the PPI analysis revealed a morphological regularity effect.However, since only two verb classes were used in the experiment, it was impossible to determine whether this effect could be associated with inflextional class frequency, productivity or defaultness.The goal of the present study is to find out whether this effect can be replicated and to explore its nature, because Russian allows teasing these factors apart.
To do so, we conducted an fMRI experiment.Our materials included the two verb classes used in (Kireev et al., 2015;Slioussar et al., 2014): the most frequent productive default AJ class and the least frequent non-productive non-default irregular verbs.We added I class to them: it is the second most frequent one after AJ class, highly productive, but not default.
We also reasoned that a genuine regularity effect should manifest itself in different experimental tasks, both in production and in comprehension.To test this prediction, we used a novel task in our study.Instead of producing verb forms, our participants were asked to read two forms of a previously presented verb and to select the one agreeing with a previously presented pronoun.More details justifying the choice of this task are given in the procedure section.

Method
Participants Twenty-four healthy volunteers (14 women and 10 men) participated in the study.All participants were native Russian speakers, 21-40 years of age, with no history of neurological or psychological disorders.All participants were right-handed, as assessed by the Edinburgh Handedness Inventory (Oldfield, 1971).They were given no information about the specific purpose of the study.All participants provided written informed consent prior to the study, and were paid for their participation.All procedures were conducted in accordance with the Declaration of Helsinki and were approved by the Ethics Committee of the N.P. Bechtereva Institute of the Human Brain, Russian Academy of Sciences.

Materials
Materials included six groups of real and nonce verbs illustrated in Table 1 (a complete list is given in Appendix).Every group contained 35 stimuli.Real verbs in the first group belonged to the AJ classthe most frequent, productive and default (AjV).The second group included verbs from the I classthe second most frequent and productive, but not default (IV).The third group contained verbs from several small non-productive classes (irregular verbs, IrrV).Only unprefixed imperfective verbs were used.Three matching groups of nonce verbs (AjNV, INV and IrrNV) mimicked the general characteristics of the corresponding real verb groups (length and phonological properties of the stem).However, we tried to avoid close resemblance to particular real words, because this would be an additional factor that is hard to control for.
AJ class verbs, irregular verbs and the corresponding nonce stimuli were the same as in (Kireev et al., 2015;Slioussar et al., 2014).Real and nonce verbs from the I class were added to them to tease apart the role of defaultness and productivity in the regularity effect observed in (Kireev et al., 2015).Token frequency was balanced for all real stimulus groups using The Frequency Dictionary of the Modern Russian Language (Lyashevskaya & Sharoff, 2009).Stimuli in all groups were matched for length (see Appendix).Unlike in (Kireev et al., 2015;Slioussar et al., 2014), we did not use nouns because the number of verb stimuli increased, and Russian nouns are not particularly useful to study morphological regularity.

Procedure
Our aim was to analyse word form processing.The most widespread experimental design to study isolated word processing is the lexical decision task.However, this design might be disadvantageous for the purposes of the present study.We focus on the processing of inflections, which encode the grammatical features of a word form.Normally, these features are necessary to establish its relations with other words in the sentence and contribute to the interpretation by themselves: for example, a Russian verb form may encode tense, aspect, mood, as well as the person and number of the subject it agrees with.But this is not particularly relevant for an isolated word form.Therefore, the processing of these features might be more superficial than is usually the case.
Notably, the only fMRI study in which regular vs. irregular verbs were compared in comprehension (Stamatakis et al., 2005;Tyler et al., 2005) used a special task that focused participants' attention on inflextional morphology.Participants listened to pairs of stimuli like jumpedjump, thoughtthink, jadejay and had to decide whether they were forms of the same word or two different words.Due to morphological differences between the two languages, this task cannot be adapted for Russian, and we came up with a different design.
In every trial, participants first saw a real or a nonce verb in the infinitive form and a pronoun ja "I" or on "he" below it (see Figure 1).They were presented for 600 ms in the middle of the screen mounted inside the magnet.Then a fixation cross appeared; the duration of this event varied equiprobably from 1700 to 2100 ms with the step of 100 ms.After that, two present tense forms of the previously presented verb appeared on the left and on the right of the screen for 1500 ms.One of the forms agreed with the previously presented pronoun, i.e. had the same number and person features, the other form did not.Participants were instructed to select the agreeing form as soon as possible by pressing a button on the MR-compatible response box.After that, a fixation cross was presented for a varying equiprobable interval of 1800-2200 ms, with the step of 100 ms, and the next trial started.
We preferred this design to showing only one present tense form (agreeing or not) because in the latter case, half of the trials would involve a morphosyntactic violation, and in general the task would focus on error detection.This would not be optimal for the present study.Below, the combination of the infinitive and the pronoun is referred to as "the first stimulus', and the pair of present tense forms participants had to choose from as "the second stimulus'.We regarded the first stimulus as auxiliary: it was introduced to avoid showing all three forms and a pronoun at once, and focused on analysis of fMRI responses associated with the second stimulus.In total, there were 210 experimental trials, which were pseudo-randomly intermixed with 105 fixation crosses presented for 6 s, used as "null-events" to increase the power of the statistical analysis of fMRI data.The experiment was divided into three consecutive runs with 2-5 min rest between them and was preceded by a short practice run.Stimulus delivery, recording of participants' responses and synchronisation with fMRI data acquisition were carried out via the MR-compatible visual system (NNL, Nordic Neuro Lab, Bergen, Norway) and the E-Prime software (version 2.0, Psychology Software Tools Inc., Pittsburgh, PA, USA).

Data preprocessing and statistical analysis
Image preprocessing and statistical analyses of the fMRI data were performed using the Statistical Parametric Mapping (SPM) 12 software (http://www.fil.ion.ucl.ac.Beta estimates obtained in the t-contrasts "condition > baseline (B)" were submitted to the group level analysis using the random effect repeated measure ANOVAs (the flexible factorial option in the SPM12).We started by running a RM ANOVA GLM model including the main effects of lexicality and inflextional class and their interaction, like in Slioussar et al. ( 2014), and went on to assess the effects of productivity and defaultnessthis is discussed in the results section.
Figure 1.The structure of a probe.Real verbs or nonce verbs in the infinitive form were presented with the pronoun ja "I" or on "he" as the first stimulus.Two present tense forms of the same real or nonce verb were presented as the second stimulus, and participants were asked to choose the form agreeing with the previously presented pronoun.
To control the rate of false positives, the voxellevel FWE (p < 0.05) was used.If the former appeared stringent, the cluster-level FWE (p < 0.05) was applied after voxel-wise uncorrected thresholding (p < 0.001).The anatomical location of the revealed significant BOLD changes was identified by the xjView toolbox (http://www.alivelearn.net/xjview).To evaluate the differences in the BOLD signal in the revealed clusters the REX toolbox (http://www.nitrc.org/projects/rex/)was used.

PPI analysis
In order to assess task-dependent changes in functional connectivity, we used the generalised psychophysiological interactions (gPPI) toolbox (http:// www.nitrc.org/projects/gppi,McLaren et al., 2012).Regions of interest (ROIs) were created by centring a 4 mm radius sphere on the local maxima revealed during the previous stage of analysis testing the main effect of inflextional class.The analysis of psychophysiological interactions was performed for each of the selected ROIs and the remaining voxels of the brain.
Multiple regression GLM models for the second stimulus of the probe included the following regressors: (1) PPI-predictors representing the interaction between physiological activity in the ROI and psychological regressors (corresponding to the second stimulus in the six experimental conditions); (2) the time series of BOLD signal changes within the ROI; (3) regressors that modelled BOLD signal changes induced by the second stimulus in the six experimental conditions and the trials in which participants failed to choose the correct answer; (4) head motion parameters revealed at the realignment stage.T-contrasts were calculated for every subject to estimate condition-specific connectivity versus baseline for both models, and resulting beta coefficients were submitted to the group level analysis.Random effect tcontrasts were calculated for the contrasts of interest associated with the second stimulus.
To identify clusters with significant changes in the PPI parameters we used the uncorrected p < 0.001 threshold at the voxel level with a subsequent FWE correction for multiple comparisons at the cluster level (p < 0.05).The anatomical locations of the observed changes in functional integration were identified using the xjView toolbox (http://www.alivelearn.net/xjview).The REX toolbox (http://www.nitrc.org/projects/rex/)was used to demonstrate the difference between beta values in the identified clusters of functional interaction changes.

Behavioural results
There were very few cases in which participants made an error or failed to respond altogether.Only one participant made 67 errors, and his data were excluded from the subsequent analysis of behavioural and fMRI data.All other participants made 19 errors at most (9.0%from all responses), at most 6 per condition (17.1% from all responses).All trials with errors were discarded from the subsequent analysis.
Behavioural results are presented in Table 2.There were no significant differences in the distribution of errors across conditions.Reaction time analysis revealed only the main effect of lexicality (F(1,132) = 24.2p < 0.001): selecting between nonce verb forms takes more time (716 ± 181 ms (mean ± SD)) than selecting between real verb forms (634 ± 159 ms (mean ± SD)).The inflextional class factor and the interaction between the analysed factors were not significant.

fMRI results: BOLD signal changes
Firstly, we assessed the main effects of lexicality (real verbs were compared to nonce verbs) and inflextional class (a three-level factor), as well as their interaction, like Slioussar et al. ( 2014) did in their study.To examine the role of lexicality, we calculated the following F-contrast: (AjV > B + IV > B + IrrV > B) vs. (AjNV > B + INV > B + IrrNV > B).The main effect of this factor was associated with relatively greater levels of the BOLD signal for nonce verbs as opposed to real verbs.This effect was observed mainly in the left inferior and superior frontal gyri (IFG/SFG) adjacent to the precentral gyrus and insula, in the supplementary motor area and the inferior parietal lobule (see Table 3).There were no clusters with relatively greater levels of the BOLD signal associated with real verbs as opposed to nonce verbs.
The following F-contrast was calculated to explore the main effect of the inflextional class: (AjV and AjNV) vs. (IV and INV) vs. (IrrV and IrrNV).Significant changes were found in the left and right IFG and SFG.We will discuss the nature of these changes below relying on pairwise comparisons between inflextional classes and a parametric contrast.No significant voxels were revealed for the interaction between lexicality and inflextional class factors.
In general, all observed effects were similar to the ones reported in Slioussar et al. ( 2014), both in terms of localisation and the direction of BOLD signal change.Slioussar et al. noted that lexicality and inflextional class effects largely overlapped and concluded that they should be explained by processing difficultyotherwise lexicality and the system of inflextional classes do not have anything in common.Using a parametric contrast, Slioussar et al. proved that BOLD signal increases from real to nonce verbs and from AJ class to irregular verbs.
We calculated a similar parametric F-contrast, only adding I class verbs: It showed that the BOLD signal increased linearly from real AJ verbs to nonce irregular verbs (see Table 3 and Figure 2A).The effect was localised in the left IFG, the supplementary motor area, the middle temporal gyrus (MTG), the inferior parietal lobule and the right middle frontal gyrus (MFG), the angular gyrus and cerebellum.Since we deal with a linear increase from the most frequent AJ class to I and then to the least frequent irregular verbs, we can conclude that processing difficulty depends on inflextional class frequency: verbs using more frequent morphological patterns are easier to process (the same is also true for real verbs as compared to nonce verbs).
Our next goal was to assess the role of productivity and defaultnessthe factors that could not be teased apart by Slioussar et al. (2014) because their study included only two verb classes.The F-contrast (AjV > B + IV > B) vs. (IrrV > B) was estimated for analysis of productivity.The voxel-wise threshold appeared very stringent for this comparison, and significant BOLD signal changes were revealed only with the cluster-level FWE correction.As a result, the comparison between the two productive classes (I and Aj) and the non-productive irregular verbs revealed relatively lower levels of the BOLD signal for the latter in three areas within the left and right IFG and SFG gyri (BA 44/45) (see Table 4 and Figure 2B).
The effect of defaultness was assessed via the F-contrast (AjV > B) vs. (IV > B + IrrV > B).It revealed clusters with decreased levels of the BOLD signal for default AJ class verbs as compared to both I class and irregular verbs.These clusters were located in the left IFG (BA 44/45) and the left MTG/STG areas (Table 4, Figure 2B).Findings from both F-contrasts were largely located within the cluster revealed in the parametric F-contrast.Since productive and non-productive, default and nondefault classes also differ by frequency, we can conclude that frequency and lexicality are two major factors affecting BOLD signal levels.

fMRI results: PPI analysis
The PPI analysis was performed for the two ROIs with a 4 mm radius located in the left IFG pars triangularis (x = −39, y = 29, z = −1), associated with BOLD effects of  both productivity and defaultness, and in the middle temporal gyrus (x = −45, y = 17, z = 20), associated with the BOLD effect of defaultness (see Table 4).In (Slioussar et al., 2014), this area was similarly more activated by irregular verb production as opposed to the production of AJ class verbs.
In (Kireev et al., 2015) and in an earlier PPI study by Stamatakis et al. (2005), significant PPI effects were found only for real stimuli, so we also focused on connectivity analysis for real verb conditions.Other comparisons were also calculated: PPI-parameters for all real and nonce verb trials were analysed using RM ANOVA with two factors: lexicality (real vs. nonce) and inflextional class ((AjV and AjNV) vs. (IV and INV) vs. (IrrV and IrrNV)).However, this analysis not yield significant results, like in the previous studies.
Analysing real verbs, firstly, we assessed the effect of class frequency via the parametric F-contrast (AjV > B) < (IV > B) < (IrrV > B) for the LIFG ROI.Significant changes in psychophysiological interactions were revealed in the left middle cingulate gyrus, left fusiform gyrus and right fusiform / parahippocampal gyrus (Table 5).However, as the estimation of effect sizes suggests (see Figure 3A), these results are primarily due to the difference between the default AJ class and two other classes, while the differences between I class and irregular verbs are very minor.Below, we confirm this observation by reporting significant defaultness effects, but no significant productivity effects.Let us also note that a subsidiary analysis demonstrated greater PPI increases associated with processing of AJ verbs as compared to both I and irregular verbs.The PPI increment is in the opposite direction of BOLD signal changes.The same is true for the results reported below.Thus, we reproduced the BOLD-PPI relationship that was found in the overt production studies by Slioussar et al. ( 2014) and Kireev et al. (2015).
To assess the effect of defaultness, we calculated the F contrast (AjV > B) vs. (IV > B + IrrV > B), which revealed clusters located in the left superior/middle frontal gyrus, left inferior frontal gyrus and insula, left parahippocampal gyrus and putamen, left fusiform gyrus, as well as right fusiform gyrus and parahippocampal gyrus, right cuneus and posterior cingulate, left and right cerebellum (Table 5, Figure 3A).As Figure 3 shows, psychophysiological interactions with the left IFG ROI were increased for AJ class verbs as compared to I class and irregular verbs.Moreover, parametric and defaultness effects substantially overlap.The F contrast   (AjV > B + IV > B) vs. (IrrV > B) associated with productivity did not yield significant results.Secondly, PPI changes for the ROI in the left MTG were revealed by the parametric F-contrast (AjV > B) < (IV > B) < (IrrV > B) in the superior and inferior frontal gyrus, inferior parietal lobule and supramarginal/angular gyrus and middle cingulate at the right brain hemisphere (Table 5, Figure 3B).The F-contrast (AjV > B) vs. (IV > B) + (IrrV > B) calculated to assess the role of defaultness yielded significant PPI increases for default AJ class verbs in the right precuneus and left middle cingulate gyrus (Table 5, Figure 3B).There were no significant PPI changes associated with productivity.In total, we can conclude that defaultness is the primary factor affecting task-dependent changes in functional connectivity.

Discussion
Summarising the results of our study, firstly, we were successful in replicating the pattern that was observed in the previous production study (Kireev et al., 2015;Slioussar et al., 2014).Secondly, we explored the nature of the morphological regularity effect identified in the PPI analysis by Kireev et al. (2015), and demonstrated that it was associated with inflextional class defaultness rather than with productivity or type frequency.We also showed that a more general processing difficulty effect observed in the BOLD signal analysis by Slioussar et al. (2014) most probably correlates with inflextional class frequency.
Let us start with the first finding.Both in production and in comprehension, the connectivity of the LIFG was strengthened in the default AJ class condition, whereas its local activity reflected in the level of the BOLD signal decreased.Looking for an explanation of this pattern, Kireev et al. (2015) took into account the results from (Stamatakis et al., 2005) the only other published study of inflextional morphology in which the PPI analysis was used.
This study was very different from (Kireev et al., 2015;Slioussar et al., 2014) and from ours in terms of materials and method.The connectivity between functionally predefined ROIs was assessed during the same/different judgment task.Stimuli were orally presented as pairs of English words and nonce words, e.g.jumpedjump, thoughtthink, jadejay.But despite all these differences, the crucial PPI finding reported by Stamatakis et al. (2005) was very similar to the results obtained by Kireev et al. (2015) and in the current study: the increased connectivity of the LIFG associated with morphological regularity.
No major model addressing the problem of morphological regularity defines this notion differently for production and comprehension.So, a genuine regularity effect is expected to be observed in both modalitieswhich was indeed the case in the three studies.The fact that this effect was replicated in two languages with relatively poor and relatively rich inflextional morphology and in the experiments using three different tasks proves that it is very robust and reliable.
There were some interesting differences in the results of the three studies.In particular, the Russian production study and the English study observed that functional connectivity between the LIFG and the anterior cingulate cortex (ACC) depended on regularity.In both studies, it was concluded that the ACC plays a monitoring role, taking part in response selection.Nothing similar was observed in the present study, presumably due to the fact that choosing between two inflected forms was a relatively easy task.
Let us turn to the analysis of the local activity levels.The same pattern was observed in Slioussar et al. ( 2014) and in the present study, while Tyler et al. (2005) reported the opposite pattern.The fronto-temporal language-related areas were activated more for irregulars in the latter study, but not in the former two.Since local activity and connectivity results went in the same direction in the English study, the authors explained both of them by morphological regularity.However, since the dissociation between the two revealed by Kireev et al. (2015) was replicated in the present study, this is not an accidental finding.
We conclude that the patterns of local activity and functional connectivity should be explained independently.Therefore, while we agree with the authors of the English study that the observed connectivity changes are associated with morphological regularity, we suggest that the differences in the local activity levels are triggered by processing difficulty.This is a very wide term, so let us indicate the specific interpretation our data point to and hypothesise how it can be used to explain the differences between the Russian and English studies.In English, regular past tense forms are morphologically complex, i.e. consist of a stem and a suffix, while most irregular forms are morphologically simple.Thus, regular verbs are expected to be more difficult in a particular way (their processing involves an extra operation of morphological decomposition).In Russian, all verb forms are morphologically complex, but we assume that if an inflextional pattern is highly frequent and productive, it is processed less effortfully and more automatically.This is why I class verbs are expected to be more difficult than AJ class verbs, but less difficult than irregular verbs.All these expectations are confirmed in the English study and the two Russian experiments.Now let us focus on what we consider to be the main finding of the present study.When the morphological regularity effect was first identified by Kireev et al. (2015), little could be concluded about its nature.The production experiment tested only two groups of verbs: from the highly frequent productive default AJ class and from several small non-productive nondefault classes (so called irregular verbs).Other verb classes that are productive, but not default, that are less frequent than the AJ class, but more frequent than irregular verbs were not included.Thus, the observed effect could be gradual and depended on type frequency; if it were categorical instead, it could be associated with productivity or with defaultness.
The main goal of the present study was to determine the nature of the regularity effect.We added the I class to the comparison and found that it does not trigger any significant connectivity changes although it is highly frequent and productive.These changes were observed only in the comparison between the default AJ class and other verb groups.
This result is far from trivial because prima facie, the special status of the AJ class is not evident In Russian.Several verb classes include thousands of verbs and have recently gained many novel verbs associated with new phenomena.AJ class is the most frequent, but its type frequency does not differ dramatically from the second most frequent I class.The inflextional pattern used in the AJ class has fewer morphophonological restrictions than other patterns, but cannot be applied to any stem.Some evidence that this class behaves as the default one has only been found in the experimental settings: when adult native speakers of Russian were asked to inflect nonce verbs, they overgeneralised this pattern more readily (Gor, 2003;Gor & Chernigovskaya, 2001;Svistunova, 2008).
Notably, only the dual route approach to inflextional morphology postulates a categorical distinction between the default inflextional class and the other classes.Other approaches also predict or at least do not exclude regularity effects in the production and processing of inflextional morphology.But these effects are expected to correlate with type frequency and productivity.The problem is that usually, it is impossible to tease these factors apart.But the system of Russian verb classes gives us a rare opportunity to do so.Summarising our findings, why do default class verbs demonstrate relatively lower activation levels, but greater connectivity?We believe that this may reflect their more effective and automatic processing.
Finally, let us compare our results to those reported by De Martino et al. (2020).They conducted a behavioural experiment and an fMRI experiment in which participants were asked to produce overtly forms from Italian verbs.Stimulus verbs belonged to three inflextional classes.The 1st class is by far the most frequent and highly productive (about 70% verbs belong to it, according to the authors).The 3rd class is the least frequent, but has some limited productivity, while the 2nd class is not productive.In addition to that, some stimulus verbs in the study did not exhibit stem alternations in any forms, while the others did.De Martino et al. termed the first group regular and the second irregular, and used inflextional class and regularity as two factors in the behavioural and fMRI data analysis.
We believe that inflextional classes and regularity in De Martino et al.'s sense can be regarded as morphological rules of different hierarchical status.Firstly, all verbs belong to different inflextional classes, and, based on our definition of regularitycategorial or gradual, relying on defaultness, productivity or type frequencythese classes may be termed (more or less) regular or irregular.This was the level investigated in the previous fMRI studies on different languages and in our study, which showed that type frequency and defaultness are relevant.In Italian, both characteristics clearly set the 1st class apart.Secondly, verbs may show more or less regular behaviour inside their inflextional class.This level had not been addressed before De Martino et al.'s study, and we did not look at it in Russian.Notably, very few 1st class verbs have stem alternations in Italian, so they were not included in De Martino et al.'s study.
If we compare De Martino et al.'s and our findings on the processing of different inflextional classes, we can see that they are largely parallel.De Martino et al. found significant differences only between the default highly frequent 1st class and the other verbs: regions in the left middle frontal gyrus, left pre-supplementary motor area and left anterior cingulate cortex were less activated during the 1st class verb processing.Nondefault infrequent 2nd and 3rd classes did not significantly differ from each other.Similar patterns observed for two different languages in different experimental tasks support the robustness and generalizability of the conclusions made in the two studies.

Conclusions
Based on the results of this study, we can conclude that processing of default class verbs in Russian is associated with a specific processing mode of the left lateralised fronto-temporal brain system responsible for inflextional morphology.We observed greater distant connectivity of the LIFG and MTG together with relatively lower activity in these regions while processing verbs from the default AJ-class, as compared to other verb classes.This not only reproduces previous experimental findings from a verb form production task, but also sheds new light on the nature of morphological regularity.
We demonstrated that inflextional class defaultness is relevant and the most notable processing differences between verb classes are triggered by it, rather than by productivity or type frequency.Of course, the default AJ-class is productive and the most frequent, but the I-class that was also included in the study is productive and highly frequent as well.Nevertheless, we observed significant differences between the AJ-class on the one hand and the I-class and verbs from very small non-productive classes on the other hand.This supports dual route (DR) approaches to morphological processing because only they introduce the concept of defaultness, while single route (SR) approaches rely on different characteristics like type frequency and productivity.Less radical rule-based approaches like Yang's (2002) model relying on multiple rules are also compatible with our findings if we assume that one rule has a special status in the system.Notes 1.In our view, De Martino et al.'s (2020) approach to regularity is definitely valid and, in fact, is more in line with the linguistic tradition, so we only wanted to stress that the term is used differently in many other experimental studies, focusing on the properties of inflectional classes.2. There are several ways to transliterate Russian words from Cyrillic to Latin alphabet.In this paper, we use the so-called scholarly transliteration system.3.For example, participants applied the AJ class pattern to nonce verbs ending in -yt' (similar real verbs belong to two very small unproductive classes) and sometimes even to the ones ending in -it' (only one real verb gnit' 'to rot' has this pattern, the absolute majority of other verbs belong to the I class).

Disclosure statement
No potential conflict of interest was reported by the author(s).

Funding
The current study was supported by the programme of fundamental studies of IHB RAS [grant number 122041500046-5].
uk/spm/software/spm12/). Before the statistical analysis, fMRI data were preprocessed in a standard way: (1) all functional images were spatially realigned to the first functional image via rigid body transformations; (2) slice time correction was applied to compensate differences in the time of acquisition for different slices; (3) individual structural images were segmented and the grey matter mask was used to calculate the parameters necessary to perform the spatial normalisation to a standard stereotactic MNI template (Montreal Neurological Institute); (4) resulting functional images were smoothed using a Gaussian filter, 8 mm full-width at halfmaximum.The statistical analysis was performed in two steps: individual and group level analysis.At the individual level, the general linear model (GLM) included: (1) six regressors of interest modelling the second stimulus presentation for the six experimental conditions (based on the onset time of the stimulus); (2) two regressors modelling the second stimulus presentation for the trials in which participants made an error (i.e.failed to choose the agreeing verb form); (3) six regressors modelling translations and rotations in the x, y and z directions from the first dynamic scan as a nuisance variable.For the regressors of interest we use labels AjV for conditions with a real AJ class verb, IV for a real I class verb, and IrrV for a real irregular verb.Repressors for nonce verbs were denoted by adding N to the corresponding trial labels: AjNV, INV, IrrNV.

Figure 2 .
Figure 2. Changes of the BOLD signal associated with processing of different verb classes.(A) Hot-scale clusters denote BOLD signal changes revealed in the current study in the parametric F-contrast modelling class frequency while selecting the relevant verb form.The green colour denotes clusters revealed in a similar parametric F-contrast in a production study by Slioussar et al. (2014).Plots with effect sizes in the clusters revealed in the current study are presented on the left side.Plots with effect sizes in the clusters revealed by Slioussar et al. (2014) are presented on the right side.(B) Green clusters denote BOLD signal changes associated with the effect of defaultnessa BOLD decrease in the AjV condition as compared to the IV and IrrV conditions.Blue clusters denote BOLD signal changes associated with the effect of productivitya BOLD decrease in the AjV and IV conditions as compared to the IrrV condition.Transparent hot-scale clusters denote BOLD signal changes revealed in the parametric F-contrast, modelling class frequency.Denotations: IFG/MFGinferior and middle frontal gyrus; STG/MTGsuperior and middle temporal gyrus; SMAsupplementary motor area; AjV -AJ class real verbs; IV -I class real verbs; IrrVirregular real verbs; AjNV -AJ class nonce verbs; INV -I class nonce verbs; IrrNVirregular nonce verbs; auarbitrary units.

Figure 3 .
Figure 3. PPI changes associated with processing of the default verb class as opposed to non-default classes.(A) PPI changes revealed for the ROI in the LIFG.(B) PPI-changes revealed for the ROI in the left MTG.Different colours depict revealed effects of PPI changes.The green colour denotes brain areas where the parametric F-contrast was significant.The red colour denotes clusters with significant PPI changes induced by processing verbs from the default AJ-class as compared with two other classes.The yellow colour denotes the overlap between the effect of defaultness and the parametric contrast.Plots depict effect sizes obtained from revealed clusters.
contains 27,409 verbs, and out of them, 11,735 are in the AJ class, 6875 in the I class, 2815 in the OVA class, 1377 in the NU class and 638 in the EJ class.

Table 1 .
The six experimental conditions with examples of stimuli and information on average token frequency (in instances per million) and length (in letters).

Table 2 .
Reaction times (in ms) and percentages of errors in different conditions.

Table 3 .
Analysis of the BOLD signal changes associated with lexicality (real vs. nonce verbs) and inflextional class.

Table 4 .
Analysis of the BOLD signal changes associated with productivity and defaultness.

Table 5 .
Changes in the PPI-parameters associated with the processing of real verbs.