Defective verbs in Portuguese: a morphomic approach

1 Introduction

Inflectional morphology is usually considered to be extremely productive; neologisms and borrowings are supposedly automatically assigned a full set of inflectional forms, just as native lexemes are. Inflectional defectiveness is the phenomenon whereby certain forms of certain lexemes can be missing. Portuguese grammars identify two major types of morphological defectivity in the verbal paradigm, which coincide with patterns of syncretic allomorphy present in a number of lexemes both in Portuguese and across other Romance languages. These patterns have been defined as autonomous morphological structures or morphomes (Aronoff 1994). There is a wealth of diachronic evidence showing how these purely paradigmatic patterns and distributions condition morphological change, indicating their cognitive reality (e.g., patterns of whole word syncretism in Gallo-Romance (Hinzelin 2012) and Russian (Baerman 2004: 810), morphophonological levelling in Ibero-Romance (O’Neill 2014) and patterns of suppletion in Romance in general (Maiden 2004c, 2018)). In addition to this diachronic evidence there is also much cross-linguistic, synchronic evidence for morphomes: Herce (2023) identifies some 110 morphomic structures across a wide variety of the world’s languages and a recent psycholinguistic study on Italian speakers (Cappellaro et al. 2024) provided synchronic experimental evidence for the psychological reality of a particular morphome (L/U-pattern morphome).

Despite such evidence, morphomes have largely been overlooked in Cognitive Linguistics (henceforth CL) due to their lack of reference to meaning: morphomes cannot be reduced to any coherent and convincing semantic feature, they are merely a subset of cells in the paradigm that are particularly cohesive formally and have a greater inter-predictability or diagnostic function with respect to the other cells. Meaning is often seen as central to CL since one of its basic tenets is that language is a system whose primary function is to convey meaning. Morphomes may not be ‘meaningful’ in the narrow sense, i.e., relating to ‘system-external information’, however, they are in the broader sense of them being ‘informative’. Morphomes are extremely informative since they exhibit systematic patterns of form that have a high predictive value within the morphological system, a value that increases in proportion to the size and diversity of the inflected forms of the system. Within such a broader conception of meaning, morphomes have essentially the same predictive value as ‘meaning-driven’ patterns. Moreover, upon the assumption that children are not born with innate concepts, it is feasible to hypothesize that the acquisition of inflectional morphology can, in the first instance, be based on form-form relations only – and in this way L1 acquisition of morphology differs radically from L2 acquisition (Clahsen et al. 2010). Systematic form-form correspondences in the linguistic input with a high predictive function are accordingly extremely informative, and therefore meaningful, for the acquisition and ultimately the organisation of morphology.

CL approaches, particularly within the Usage-based tradition, place a primary focus on communication, rather than on a kind of abstract, disembodied ‘meaning’ and so they are well suited to model morphomes. Also, learning is a central component of/prerequisite for the use of a communication system, and any systematic correlations, whether these be system-external (ostensibly meaningful) or system-internal (morphomic), are of predictive value and thus contribute to a learner’s ability to extrapolate beyond the input that they directly encounter. Moreover, according to Schmid (2016: 543), the three main premises which motivate ‘the cognitive-linguistic enterprise’ are (a) that language interacts with and follows the same principles as other domains of cognition, (b) that grammatical structure derives from usage, and (c) that grammar is emergent and continuously reorganised under the influence of language use. Usage, however, has produced various purely morphological phenomena which make no reference to information outside the morphological system, the classic example being inflectional classes (for other examples and a discussion see Enger 2019). Morphomes are another instance of such purely morphological phenomena, created by recurrent patterns of form across lexemes. Note that pattern recognition is a fundamental aspect of human cognition (Hawkins and Blakeslee 2005; Koch 2004; Neisser 2014; Shepard 1987). In this article, I present further synchronic evidence in favour of morphomes, claiming that they are responsible for patterns of defective verbs in Portuguese. I also suggest that CL should be open to the possibility of the importance of form-form relationships, which are extremely meaningful in the broader sense of them being informative.

2 The Portuguese data

Portuguese grammars identify two major types of defectivity in the verbal paradigm characterised by the verb abolir ‘abolish’ in Table 1 and the verb blandir ‘brandish’ in Table 2. This latter verb supposedly does not possess any inflectional forms for the 1sg.prs.ind, all of the prs.sbj and the imperative forms syncretic with the subjunctive; this type of defectivity will be called L-pattern defectivity. The other type of defectivity, as characterised by the verb abolir, combines this pattern but also includes the 2sg, 3sg, 3pl prs ind and the syncretic imperative forms, this will be referred to as N&L-pattern defectivity. These names, as will be elaborated on, refer to predictive patterns of syncretism (morphomes) present in a number of lexemes across the Romance languages.

Cross-linguistically, defective forms have been attributed to semantic and/or pragmatic, or even phonological reasons (see Sims (2016: 58–63) and Bermel and Brown (2023) for a general overview of morphological defectivity). However, on the basis of comparative evidence from the extremely closely related language, Spanish (Maiden and O’Neill 2010; O’Neill 2009, 2010), it would seem that the defectiveness in Portuguese should be given the same explanation. Spanish defective verbs are strikingly similar to those reported in Portuguese grammars: the lexemes are all of relatively low frequency, they belong overwhelmingly to the -ir class, they display similar or identical patterns of defectivity and, at times, the exact same lexemes are noted as defective in exactly the same way. For example, the Spanish verbs abolir ‘abolish’ and brandir ‘brandish’, have exactly the same patterns of defectivity as the Portuguese verbs above. I therefore focus my discussion on the aforementioned accounts for defectiveness in Spanish and assess these against the Portuguese data. These accounts can be classed as structural; they see synchronic gaps as not accidental but systematic, reflecting the organization of larger morphological patterns. Indeed, the missing forms are intimately related to parallel patterns of allomorphy within the language. These patterns are defined as autonomous morphological structures or morphomes (Aronoff 1994). Models that do not recognize an analogue of morphomic structure are largely compelled to describe these patterns in phonological terms, resulting in high levels of phonological abstraction and a loss of generalization (see O’Neill 2024 §2.2.1.2). As will be discussed, complex abstract phonological factors have also been put forward for defective verbs in Portuguese (Postma 2013), and this explanation has also received validation from an experimental study (Nevins et al. 2014).

The descriptive grammatical treatment of defectivity in Portuguese, however, is characterised by a lack of consensus and contradictions. As shown in Figure 1, there is a great disparity among grammarians as to the number of defective verbs in Portuguese and, as illustrated in Table 3, there is little agreement between scholars over which particular cells of particular lexemes are defective.

Figure 1:

Bar chart of the number of alleged defective verbs in Portuguese according to different grammars. Note that the Cunha and Cintra edition consulted is from (Cunha and Cintra 1984).

Such discrepancy invites one to pose the question as to whether defectivity in Portuguese constitutes a psychological reality for speakers or is just an invention of grammarians, possibly influenced by Spanish grammars. The present article answers this question via a statistical analysis of corpus data in order to validate the putative defective lexemes and patterns listed in the literature. As will be seen, the results invalidate the phonological-based theory of defectivity for Portuguese and support the idea of a common ‘morphomic’ explanation for defective verbs in Spanish and Portuguese.

3 The phonological account of defectivity in Portuguese

Postma (2013) has argued that defective verbs in Portuguese are phonologically conditioned by the appearance of a coronal sonorant after the vocalic root vowel. This phonological explanation of defectivity depends on the assumption, dating back to early versions of generative phonology, that the vocalic alternations in Portuguese -ir verbs (as illustrated in Table 9 further down) all derive from the same underlying root which undergoes different phonological rules (Harris 1974). The general strategy for assigning a phonological analysis to a morphological phenomenon involves reconstituting an environment that conditions a morphologized alternation (often mimicking historical origins) at a non-surface level. This idea is expressed in the treatment of Portuguese mid-open vowels in the 1/3sg.prs.ind and the high-vowels in the 1sg.prs.ind and all of the prs.sbj (e.g., d[ɔ]rme, s[ɛ]rve (3sg.prs.ind) and durma, sirva (1/3sg.prs.sbj) respectively for dormir ‘sleep’ and servir ‘serve’) which are assumed to be the result of a processes of vowel harmony with an underlying thematic vowel. Note that this vowel is deleted in the surface representation (dorm+i+e/a, serv+i+e/a). The different vowel raisings in these forms are attributed to some type of auto-segmental spreading from the deleted theme vowel to the preceding rhizotonic vowel. The claim is that coronal sonorants which occur immediately after this proposed thematic vowel (e.g., in the verb colorir ‘colour’) compete with the theme vowel in the suprasegmental docking and the result is defective forms. In order to test the hypothesis that defectiveness arises through slot competition between a surfacing coronal and a postulated harmony-triggering theme vowel, Nevins et al. (2014) asked speakers to produce the 1pl.prs.ind and 3sg.prs.sbj of the defective forms and to rate their confidence in their own productions. A statistical model was then created to establish, based on the confidence rates, what forms might be classed as defective. The results were that the majority of defective verbs, except the verbs imergir ‘emerge’ and ruir ‘collapse’, contained a coronal consonant, providing independent support for Postma’s (2013) phonologically conditioned account.

There are a number of issues with both studies, however, I will concentrate on the experimental methodology of Nevins et al. (2014). The main problem with this study is that the list of defective verbs which formed the basis of their experiment was extremely reduced. Firstly, these authors, who restricted their attention to Brazilian Portuguese, consulted a single source for the list of possible defective verbs in Portuguese, the 40 verbs listed in Cunha and Cintra (1984) and they only tested for L-pattern defectivity, not L&N pattern defectivity. This is problematic because, as noted, within Portuguese grammars, there is much discrepancy over the actual number of defective verbs and the types of defectiveness they display. Secondly, the authors eliminated many defective verbs from their study. They used the judgements of just two PhD students to vet the list of 40 defective verbs in Cunha and Cintra (2013). Almost half of these verbs were eliminated from the study because the students did not know the meaning of the verbs. Moreover, more verb forms were removed if the participants in the study rated their familiarity with the lexemes as below 2.5 on a 1–5 scale. This removal of data from the experiment is a serious issue, not only methodologically but also theoretically, since, cross-linguistically it has been noted that defectivity seems to occur in lexemes of a very low frequency (Sims 2019). Taken together, these extensive omissions and exclusions reduce the empirical coverage of the study to the point that no conclusions of any generality can be drawn from any results that it purports to obtain.

4 A purely-morphological/morphomic account of defectivity

A subset of cells in a paradigm or class constitute a morphome when they exhibit a high degree of cohesion and inter-predictability that allows them to serve a diagnostic function. A salient characteristic of morphomes is that the recurrent patterns of allomorphy that they display cannot be attributed to common semantic or phonological properties of the cells, but instead reflect a shared morphological structure.

The morphomes relevant to the present discussion are those termed the N-pattern and the L-pattern, both arbitrary^[1] labels coined by Maiden (2004a). These labels denote an alternation within the verbal paradigm whereby an allomorph distinct from the rest of the paradigm is shared by the collection of semantically diverse cells defined by the pattern; those of the L-pattern are 1sg.prs.ind and all of the prs.sbj and syncretic imperative forms, those of the N-pattern are all singular and 3pl forms of the prs.ind and sbj and 2sg.imper. Witness the examples of L-pattern allomorphy in Table 4 for Spanish and in Table 5 for Portuguese, and N-pattern allomorphy in Table 6 for Spanish and Table 7 for Portuguese; the shaded cells correspond to the morphomic cells. For a full discussion of the origins of these patterns, their different instantiations in the different Romance languages and justifications as to their autonomous morphological status, see Maiden (2018) and references therein; see O’Neill (2024) for arguments against the Spanish allomorphy being phonologically conditioned.

It should also be noted that, in Portuguese, the N-pattern interacts with the L-pattern, which effectively dominates it and reduces the N-pattern to the 2sg, 3sg, 3pl and relevant imperative forms, thereby creating a new pattern, which I have termed the Prolific-Portuguese Pattern. This name is due to this pattern being extremely prominent in the Portuguese verb; nearly all^[2] -er and -ir verbs which display an orthographic mid-vowel as the root-vowel exhibit allomorphy which alternates in accordance with this pattern. In -er verbs, as illustrated in Table 8, the L-pattern cells display a high-mid vowel in the root which alternates with an open-mid vowel in the reduced N-pattern cells. In -ir verbs, the root of the reduced N-pattern cells also displays an open-mid vowel, but the vowel in the L-pattern is a high vowel; witness the examples in Table 9. Remember also that a number of lexemes are reported as defective exclusively in the cells of this morphome (see abolir ‘abolish’ in Table 1).

Such morphomes have been used to argue that systematic form-form correspondences should be reflected in any theory of grammar (Aronoff 1994: 25) especially since there is much diachronic evidence to suggest that they are psychologically real for speakers (Maiden 2018) That is, the historical perseverance and resilience of morphomic patterns suggest that they perform a function that speakers seem willing to preserve even when languages undergo significant restructuring. This function is that they are informative or ‘meaningful’ system-internally.

Within this context, Maiden and O’Neill (2010: 121) noted for Spanish that although the cells of the morphomes have, amongst themselves, a high inter-predictive function, the particular allomorphy within these cells is not readily predictive from cells outside the morphome. They proposed that the defectivity in Spanish is due to speakers being ‘paranoid’^[3] about the allomorphy in these particular cells and they therefore avoid the forms even where there are no reasonable grounds to expect allomorphy to occur. This inexact notion of ‘paranoia’ is clarified and substantiated in a later study by O’Neill (2010). It is claimed that due to the high number of -ir verbs that display N-pattern and L-pattern allomorphy, it has become a ‘rule’ for the language that all –ir verbs have a lexicalized root exclusively for these patterns. Essentially, the dominant pattern of frequent verbs in an inflectional class has been analogically extended to all verbs of the class. Defective verbs are therefore verbs which have not had the roots for these verbs committed to memory due to the lexemes being low frequency and the particular idiosyncratic history of the verbs. According to a diachronic corpus study (O’Neill 2009), these verbs are historically those which (i) did have a full paradigm but which were increasingly used in the past-participle form as verbal adjectives (e.g., compungido ‘remorseful’, embutido ‘stuffed’, tullido ‘crippled’) and so their N&L-pattern forms were lost and (ii) borrowings from Latin or French which entered the language in the infinitive or past participle (abolir ‘abolish’, blandir ‘brandish’, precaver ‘guard against’) and which never possessed forms for the N&L patterns and which still today have not been created.

The following sections investigate whether a parallel morphomic analysis can be applied to Portuguese. The first step towards developing a viable analysis of the general phenomenon of defectivity in Portuguese involves a careful evaluation of the competing claims about the number and type of defective cells in Portuguese. By grounding an analysis in a principled enumeration of cases, it is also possible to evaluate more extensively the proposed phonological account for defectivity: the influence of coronal sonorants. In what follows, therefore, I propose a list of defective verbs based on a corpus study.

5 Corpus study and statistical analysis of defective verbs in Portuguese

To establish the defective verbs of Portuguese, I followed a similar methodology used for Spanish (O’Neill 2009). I carried out corpus searches on the list of possible defective verbs and established whether these lexemes could be categorized as defective based on a predictive statistical model. This model was based on the frequency distribution of the forms of 100 non-defective lexemes and compared their distribution of forms with that of the putative defective forms. The statistical model provides a nuanced, gradient analysis of defectivity, in place of an ‘all or nothing’ account. Lexemes are classified as defective if they contain cells that are realized by forms that show a frequency of occurrence that is significantly below the expected frequency. I concentrated exclusively on European Portuguese and used the Centempúblico corpus, with some 180 million words, comprised of texts published in the Portuguese newspaper El Público.^[4]

Firstly, a list was compiled of all the verbs classed as defective in the different Portuguese grammars (Figure 1). Secondly, these lexemes were then subject to corpus searches in which, in addition to the overall frequency of each lexeme, the frequency of the following inflectional forms for these verbs were noted:^[5] (a) 3sg.prs.ind, (b) 1pl.prs.ind, (c) all the prs.sbj, (d) pst.ptcp. The 3sg.prs.ind was used as a diagnostic of the modified N-pattern and the forms of the prs.sbj as a diagnostic for the L-pattern. The forms of the 1pl.prs.ind were noted to check whether it was the case that a verb was simply defective in the present tense, as opposed to it being defective in accordance with the different morphomic patterns. The forms of the pst.ptcp were logged since for Spanish it had been noted (O’Neill 2009) that a number of the supposed defective verbs were not verbs but adjectives whose forms coincided with the pst.ptcp forms of the putative verbs. The overall frequency of the purported defective verbs was noted in order to make testable predictions based on a mathematical model extracted from corpus data of verbs which were not classed as defective.

For this model, a sample of 100 verbs was taken from the 1,000 most frequent verbal lexemes according to the same corpus; 50 from the top end of the list and 50 from the bottom end. For these lexemes, in addition to their overall frequency in the corpus, the frequency for the following forms was noted: (a) 3sg.prs.ind, (b) 1pl.prs.ind and (c) all the prs.sbj. ^[6] The numerical values for these forms were then converted into logarithmic values and a number of linear regressions were calculated using the software R (R Core Team 2021). The result was a statistical model which, from the overall frequency of a non-defective Portuguese verbal lexeme, could predict not only the expected values for (a)–(c) above but also a maximum and minimum value for each form according to the natural variance of the model.

This statistical model was subsequently applied to the attested overall frequency values of the purported defective verbs in order to generate a range of values for the forms (a)–(c) above. These values were then compared with the actual attested values of these forms in the corpus. If the frequency of a particular verb form fell below the minimum value predicted by the statistical model, then this form was classed as defective, converting raw patterns of continuous variation into a discrete classification of lexemes and cells. To illustrate how this process worked, compare in Table 10 the values generated by the mathematical model for two supposed defective verbs: abolir ‘abolish’ and punir ‘punish’.

From this table, it can be concluded that the verb abolir is defective in 3sg.prs.ind and all the prs.sbj forms, but not in the 1pl.prs.ind. This verb was therefore classed as defective according to the N&L-pattern. Contrastively, the verb punir was not classed as defective since its attested forms in the corpus fell within the predicted maximum and minimum values for this verb, given its overall frequency. As with other statistical analyses (including those describing Zipfian power-law distributions), this methodology does not, in general, induce appropriate thresholds for statistical outliers, such as hapaxes and low-frequency lexemes. Witness in Table 11, the data for another alleged defective verb, the verb polir ‘polish’.

From this data it is unclear whether the verb is defective according to the N&L-pattern or just defective in the entire present tense since, on the basis of the verb’s overall frequency, the probable range of values for the 1pl.prs.ind form was between 0.01 and 1.07, with a predicted expected value of 0.10. To compensate in cases that exhibit a skewed or ambiguous distribution the Poisson distribution was used to calculate if the form should be classed as defective, since equation P(0) = G0e-G/0! can predict, from the minimum occurrence value (G), the probability (P) that the actual frequency in the corpus is zero. The result for polir was a p-value of 0.99 that the 1pl.prs.ind should produce a result of zero in the corpus. This form was therefore considered not defective and the lexeme, overall, was considered defective according to the N&L-pattern.

The general problem of extrapolating from the distribution of sparsely-attested lexemes is, however, intrinsic to the kind of statistical methodology adopted in this study. Extremely infrequent lexemes will be predicted as having a threshold value of zero in the corpus, and, counterintuitively, will be classified as non-defective items with full paradigms. Therefore, for the purposes of the present study, and on the basis of results from the Spanish data, two types of lexemes were excluded from the statistical analysis: (a) those whose overall frequency in the corpus was less than 20 and (b) those which were just attested in the pst.ptcp/adj form. The former were classed as infrequent and their inflectional forms were noted, the latter were classed as pst.ptcp /adjectives.

6 Purely morphological/morphomic account of defectivity II

At first glance, the Spanish and Portuguese data seem confusing since it makes little sense that the Spanish verb ungir ‘anoint’ and the Portuguese banir ‘banish’ should be defective, since on the model of regular word formation for -ir verbs in these languages the 3sg.prs.ind and prs.sbj forms should be un[x]e, un[x]a and bane, bana respectively; they are entirely predictable. However, the morphomic explanations for defectivity in Spanish (Maiden and O’Neill 2010: 121; O’Neill 2009, 2010) can explain this conundrum. The basic argument is the following: the cells of the morphomes were, for historical reasons, the locus of different types of allomorphy in a core set of frequently occurring -ir verbs. This allomorphy could not be predicted based on verb-forms outside the morphome and thus the allomorphs were lexically listed. This lexically listed nature of the morphomic forms of a number of verbs with a high token frequency was then analogically extended to all verbs of this class. The defective verbs are those verbs for which speakers have not heard and thus memorized a form for the different patterns.

The claim that the functional value of morphomic N&L patterns contributes to the preservation of patterns in Spanish makes testable predictions for closely-related languages. If defectivity is a product of the N&L-pattern forms of -ir verbs being largely unpredictable due to them being the locus of different types of allomorphy, then the prediction would be that defective verbs should exist in related varieties in which the N&L-patterns are attested and they are also the locus of different types of lexically bound allomorphy. Likewise, defective verbs should not be attested in those varieties in which either the N&L-patterns are not active or the allomorphy is largely predictable. In this context, it is instructive to analyse Catalan, a closely related language to Spanish but one which has no listed defective verbs in the N&L-patterns.

Catalan has two different sub-paradigms for the present tense forms of –ir verbs. The first, traditionally termed conjugation IIIa, exemplified by the verb sentir ‘feel’, has rhizotonic stress and historically would have been the locus of extensive vowel allomorphy in the N&L-Pattern cells but, for a number of complex historical reasons, this allomorphy was levelled (see Wheeler (2015) for an overview). In the second sub-class of –ir verbs, traditionally termed conjugation IIIb and exemplified by servir ‘serve’ and abolir ‘abolish’ in Table 14, the stress falls on the augment –eix (Maiden 2004b) which occurs after the root and is present for all verbs of this sub-paradigm. Therefore, whilst the Catalan verbs do display allomorphy in the N-pattern, unlike Spanish, for the IIIb conjugation this allomorphy is always predictable on the basis of any other form of the verb.

It appears to be no coincidence, that the Catalan cognate forms of the attested Spanish defective verbs^[7] all belong to this IIIb conjugation, which display the augment –eix in the N-pattern cells. These verbs all have a full paradigm even if they are extremely infrequent or, used mostly in the past participle form as adjectives. The reason why Catalan, unlike Spanish, has no similar defective verbs is because, for the Catalan IIIb class, knowledge of inflected forms outside the N-pattern predicts the forms of the N-pattern for all verbs.

Portuguese, on the other hand, does have defective verbs as attested in section 5.1. And, as expected, Portuguese has extensive L-pattern and (modified) N-pattern allomorphy as can be appreciated in Table 5, Table 8 and Table 9 and summarized in Table 15 below for -ir verbs. The explanation proposed for the Spanish data thus extends straightforwardly to the Portuguese patterns. The allomorphy in these cells was so unpredictable that it became a generalisation for this conjugation that the N&L-pattern forms were listed: the defective verbs are low frequency lexemes which, for historical reasons, have either lost or not acquired their N&L-pattern forms.

Confirmation of the lexical nature of this allomorphy in –ir verbs comes from experimental studies conducted with entirely independent grounds for distinguishing regular from irregular patterns, both in Portuguese (Veríssimo and Clahsen (2009) and Spanish (Linares et al. 2006: 113; Rodriguez-Fornells et al. 2002; Yang 2016: 149). Yang (2016: 150) specifically notes for Spanish that ‘for every item in the third conjugation, learners must have positive evidence for its stem alternation because it is lexically arbitrary. And they will be at a loss for verbs that do not have attested forms, resulting in gaps.’.

In sum, the essential particularity of –ir verbs in both Spanish and Portuguese can be correlated with the high token and type frequency of lexemes in this class which display allomorphy precisely in these patterns. These lexemes forge out a structure for this class of verbs whereby the N&L pattern forms are not predictable by the other forms of the paradigm, but must be memorized. This pattern, which is only explicitly present in a number of ‘irregular’ lexemes, is then adopted by all verbs of this class, even those classed as ‘regular’.

Note also, that the general idea that irregular forms with a high token frequency can act as a type of cognitive anchor or prototype around which to organize less-frequent forms, is also the conclusion for a number of studies in CL (Goldberg et al. 2004; Strack and Mussweiler 1997) and is supported by historical data (Fertig 2013: §7.3.6, §8.6.1; Sims-Williams 2022). Additionally, there is historical (O’Neill 2014: 58–65) and computational evidence (Pirrelli et al. 2007) to support the specific claim that the patterns of inflection of frequently occurring and irregular verbs can be adopted by/are valid for verbs classed as regular, thus challenging the regular/irregular dichotomy.

7 Discussion

The present analysis of defective verbs in Spanish and Portuguese assumes a notion of the morphome that plays a central role in the structure of these Romance languages and assumes that they constitute a psychological reality for their speakers. Morphomes in this sense are intra-morphological meaningful structures, providing information about other forms, rather than about language-external elements. Hence, morphomes lack counterparts in models that approach the description of language in terms of associations between ‘units of meaning’ and ‘units of form’, and have provoked resistance from advocates of these types of ‘constructive’ models (see Blevins (2006) for the constructive/abstractive distinction and Luis and Bermúdez-Otero (2016) and O’Neill (2022, 2024) for the polemic nature of morphomes), which have the morpheme as the basic unit of storage and processing, and derive word-forms based on symbolic rules associated with these underlying meaning-form pairings.

Models of morphology within CL are not constructive but, rather, abstractive: the minimal meaningful unit and basic element of lexical storage is the word, and complex word forms are stored in their entirety in the lexicon, where interconnections between words ‘provide[s] generalisations and segmentation at various degrees of abstraction and generality whereby units such as morpheme, arise from the relations of identity and similarity that organize representation’ (Bybee 2001: 7). Abstractive models of morphology have been argued as being more suitable to model morphomes (Blevins 2010, 2016; O’Neill 2014), and, in fact, the historical diagnostic for the psychological reality of morphomes (the fact that they act as a domain within which morphophonemic alternations are levelled) was directly borrowed from Bybee’s model of morphology. Nevertheless, morphomes have largely not been recognized in the CL literature due to their lack of reference to system-external information, i.e., meaning. However, as noted throughout this article, morphomes are meaningful in the broader sense of being informative: they exhibit systematic patterns of form that have a high predictive value within the morphological system.

Meaning is central to CL in which morphological structure is conceived as being symbolic in nature (Langacker 2019), and as the union of semantic and phonological connections. These two types of connections do not have equal status, however; semantic relations always take precedence over purely formal ones. This point is reiterated and stressed on a number of occasions by Bybee (2001: 117, 1985: 118), in whose model of morphology, a fundamental tenet of the lexical organisation of whole word forms is that form is subordinate to meaning to the extent that “the formal organisation of a paradigm diagrams the semantic organisation so that forms that are more similar semantically will be more similar in morphophonemic shape.” (Bybee and Brewer 1980: 58). The central point of morphomes, however, is that one can have consistent patterns of similar form which do not diagram any type of system-external semantic organisation but have an essential system-internal predicative function.

At this point, it is informative to review Bybee’s views and interpretation of velar allomorphy in Spanish (defined as the L-pattern morphome in Table 4). Bybee (1985: 68) admits the lack of semantic relations between the L-pattern cells and defines them as corresponding to a ‘coherent morphological set’ (Bybee 1985: 71). However, whilst she acknowledges that such purely morphological patterns were ‘originally morphologically arbitrary’ (Bybee 1985: 71), she is reluctant to accept that they could endure in the language without any reference to system-external information and claims such forms ‘can be preserved if they coincide or can be made to coincide with the morphological relations among forms.’ (Bybee 1985: 68). Thus, the purely morphological distribution is justified with reference to issues relating to markedness, but in a way very specific to Spanish since she acknowledges (Bybee 1985: 78n9) that markedness alone cannot explain the distribution of this allomorphy. This markedness-based motivation for the L-pattern suffers from a number of problems (O’Neill 2015: 493n4), however, especially regarding the actual data.^[8]

Abstracting away from the specific data, however, what we have here is an acknowledgement that morphophonemic alternations need not always diagram semantic ones but a reluctance to accept that purely morphological relations can structure morphology without any reference to system-external factors. This stance is typical of CL approaches to morphology whereby there is a tacit assumption that form-form relations are secondary and a tendency to (a) try to explain what seems to be purely morphological phenomena in terms of semantics^[9] and (b) to only recognize purely morphological factors as a last-resort option when all other possible analyses have been exhausted. For example, one anonymous reviewer suggested that the judgements of relatedness and unrelatedness between the Romance morphomes are impressionistic and pointed out that there are more objective methods for quantifying relatedness of paradigm cells via distributional semantics (e.g., Chuang et al. (2022)).

CL should be open to the possibility of the importance of form-form relationships. Two central tenets of cognitive linguistics are that language is a system whose primary function is to convey meaning and that grammar emerges through usage. As noted, however, usage has produced various purely morphological phenomena which have no relation to established semantic meanings. The classic example is inflectional class. It is widely acknowledged that alongside languages with classes whose membership is correlated with semantic meaning, there are numerous languages in which membership is totally arbitrary and different lexemes have different, but systematic, morphological alternations depending on their random classification. Enger (2019) notes that ‘Langacker (1987: 422) is completely unfazed by arbitrary distributional classes, whose existence he simply acknowledges; they do not violate his “Content Requirement”’. Also, within the broader understanding of meaning as adopted in this article, these arbitrary different classes can be classed as meaningful: they are informative system-internally and have a high predictive value despite their lack of ‘system-external information’.

Additionally, whilst it is true that other methods for quantifying relatedness of paradigm cells via distributional semantics exist, and these, in fact, have even been used to motivate defectivity in Russian nouns semantically (Chuang et al. 2022), one must question, however, such methods being classified as objective and also their suitability as models of human cognition. As is typical of many mathematical and statistical models, their results are often an artefact of the researchers’ choice of representation, and so they are not therefore objective. A case in point is the distributional semantic analysis of morphomic structures in Spanish by Herce and Tang (2023) and Herce (2022), whose methodology and experimental design are flawed.^[10]

Even for models with a sound methodology, the ecological validity of these mathematical models remains to be established, in the form of external evidence that supports the claim that they model human cognition, their mechanisms are cognitively viable and their conclusions are linguistically relevant for speakers. The fact that a computer can detect semantic and syntactic generalisations in linguistic data, does not mean (a) that humans necessarily have detected the same patterns and (b) that such patterns are relevant to linguistic structure in that there exists a cause and effect relationship (see also Divjak (2015)). Indeed, the conclusions of the present analysis of defective verbs in Spanish and Portuguese are similar to those of Yang (2016). His conclusions were based on a mathematically defined principle, the Tolerance Principle (Yang 2016: §73), which claims that speakers have a certain level of tolerance to the amount of exceptions to a rule; once this level is passed it is hypothesized that, instead of searching for other abstract generalisations, speakers simply store forms.

The central idea behind the morphome is that speakers have detected systematic, system-internal patterns of form across stored words and these patterns are psychologically real for speakers, as attested by the way that they can condition historical change (Maiden 2018). I would like to emphasize the importance of patterns independently of the actual form which expresses these patterns and that the importance of such patterns has not been fully considered in abstractive theories of morphology. For example, Bybee makes specific reference to the morphomic vowel alternations of Spanish in Table 6 (these are of different types but the paradigmatic pattern is the same across lexemes) and whilst she does not reject the idea that lexical connections can be made on the basis of formal identity alone, she states: ‘In many cases, lexical connections among paradigms with similar alternations are not justified’ (Bybee 1985: 131). Bybee is not wrong to note that consistent form in a semantically unmotivated set of paradigmatic cells could be a mere coincidence, thus invalidating a justification of lexical connections. However, as referenced in the introduction, there is much synchronic and diachronic evidence for morphomes and the current analysis of defective verbs in Portuguese adds to this evidence.

Morphomes are essentially abstract patterns and schemas. Schemas play a central role in morphology in CL and are defined as ‘output-based generalizations over forms occurring in utterances’ (Nesset 2019). Such schemas, which are internally predictive, offer much potential to model morphomes. This is especially true of ‘second-order schemas’ (Booij 2016; Kapatsinski 2013; Nesset 2008), which represent the implicative relationships between different paradigmatic forms, often across different inflectional classes. Such second order schemas have been used to model the purely paradigmatic relationship that exists between the 3pl.prs. and the active participle in Russian, which, cross-conjugationally, always share the same root and thematic vowel: the forms X[ut] and X[at] in the 3pl.prs alternate with X[uʃ̩ ʲ:] and X[aʃ ʲ:] in the participle (e.g., razgovarivajut ‘they converse’ vs. razgovarivajuščij ‘conversing’; prosjat ‘they ask’ vs. prosjaščij ‘asking’ (Nesset 2008: 60–61, 2019: §9)). In fact, the relationship between these two inflectional forms, has been related to the concept of the morphome by Nesset (2008: 60), who discusses it in the context of other formalizations of how to deal with such phenomena and states that “An evaluation of the merits of these proposals is beyond the scope of this study, but it is interesting to notice that cognitive linguistics does not need any additional machinery in order to accommodate basic-derived relationships”. The question is, however, whether such schema can handle the central insight of the morphome regarding the psychological reality of phonologically heterogenous allomorphy with the same paradigmatic distribution/pattern. I leave this to future research and hope that in this article I have made a case for the validity of morphomes and the potential they have to be modelled in and inform discussions about models of morphology within CL.

8 Conclusions

In this article a statistical analysis of corpus data from European Portuguese has been used to establish that defective verbs do indeed exist in this variety of Portuguese and are extremely similar to those attested for Spanish. The morphomic account of defectivity for Spanish verbs has been extended to Portuguese and this account has also been expanded and developed further. The claim is that the verbs in Portuguese and Spanish which are defective are so due to them not having a stored form for the N&L patterns, and the morphological generalization of -ir verbs is that the forms of these patterns need must be lexically retrieved and usually not created via knowledge of inflectional forms outside these patterns. This latter idea is also independently supported by a number of experimental and theoretical studies on Spanish and Portuguese -ir verbs (Linares et al. 2006: 113; Rodriguez-Fornells et al. 2002; Veríssimo and Clahsen 2009; Yang 2016: 149) and the overall view of defectivity in these languages is compatible with the views of Gorman and Yang (2019), who see defectiveness as the result of when a number of rules are in competition and none of them can be defined as productive.^[11]

The hypothesis put forward here is that, in acquisition, the frequent lexemes of the -ir class provided a blue-print for the geometry of the entire class whereby the N&L-pattern forms had to be memorized. This was due to the fact that, in a number of frequent lexemes, the different types of allomorphy was unpredictable based on other inflectional forms of the verbs. This tendency for the paradigmatic patterns of a high-frequency group of irregular verbs to be relevant to regular verbs is also attested diachronically for other morphomes (O’Neill 2014) and has been modelled computationally (Pirrelli et al. 2007: 285) for the acquisition of morphology.

In CL little attention has been paid to morphomic structures on account of their lack of reference to meaning and the central role which meaning plays in CL. Here it has been argued that morphomes are meaningful in a broader sense of them being informative: the forms in a morphome exhibit a high degree of inter-predictability that allows them to serve a mutual diagnostic function. That is, they may not be informative (=meaningful) in extra-morphological ways but they are crucial for the organisation of the morphology as a complex inter-predictive network which enables speakers to extrapolate from their linguistic input and thus produce forms they have never heard before.

Interestingly, in the case of Spanish and Portuguese, however, the organisation of this network is such that it also encourages speakers not to produce forms that they have not heard before, resulting in the attested defective verbs in these languages.