Reassessing inflectional regularity in Modern Greek conjugation Stavros Bompolas Marcello Ferro Claudia Marzi University of Patras, Greece ILC-CNR Pisa, Italy ILC-CNR Pisa, Italy stavros.bompolas@gmail.com marcello.ferro@ilc.cnr.it claudia.marzi@ilc.cnr.it Franco Alberto Cardillo Vito Pirrelli ILC-CNR Pisa, Italy ILC-CNR Pisa, Italy francoalberto.cardilllo@ilc.cnr.it vito.pirrelli@ilc.cnr.it 1 Introduction Abstract Issues of morphological (ir)regularity have tradi- Paradigm-based approaches to word pro- tionally been investigated through the prism of cessing/learning assume that word forms morphological competence, with particular em- are not acquired in isolation, but through phasis on aspects of the internal structure of associative relations linking members of complex words (Bloomfield, 1933; Bloch, 1947; the same word family (e.g. a paradigm, Chomsky and Halle, 1968; Lieber, 1980; Selkirk, or a set of forms filling the same para- 1984; among others). Within this framework, one digm cell). Principles of correlative of the most influential theoretical positions is learning offer a set of dynamic equations that morphologically, phonologically, or/and se- that are key to modelling this complex mantically transparent words are always pro- dynamic at a considerable level of detail. cessed on-line through their constituent elements, We use these dynamic equations to simu- whereas irregular, idiosyncratic (non-transparent) late acquisition of Modern Greek conju- forms are stored (and retrieved) as wholes in the gation, and we compare the results with lexicon (Pinker and Prince, 1994). Likewise, evidence from German and Italian. Simu- Ullman and colleagues (1997) assume that the lations show that different Greek verb past tense formation of regular verbs in English classes are processed and acquired differ- requires on-line application of an affixation rule entially, depending on their degrees of (e.g. walk > walk+ed), while irregular past tense formal transparency and predictability. forms, involving stem allomorphy (e.g. drink > We relate these results to psycholinguis- drank), are retrieved from the lexicon. tic evidence on Modern Greek word pro- Modern Greek introduces an interesting varia- cessing, and interpret our findings as tion in this picture. First, stem allomorphy and supporting a view of the mental lexicon suffixation are not necessarily mutually exclu- as an emergent integrative system. sive processes, but coexist in the same inflected forms (e.g. lin-o ‘I untie’ > e-li-s-a ‘I untied’, Secondo l’approccio paradigmatico allo aɣap(a)-o ‘I love’ > aɣapi-s-a ‘I loved’). Sec- studio dell’elaborazione e dell’appren- ondly, affixation rules may select unpredictable dimento lessicali, le parole di una lingua stem allomorphs: aɣap(a)-o ‘I love’ > aɣapi-s-a non sono acquisite in isolamento, ma at- ‘I loved’, for(a)-o ‘I wear’ > fore-s-a ‘I wore’, traverso legami associativi tra membri xal(a)-o ‘I demolish’ > xala-s-a ‘I demolished’. della stessa famiglia morfologica, la cui These cases suggest that inflectional dinamica è modellata dalle equazioni (ir)regularity is not an all-or-nothing notion in dell’apprendimento correlativo. Il pre- Greek. Different inflectional processes may sente contributo offre una serie di espe- compound in the same words to provide a chal- rimenti nei quali l’apprendimento del si- lenging word processing scenario (Tsapkini et stema verbale del greco moderno è simu- al., 2004). From this perspective, Modern Greek lato come un processo di auto- offers the opportunity to test traditional hypothe- organizzazione dinamica di parole me- ses of grammar and lexicon interaction in word morizzate in modo concorrente. I risultati processing and learning, to explore the potential mostrano chiari effetti di interazione di- of single, distributed mechanisms in addressing namica tra trasparenza e regolarità mor- word processing challenges (Alegre and Gordon, fologica nell’acquisizione di classi di 1999; Baayen, 2007). forme del verbo greco. 1.1 The evidence (i) an affix-based class, requiring the presence of the aspectual marker -s-, and including Modern Greek conjugation is stem-based, each verbs with a predictable phonological stem- fully inflected verb form requiring obligatory allomorph (e.g., lin-o ‘I untie’ ~ e-li-s-a suffixation of person, number and tense markers ‘I untied’, ɣraf-o ‘I write’ ~ e-ɣrap-s-a that attach to either a bare or a complex stem in ‘I wrote’); both regular (aɣap-o ‘I love’ ~ aɣapis-a (ii) a mixed class where active perfective past ‘I loved’) and irregular verbs (pern-o ‘I take’ ~ tense forms are produced by affixation of pir-a ‘I took’). Unlike English speakers, Greek the aspectual marker -s- to a systematic speakers must always resort to an inflectional morphological stem-allomorph (e.g., mil-o process to understand or produce a fully inflected ‘I speak’ ~ mili-s-a ‘I spoke’); form, no matter how regular the form is (Terzi et (iii) an idiosyncratic verb class whose forms are al., 2005: 301). based on non-systematic stem-allomorphy Classifying a Greek verb as either regular or (requiring stem-internal alternation or irregular thus requires observation of the stem suppletion) or no stem-allomorphy at all, formation processes on whose basis agreement and no (sigmatic) aspectual marker (e.g., and tense suffixes are selected. Accordingly, it is pern-o ‘I take’ ~ pir-a ‘I took’, tro-o ‘I eat’ assumed that the presence or absence of the as- – e-fag-a ‘I ate’, krin-o ‘I judge’ – e-krin-a pectual marker is a criterion for assessing the ‘I judged’). degree of regularity of a Greek verb. In particu- lar, so-called “sigmatic” past-tense forms (e.g. It should be noted that, in regular Greek verbs, aɣap-o ~ aɣapis-a) are traditionally considered transparency/systematicity and predictability are to be regular, in that they involve a segmentable not mutually implied. The morphologically- marker (-s-) combined with phonologically pre- conditioned allomorphy of class-(ii) verbs re- dictable or morphologically systematic stem- quires a systematic pattern of perfective stem allomorphs. Asigmatic past-tense forms (e.g. formation, namely X(a) ~ X + V (e.g. aɣap(a)- > pern-o ~ pir-a), in contrast, exhibit typical prop- aɣapi-), where ‘X’ is a variable standing for the erties of irregular inflection, since they involve bare stem, ‘V’ stands for a vowel, and the sub- unsystematic stem allomorphs (in some cases scripted ‘(a)’ indicates an optional ‘a’, forming a suppletive stems), and no segmentable affixes Modern Greek free variant of the imperfective marking perfective aspect. This distinction has stem (e.g. aɣapo ~ aɣapao, see Ralli, 2005, also been supported by psycholinguistic evidence 2007). The variable V in the perfective stem can (Stamouli, 2000; Tsapkini et al., 2001, 2002a,b,c, be instantiated as an i, e or a, and cannot be pre- 2004; Mastropavlou, 2006; Varlokosta et al., dicted from the bare stem. On the other hand, the 2008; Stavrakaki and Clahsen, 2009a,b; Statho- phonologically-conditioned allomorphs of class- poulou and Clahsen, 2010; Stavrakaki et al., (i) verbs (e.g. lin- > e-li-s-) are the outcome of 2012; Konstantinopoulou et al., 2013; among exception-less phonological rules, which none- others), suggesting that sigmatic past-tense forms theless obfuscate a full formal correspondence are typically produced on-line by rules, and (transparency) between the imperfective stem asigmatic forms are stored and accessed from the and the perfective stem. mental lexicon. Evidence from language acquisition and ex- However, careful analysis of the Greek verb perimental psycholinguistics shows that percep- system appears to question such a sharp pro- tion of formal transparency between imperfective cessing-storage divide. In particular, Greek data and perfective Greek stems plays a prominent provide the case of a mixed inflectional system role in human word processing strategies (Tsap- where both stored allomorphy and rule-based kini et al., 2002c: 116, 2004: 616; Stavrakaki and affixation are simultaneously present in the for- Clahsen, 2009a: 117; Stathopoulou and Clahsen, mation of past tense forms. Ralli (2005) provided 2010: 872). More specifically, lack of full formal a classification of verb paradigms which is based nesting between imperfective and perfective on two criteria; firstly, the presence vs. absence stems (compare aɣap-o ‘I love’ ~ aɣapi-s-a of the sigmatic affix and, secondly, the presence ‘I loved’ vs. ðulev-o ‘I work’ ~ ðulep-s-a vs. absence of (systematic) stem allomorphy. As ‘I worked’) appears to have an extra processing a result, we can define the following three clas- cost (Tsapkini et al., 2002c: 116). ses (see also Tsapkini et al., 2001, 2002a,b,c, To sum up, analysis of Greek data offers evi- 2004): dence of graded levels of morphological regulari- ty, based on the interaction between formal words (Marzi et al., 2014, 2016). In particular, transparency (degrees of stem similarity) and high-frequency words tend to recruit specialised (un)predictability of stem allomorphs. The evi- (and stronger) chains of BMUs, while low- dence appears to question a dichotomous view of frequency words are responded to by more storage vs. rule-based processing mechanisms. In “blended” (and weaker) BMU chains. In what fact, no sharp distinction between affix pro- follows, we report how well a TSOM can ac- cessing and allomorph retrieval can account for commodate the complexity of the Greek verb the interaction of formal transparency and pre- system, by controlling factors such as word fre- dictability in Greek word processing. On the one quency distribution, degrees of inflectional regu- hand, rule-based mechanisms are called for to larity and word length. account for transparency effects of stem allo- morphy on word processing. On the other hand, 2.1 The experiment storage is required if the same allomorphs cannot To allow pairwise comparison with existing ex- be predicted. In the remainder of this paper, we perimental evidence on German and Italian test the hypothesis that this evidence is compati- (Marzi et al., 2016), the Greek training dataset ble with a parallel processing architecture (a was designed to contain 50 top-ranked paradigms Temporal Self-organising Map) where pro- by cumulative token frequency, for a total of 750 cessing and storage are in fact mutually implied. verb forms, whose frequency distributions were sampled from the FREQcount section of the 2 TSOMs Greek SUBTLEX-GR corpus (BCBL, 2016). Temporal Self-organising Maps (TSOMs, Ferro From each paradigm, 15 inflected forms were et al., 2011; Marzi and Pirrelli, 2015; Pirrelli et extracted: the full set of present indicative (6) al., 2015) are unsupervised artificial neural net- and simple past tense (6) forms, and the singular works that learn to dynamically memorise input forms of simple future (3). As we were mainly strings as chains of maximally-responding pro- interested in effects of global paradigm-based cessing nodes (Best Matching Units or BMUs), organisation of active voice indicative forms, we whose level of sensitivity to input symbols in excluded paradigms with systematic gaps, imper- specific contexts is a continuous function of the sonal verbs, and deponent verbs. We included distributional regularities of the input symbols high-frequency paradigms with suppletive forms during training. In a TSOM, each processing or/and non-systematic allomorphy (Ralli, 2007, node has two layers of synaptic connectivity: an 2014) as attested in the training set. input layer, connecting the node to the current The dataset was administered to a 42x42 node input stimulus (e.g. the letter of a written word), map for 100 learning epochs. Word frequencies and a (re-entrant) temporal layer, connecting the in the training data were a function of the real node to all other nodes. word frequency distribution in the reference cor- Given the BMU at time t, the temporal layer pus, fitted in the 1-1000 range. To control for encodes the expectation of the current BMU for random variability, we repeated the experiment 5 the node to be activated at time t+1. The strength times. of the connection between consecutively activat- For each repetition, we then assessed how well ed BMUs is trained through the following princi- the map could acquire the 750 input forms, using ples of correlative learning (compatible with the task of Word Recall as a probe. Word recall Rescorla-Wagner (1972) equations): is defined as the process of retrieving a word Given the input bigram ab, the connection form from its chain of BMUs. Successful recall is strength between BMU of a at time t and BMU of possible if inter-node connections on the tem- b at time t+1 will poral layer are finely tuned to the distribution of symbols in the training data. The more accurate the re-entrant temporal coding is, the easier for  increase if a often precedes b in training the map to retrieve the symbols of a word in their (entrenchment) appropriate order. We make the further reasona-  decrease if b is often preceded by a sym- ble assumption that a word is acquired by a bol other than a (competition). TSOM when the map is in a position to recall the word accurately and consistently from its BMU The interaction between entrenchment and com- chain. Average recall accuracy at epoch 100 petition in a TSOM accounts for important dy- turned out to be considerably high: 99.6 % (std = namic effects of self-organisation of stored 0.1%). 3 Data analysis fect, and explains why long regular forms tend to be acquired (on average) more easily than long Results were analysed using Linear Mixed Ef- irregular forms. fects (LME) models with experiment repetitions and training items as random variables. Figure 1 shows the marginal plot of the inter- action between word length and regular vs. ir- regular verb classes for German, Italian and Greek, using an LME model fitting word learn- ing epochs, with (log) word frequency, inflec- tional class and word length as fixed effects. In German and Italian, the distinction between regular and irregular paradigms is based on the criterion of absence vs. presence of stem allo- morphy across all forms of a paradigm (Marzi et al., 2016). In Greek, we consider regular all par- adigms showing a sigmatic perfective stem, and irregular those with an asigmatic perfective stem. Unlike German and Italian (Figure 1, top and middle panels), where irregulars tend to be ac- quired systematically later than length-matched regulars are, and no significant interaction is found, Greek data (Figure 1, bottom panel) show an interesting crossing pattern: shorter irregulars are acquired earlier than length-matched regulars of comparable frequency, but long irregulars are acquired later than long regulars. Marzi and colleagues (2016) account for earli- er learning epochs of both German and Italian regulars as an effect of stem transparency on cu- mulative input frequencies. With German and Italian regular verbs, stems are shown to the map consistently more often, since they are transpar- ently nested in all forms of their own paradigm. This makes their acquisition quicker, due to spe- Figure 1. Marginal plots of interaction effects between word cialised chains of stem-sensitive BMUs getting length and inflectional regularity in an LME model fitting more quickly entrenched. Once a stem is ac- word learning epochs in German (top), Italian (middle) and quired, it can easily be combined with a common Greek (bottom). Solid lines = regulars, dotted lines = irregu- pool of inflectional endings for tense and agree- lars. ment, simulating an effect of (nearly) instantane- To further investigate the impact of degrees of ous (or paradigm-based, as opposed to item- formal transparency on the processing of Greek based) acquisition. In contrast, Greek verb clas- verb forms, we conducted an LME analysis of ses always present stem allomorphy throughout the interaction between word length and classes their paradigms, no matter whether allomorphy is of (ir)regularity in word recall (Figure 2). When systematic, phonologically motivated or unsys- we control for length, regular verbs with system- tematic. In regular verbs, where perfective stem atic morphological allomorphs (e.g. aɣap(a)-o ~ formation requires -s- affixation, perfective aɣapi-s-a, solid line in the plot) are recalled more stems are systematically longer than their imper- easily than regular verbs with phonological allo- fective counterparts, and are acquired after them. morphs (e.g. ðulev-o ~ ðulep-s-a, dashed line in Nonetheless, since imperfective stems are redun- the plot). Notably, both classes are easier to re- dantly embedded in perfective stems, learning a call than asigmatic (irregular) verbs (dotted line long regular perfective form is easier (i.e. it takes in the plot), which show, in most cases, formally a comparatively shorter time) than learning an more opaque allomorphs (e.g. pern-o ~ pir-a). irregular perfective form of comparable length. As shown by the difference in slope between the This is, again, a regularity-by-transparency ef- solid line and the other two lines of Figure 2, facilitation increases with word length, support- presuppose a hardly tenable subdivision of work ing our interpretation of the crossing pattern in between storage and processing. the bottom panel of Figure 1. References Alegre, M. and P. Gordon. 1999. Frequency effects, and the representational status of regular inflec- tions. Journal of Memory, and Language, 40: 41- 61. Baayen, H. R. 2007. Storage, and computation in the mental lexicon. In: G. Jarema and G. Libben (eds.), The Mental Lexicon: Core Perspectives. Amster- dam: Elsevier, 81-104. Figure 2. Marginal plot of interaction effects between length BCBL. 2016. SUBTLEX-GR: The corpus. Donostia: (x axis), and degrees of stem regularity in an LME model Basque Center on Cognition, Brain, and Language. fitting Difficulty of Recall (y axis) by TSOMs trained on [Retrieved from: http://www.bcbl.eu/subtlex-gr/, Greek verb forms 12-04-2016]. Bloch, B. 1947. English verb inflection. Language, 4 Conclusions 23: 399-418. Data analysis highlighted a hierarchy of regulari- Bloomfield, L. 1933. Language. New York: Henry ty-by-transparency effects that appear to have Holt, and Co. consequences on morphological processing. In Chomsky, N. and M. Halle. 1968. The sound pattern particular, the evidence offered here emphasises of English. New York: Harper and Row. the role of formal preservation of the stem (or stem transparency) in the paradigm as a key fa- Ferro, M., Marzi, C. and V. Pirrelli. 2011. A Self- Organizing model of word storage and processing: cilitation factor for morphological processing. implications for morphology learning. Lingue e Our case study is focused on a distinguishing Linguaggio, X(2): 209-226. characteristic of Greek conjugation: all verb par- adigms, both regulars and irregulars, involve Konstantinopoulou, P., Stavrakaki, S., Manouilidou, (unpredictable) stem allomorphy in past-tense C. and D. Zafeiriou. 2013. Past tense in children formation. Hence, the difference between regular with focal brain lesions. Stem-, Spraak- en Taalpa- thologie, 18(1): 90-94. and irregular verbs could not be attributed to the presence or absence of stem allomorphy as is the Lieber, R. 1980. On the organization of the lexicon. case with other languages, such as English and PhD thesis. Cambridge: MIT. Italian (and, to a lesser extent, German), but ra- Marzi, C., Ferro, M. and V. Pirrelli. 2014. Morpho- ther to the type of stem allomorphy itself. The logical structure through lexical parsability. Lingue finding (Figure 2) that stem-final systematic e Linguaggio, XIII(2): 263-290. change, as in the case of regulars, led to Marzi, C. and V. Pirrelli. 2015. A neuro- significantly easier recall than stem-internal computational approach to understanding the Men- vowel changes and non-systematic/non- tal Lexicon. Journal of Cognitive Science, 16(4): predictable stem-final change, as is the case of 493-534. irregulars, lends support to the conclusion that Marzi, C., Ferro, M., Cardillo, F. A. and V. Pirrelli. the type of stem allomorphy is what determines 2016. Effects of frequency, and regularity in an in- the different levels of morphological regularity in tegrative model of word storage, and processing. Greek. Crucially, this seems to involve a regular- Italian Journal of Linguistics, 28(1): 79-114. ity-by-transparency interaction, with predictabil- ity playing second fiddle. Mastropavlou, M. 2006. The effect of phonological saliency, and LF-interpretability in the grammar of Our data meet growing psycholinguistic evi- Greek normally developing, and language im- dence pointing in the same direction, to empha- paired children. Ph.D. thesis. Thessaloniki: Aristo- sise the importance of formal redundancy for tle University of Thessaloniki. speakers’ perception of morphological structure. Furthermore, it paves the way to the definition of Pinker, S. and A. Prince. 1994. Regular, and irregular morphology, and the psychological status of rules a performance-oriented notion of inflectional of grammar. In: S. D. Lima, R. L. Corrigan and G. regularity that may ultimately cut across tradi- K. Iverson (eds.), The reality of linguistic rules. tional competence-based classifications, which Amsterdam: Benjamins, 321-351. Pirrelli, V., Ferro, M. and C. Marzi. 2015. Computa- Amsterdam/Philadephia: John Benjamins, 315- tional complexity of abstractive morphology, In: 335. M. Baerman, D. Brown and G. Corbet (eds.), Un- Tsapkini, K., Jarema, G. and E. Kehayia. 2002c. Reg- derstanding and Measuring Morphological Com- ularity revisited: Evidence from lexical access of plexity. Oxford: Oxford University Press, 141-166. verbs, and nouns in Greek. Brain, and Language, Ralli, A. 2005. Morfologia [Morphology]. Athens: 81: 103-119. Patakis. Tsapkini, K., Jarema, G. and E. Kehayia. 2004. Regu- Ralli, A. 2007. On the role of allomorphy in inflec- larity re-revisited: Modality matters. Brain, and tional morphology: evidence from dialectal varia- Language, 89: 611-616. tion. In: G. Sica (ed.), Open problems in linguis- Ullman, M., Corkin, S., Coppola, M., Hickok, G., tics, and lexicography. Milano: Polimetrica, 123- Growdon, J. H., Koroshetz, W. and S. Pinker. 152. 1997. A neural dissociation within language: evi- Ralli, A. 2014. Suppletion. In: G. K. Giannakis (ed.), dence that the mental dictionary is part of declara- Encyclopedia of Ancient Greek language, and lin- tive memory, and that grammatical rules are pro- guistics, Vol. 3. Leiden, Boston: Brill, 341-344. cessed by the procedural system. Journal of Cogni- tive Neuroscience, 9(2): 266-276. Selkirk, E. 1984. Phonology, and Syntax. The MIT Press. Varlokosta, S., Arhonti, A., Thomaidis, L. and V. Joffe. 2008. Past tense formation in Williams syn- Stamouli, S. 2000. Simfonia, xronos ke opsi stin elin- drome: evidence from Greek. In: A. Gavarro and iki idiki glosiki diataraxi [Agreement, tense, and M. Freitas (eds.), Proceedings of GALA 2007. aspect in specific language impairment in Greek]. Cambridge: Cambridge Scholars Publishing, 483- In: Proceedings of the 8th symposium of the Pan- 491. hellenic Association of Logopedists. Athens: El- linika Grammata. Rescorla R. and A. Wagner. 1972. A theory of Pavlo- vian conditioning: variations in the effectiveness of Stathopoulou, N. and H. Clahsen. 2010. The perfec- reinforcement and non-reinforcement. In: A. H. tive past tense in Greek adolescents with Down Black and W. F. Prokasy (eds.), Classical condi- Syndrome. Clinical Linguistics & Phonetics, tioning ii. New York: Appleton-Century-Crofts, 24(11): 870-882. 64-99. Stavrakaki, S. and H. Clahsen. 2009a. The perfective past tense in Greek child language. Journal of Child Language, 36: 113-142. Stavrakaki, S. and H. Clahsen. 2009b. Inflection in Williams Syndrome. The perfective past tense in Greek. The Mental Lexicon, 4: 215-238. Stavrakaki, S., Koutsandreas, K. and H. Clahsen. 2012. The perfective past tense in Greek children with specific language impairment. Morphology, 22: 143-171. Terzi, A., Papapetropoulos, S. and E. D. Kouvelas. 2005. Past tense formation, and comprehension of passive sentences in Parkinson’s disease: Evidence from Greek. Brain, and Language, 94: 297-303. Tsapkini, K., Jarema, G. and E. Kehayia. 2001. Mani- festations of morphological impairments in Greek aphasia: A case study. Journal of Neurolinguistics, 14: 281-296. Tsapkini, K., Jarema, G. and E. Kehayia. 2002a. A morphological processing deficit in verbs but not in nouns: A case study in a highly inflected lan- guage. Journal of Neurolinguistics, 15: 265-288. Tsapkini, K., Jarema, G. and E. Kehayia. 2002b. The role of verbal morphology in aphasia during lexical access: Evidence from Greek. In: E. Fava (ed.), Clinical linguistics, and phonetics: Theory, and applications in speech pathology, and therapy.