1 Introduction

Morphological predictability relations in paradigms have been explored for a long time. In classical language pedagogy, principal parts allowed learners of Latin, for example, to accurately predict any inflected form of a lexeme's paradigm from a small subset of word forms (see Finkel & Stump, 2009). Morphologists have long known, thus, that “one inflection tends to predict another” (Matthews, 1991: p. 97), so that Lat. NOM.SG templum predicts GEN.SG templī, NOM.PL templa, etc. Computational and theoretical advances in the last few decades, however, have enabled more exhaustive, systematic, and faster analyses of inflectional systems than ever before.

In recent years, the complexity of inflectional systems has started to be analyzed more and more often from the perspective of the Paradigm Cell Filling Problem (PCFP, Ackerman et al., 2009), which is the name given to the challenge that speakers of languages with inflection classes face to produce essentially any form of any lexeme, no matter how infrequent the lemma or paradigm cell. Although high frequency forms could be stored in memory and learned by rote, given the rank frequencies of words in natural languages (Zipf, 1932), with a long tail of extremely low or hapax (i.e. single-occurrence) words, most inflected words must be built online, with their forms predicted by language users on the basis of other more frequent forms in their paradigm, and/or on the basis of the equivalent forms or patterns as encountered in other more frequent lexemes.

The methods to explore these morphological predictability relations within inflectional or derivational paradigms have been expanding in recent years. Traditional philological- comparative qualitative methods (e.g. Malkiel, 1966, Maiden, 1992, 2018, O’Neill, 2018, Esher, 2022, Herce, 2022) continue to be pursued alongside quantitative computational ones involving Set Theory, Graph Theory, Machine learning, and Information Theory (e.g. Ackerman & Malouf, 2013; Beniamine, 2018; Elsner et al., 2019; Malouf, 2017; Sims, 2020; Stump & Finkel, 2013). Entropy (a measure of uncertainty, Shannon, 1948) is the key notion of Information Theory. Applied to paradigmatic forms and the PCFP, conditional entropy can be used to measure the (un)predictability of one form given another. Software has also been developed to calculate different complexity measures automatically over whole systems: Beniamine’s Qumin (https://sacha.beniamine.net/software/qumin/), Finkel’s Principal Parts Analyzer (https://www.cs.uky.edu/~raphael/linguistics/analyze.html), and Sims’ Inflectional Networks scripts (https://github.com/sims120/inflectional-networks) among others.

Alongside these methodological, theoretical, and software solutions, progress has also been made on the side of resources and databases (see e.g. Kirov et al., 2018). Most crucial to the PCFP are inflected lexicons, where all inflected forms of hundreds or thousands of lemmas are listed, preferably in phonological form. This is the data that is required for the quantitative investigation of morphological predictability and complexity. Most underdocumented at present, and hence most urgent to describe, are inflectional systems of non-Indo-European and non-WEIRD (non-Western European Industrialized Rich and Democratic, see Henrich et al., 2010), low-resource languages (Malouf et al., 2020). Better (i.e. larger, phonemized, well-curated) resources, and comparable morphological predictability analyses, however, are also needed for many national standard Indo-European languages. Focusing on Romance, arguably the language family where more attention has been paid to paradigmatic morphology, we have some family-wide but comparatively small inflected lexicons (see Maiden et al., 2010; Beniamine et al., 2020), as well as large lexicons and PCFP-analyses for some of the major languages in the family [namely French (Bonami et al., 2014), Latin (Pellegrini & Passarotti, 2018), Italian (Pellegrini & Cignarella, 2020), Portuguese (Beniamine et al., 2021), and Spanish (Herce, 2023)]. No such inflected lexicon and analysis exists, however, for Romanian (see Diaconescu et al., 2015, and Lőrincz et al., 2022. for lexicons not specialized in the coverage of inflected forms). This is the purpose of the present paper. Section 2 will explain the creation of a verbal inflected lexicon for Romanian in phonological form (VeLeRo), annotated with lemma and cell frequencies. Section 3 presents a morphological predictability analysis of Romanian verbal inflection on the basis of this resource, and discusses the results briefly, particularly how they compare to extant analyses of other Romance languages. Section 4 summarizes the main highlights of the paper and proposes avenues for future research.

2 Building VeLeRo

VeLeRo is an inflected lexicon of Romanian verbs in phonological form. It is based on Barbu’s (2008) lexical database RoMorphoDict, which contains the orthographical inflected forms, lemmas and morpho-syntactic descriptions of around 700,000 Romanian words. The verbal lemmas and their inflected forms were extracted from this dataset and transcribed phonologically using Epitran (https://github.com/dmort27/epitran). Epitran is a massively multilingual, rule based G2P (grapheme to phoneme) system with support for 61 languages and distributed as open source software (a Python library) under an MIT license (Mortensen et al., 2018). Epitran’s rule-based conversion provided a broad phonemic transcription, but certain aspects required further refinement. Thus, the initial conversion was only used as the starting point for later adjustments. Overall, 51% of the words were modified after the G2P conversion to represent the Romanian phonological system more accurately (e.g. regarding diphthongization, palatalization, and the rest of topics discussed in the upcoming sections). Using the Ratcliff/Obershelp string matching algorithm provided by the difflib Python library, there was a 93% match between the initial Epitran phoneme conversion and the final version of the lexicon.

Our inflected lexicon of Romanian verbs (which can be found freely available online at https://osf.io/kqrjg/?view_only=f1583503953d45028d3bb85a2e1c6d01) adopts a wide format, with each row corresponding to a lexeme (identified through the infinitive in orthographic form (e.g. face ‘do’), and columns representing all 39 forms of every lexeme in phonological form (e.g. inf fáʧe, ind.prez.1sg fák, ind.prez.2sg fáʧʲ, etc.).Footnote 1 The following sections outline the procedures and decisions involved in the phonemization. A full inventory and description of phonemes can also be found online.

2.1 Stress assignment

In Romanian, stress can be oxytonic (final syllable), paroxytonic (penultimate syllable) or proparoxitonic (antepenultimate syllable) and is not marked orthographicaly. Some authors (Chitoran, 1996) have claimed that stress is highly predictable, depending, among others, on the part of speech of the lexical item, but some others (Dindelegan & Maiden, 2013), have claimed that stress is largely unpredictable and highly mobile, especially in verbal inflections. RoMorphoDict was used as the point of departure for building the present resource because it indicates stress consistently on polysyllabic verbs. For consistency (i.e. to avoid spurious morphological contrasts), we added stress markers on monosyllabic verb forms as well (on the only vowel when a word had only one, or on the most open vowel in case the word contained a diphthong or triphthong).

2.2 Diphthongs vs hiatuses

Another important point when it comes to Romanian phonology is the representation of the tautosyllabic and heterosyllabic sequences of vowels (i.e. diphthongs/triphthongs and hiatuses). According to Chitoran (2002) Romanian has two non-controversial diphthongs, namely /e̯a/ and /o̯a/. Apart from these, the glides /j/ and /w/ can combine with most vowels to create additional ones. For the creation of this lexicon, we adopted Chitoran’s (2002) treatment of diphthongs, where /j/ and /w/ have phonemic status and can be predicted by looking at syllable boundaries. Syllable boundary contrasts are not handled by Epitran, so these had to be encoded manually. For this, we used Barbu’s (2008) RoSyllabiDict database as a reference for the syllabification of verbs and made the necessary adjustments by hand. For example, in words such as “a.ban.do.nea.ză”, the sequence “ea” is homosyllabic, and was hence coded as a diphthong (e̯a). On the other hand, in words like “a.gre.a.se”, were “e” and “a” are heterosyllabic, “ea” was transcribed as a sequence of vowels /ea/. The lexicon contains a total of 16 diphthong and 7 triphthong sequences comprising different glide—vowel combinations plus /e̯a/ and /o̯a/.

2.3 Diphthong reduction

Despite varying or inconsistent phonemic transcription practices elsewhere (e.g. in Maiden et al., 2010’s Oxford Database of Romance Verb Morphology), sequences that involve postalveolars (e.g. /ʃ/) followed by /j/ or non-syllabic /e̯/ have been uniformly transcribed without this second segment here (e.g. /ziʧám/, rather than /ziʧe̯ám/, for ziceam ‘say.1SG.IPF.IND’). This is justified by (i) the absence of an audible front vowel in these sequences, (ii) the absence of minimal pairs based on these sequences (for example /ʧe̯a/ vs /ʧja/ vs /ʧa/), and (iii) by the phonemic transcription conventions in parallel cases in related Romance languages, for example Italian /diʧámo/, rather than *diʧjámo for diciamo ‘say.1PL.PRS.IND’, or Spanish /riɲó/ rather than *riɲjó for riñó ‘scold.3SG.PST.IND’ (note that in this latter case the standard spelling reflects pronunciation accurately).

2.4 Palatalization

Palatalization is a prevalent feature of Romanian phonology and generally uncontested by most grammars. According to Chitoran (2002: p. 173), palatalization in Romanian can be phonologically or morphologically conditioned. The former refers to those instances where a change in articulation occurs automatically in a given phonetic environment (i.e. is allophonic). The latter refers to the cases where this environment has disappeared, thus leaving the palatalized consonant not predictable from its phonetic environment (i.e. phonemic). An instance of the first type is the palatalization of velars like /k/ and /g/ before a front vowel, where they are realized as [c] and [ɟ] (Dindelegan & Maiden, 2013) or [kʲ] and [gʲ] (Chitoran, 2002) depending on the source. Because this inflected lexicon is aimed at capturing phonemic representations, this automatic allophonic palatalization has not been represented when morphologically inconsequential.Footnote 2 Only the latter type of palatalization (i.e. phonologized, unpredictable), thus, is relevant in the context of this inflected lexicon. The clearest case of this type is represented by the (orthographic) ending − i occurs in the second person (singular and plural) in verbs, and often indicates a palatalization of a word-final consonant rather than a full word-final vowel /i/. Epitran, again, does not account for this, so modifications were necessary (e.g. to transcribe the form dați ‘give.ind.prez.2p’ as /dáʦʲ/ rather than /dáʦi/). Regarding this palatalization, the choice was made to distinguish between post-alveolar (e.g. /ʃ/) and post-alveolar palatalized (/ʃʲ/) phonemes. Although there is some research suggesting that these are extremely close acoustically (see Spinu et al., 2012, 2019), most speakers appear to be able to distinguish the two sounds (Spinu, 2018). Thus, a form like ziseși “say.2SG.PST” was transcribed as /ziséʃʲ/. A full vowel /i/ (also /u/ in the 1SG.PRS) has been preserved in Romanian pronunciation, and hence transcribed as ‘I’ here, only after segment sequences of “muta cum liquida” (e.g. /íntru/ ‘I enter’, /íntri/ ‘you enter’, vs /egzíst/Footnote 3 ‘I exist’, /egzíʃtʲ/ ‘you exist’).

2.5 Resource overview

After all these steps, we arrived at a consistent phonemization of the complete paradigms of 7297 Romanian verbs, for a total of 284,583 word forms. This is comparable (see Table 1) to the size of extant inflected lexicons in phonological form from other national standard Romance languages. Although this feature will not be used in the second part of this paper, our resource also includes overabundant (Thornton, 2012) or substandard/dialectal forms (e.g. /fúrəm/ occurs, alongside standard /fusérəm/ as the ind.perf.1pl form of the verb fi ‘be’). Given the importance of usage frequency for morphological learnability, predictability, and the PCFP, this information has also been supplied. The frequencies of lemmas in the corpus CoRoLa (Tufiș et al., 2019), a digital lemmatised corpus with over 1 billion words, have been added to VeLeRo and pattern in the way illustrated in the left-hand panel of Fig. 1, with 9 verbs above 1 million tokens, 152 above 100,000, 727 above 10,000, 2146 above 1000, 4109 above 100, 5601 above 10, and 6688 verbs with at least 1 token in CoRoLa, while 609 verbs from our dataset, are completely unattested in that corpus. The frequency of paradigms cells (based also on CoRoLa)Footnote 4 is shown through the rank order plot in the right-hand panel of Fig. 1. These also vary widely between the 3,152,827 tokens of the ind.prez.3sg (the most frequent value) and the 240 of the ind.perf.2pl (the least frequent one).

Table 1 Size of VeLeRo and comparable inflected lexicons in other Romance varieties
Full size table
Fig. 1
figure 1

Rank frequency of the 7296 verbs in VeLeRo (left) and of its 39 cells (right)

Full size image

3 A quantitative analysis of the PCFP in Romanian verbal inflection

To keep computational times within reasonable limits, and to allow for comparability with extant quantitative analyses of the PCFP in other Romance languages, we decided to select for further analysis those (3564) verbs with 200 or more tokens in CoRoLa. This measure/threshold is justified due to two main reasons. The first is that many of the lower frequency verbs are unknown to even highly-educated native speakers of Romanian, and hence probably do not really form part of the average acquired inflectional system of the language, despite the presence of these verbs (and many others) in dictionaries and grammars that are understandably aimed at exhaustivity. The second is that this number of verbs is closer to the average number of items analyzed with identical methods in the extant literature on other Romance languages (see Bonami et al., 2014, Pellegrini & Passarotti, 2018, Pellegrini & Cignarella, 2020, Beniamine et al., 2021, and Herce, 2023), which will enable us to draw more meaningful cross-linguistic comparisons.

After the mentioned exclusions (which included nonstandard overabundant forms), all remaining verbs and forms were analyzed in Qumin (Quantitative Modelling of Inflection, Beniamine, 2018). This is a set of Python scripts that automatically extracts morphological alternations between all possible pairs of word forms in all lemmas. Due to the sheer number of combinations, 1482 (= 39 × 38) per verb, this is a task that can only be performed computationally. The algorithm finds maximally generalizable morphological alternations,Footnote 5 and derives a knowledge of which verbs show the same contrasts (i.e. belong to the same inflection class) and which have different ones (i.e. belong to different classes).

Consider, as an illustrative example in Table 2, some of the ways in which the 2SG present indicative can differ in Romanian from the 2SG imperative. While in some verbs (forma and facilita) these forms differ in identical ways (/ézʲ/ in the former value has to be replaced by /e̯ázə/ to form the latter and vice versa), these forms contrast in different ways in other verbs (e.g. duce adds ʧʲ to the 2SG imperative to form the 2SG present indicative, but the same rule does not apply to derive the same form in comunica). For the purposes of this pair of cells, thus, the former verbs (i.e. forma and facilita) belong to the same class, while the latter verbs (i.e. duce and comunica) belong to different morphological classes. To aid with sequence-to-sequence alignment, and the interpretability of morphological alternations, the Qumin algorithm also makes use of distinctive phonological features. A separate file needs to be supplied which defines the language’s phonemes and their features (e.g. + or−voiced, + or−nasal, + or−velar, etc.). This file can be found online, along with VeLeRo itself.

Table 2 Extracted morphological alternations between two cells in eight verbs (A)
Full size table

As in the eight illustrative word-form pairs in Table 2, patterns of morphological alternations are extracted for all word-form pairs of all verbs in our sample (1482 × 3564 = 5.3 million alternations). This is a demanding computational process that can last several hours. After they are extracted, the patterns can be inspected for quality control (making sure, for example, that infrequent or exceptional patterns are not due to mistakes or inconsistencies in either the original inflected lexicon or its subsequent phonemization). The extracted patterns also constitute the basis for various other scripts and functions within Qumin that allow to calculate further measures like conditional entropies, group inflection classes, etc.

In the extant literature on the PCFP in Romance and beyond, the analysis of morphological predictability within paradigms often starts with a presentation of which values or word forms are mutually interpredictable with complete certainty. Thus, although the morphological relationship between some cells (e.g. ind.prez.2sg and imper.2sg in Table 2) is a heterogeneous one, in the sense that it varies unpredictably from verb to verb, the morphological difference between other cells is the same across all verbs. This is the case, for example, of the ind.imperf.1sg and ind.imperf.2sg. As Table 3 shows, the former can be reliably transformed into the latter by replacing a word-final /m/ with /j/, and, conversely, the latter can be transformed into the former by changing this final /j/ into /m/.

Table 3 Extracted morphological alternations between two cells in eight verbs (B)
Full size table

These paradigmatic domains of interpredictability, comparable to the notions of ‘stem space’ (see Montermini & Bonami, 2013) or ‘distillations’ (Stump & Finkel, 2013), provide a first measure of the complexity of an inflectional system. From the 1482 pairs of cells in a Romanian verbal paradigm, 236 (15.9%) involve no uncertainty (i.e. they have a conditional entropy of zero). In terms of concrete paradigm cells (see Fig. 2), the 39 cells of the Romanian verbal paradigm can be classified into 14 areas of interpredictability.

Fig. 2
figure 2

Areas of morphological interpredictability in Romanian verbal inflection

Full size image

As Fig. 2 shows, many of these areas (Z1, Z5, Z6, Z12, Z13, Z14) are single-cell ones (e.g. Z1 is the imper.2sg) and hence trivial "areas" to some extent, since they merely indicate that some cells are not interpredictable with any other cell. A similar case is the one represented by those forms which contrast in values (e.g. ind.prez.1sg and conj.prez.1sg) but never in their form (dáw dáw, jáw jáw, forméz forméz, faʧilitéz faʧilitéz, dúk dúk, etc. for the verbs in Table 2). Systematic syncretisms like this account for areas Z3, Z4, and Z7. Remaining areas (i.e. Z2, Z8, Z9, Z10, and Z11) are the ones based on predictable morphological alternations like the one in Table 2).

Notable commonalities can be identified between Romanian verbal inflection and that of the other major Romance languages analyzed with this same methodology to date (Beniamine, 2018, Pellegrini & Cignarella, 2020, Beniamine et al., 2021, and Herce, 2023). The number of interpredictability areas (14 in Romanian, vs 15 in Italian [and Latin], 14 in Spanish and French, and 12 in Portuguese), and their distribution (e.g. most areas in the present indicative) are very similar to those found in other Romance languages. As in all other Romance languages analyzed so far, the 1SG present indicative, and the past participle(s) constitute areas of their own. Some other aspects are shared with most but not all other Romance languages, for example, the fact that the 2SG imperative is also a one-cell area of its own (shared with all of Romance except Portuguese), or the fact that all the imperfective indicative cells constitute another area together, to the exclusion of all other cells (shared with all except French).

Our results also show, of course, some differences to the patterns found in other Romance languages. A somewhat trivial one concerns the raw number of values a verb can inflect for, which is less in Romanian than in the other major Romance languages, mainly due to the absence of the synthetic future and conditional tenses, which did not emerge outside Western Romance (i.e. Portuguese, Spanish, French, Italian). In the cognate persons and tenses, another well-known difference is the one derived from the fact that the morphological subjunctives of the first and second person have been replaced by the corresponding indicative forms. This has generated a morphological overlap between present indicative and subjunctive not seen generally in other Romance languages.

Regarding more aspects where there is within-Romance variation, Romanian patterns like Spanish and French (also like Latin), but unlike Italian or Portuguese, concerning the interpredictability of 1PL and 2PL present indicative. Regarding the absence of morphological interpredictability between the 2SG and the 3SG present indicative, and between the 1PL present indicative and the infinitive, Romanian patterns like Italian, and unlike Spanish and Portuguese.

Although this goes beyond the goals of the present paper, these properties (e.g. whether any pair of cells is (1) or is not (0) mutually predictable) could be represented as vectors of (binary) values for each language, and between-language similarity could be explored via Hamming distances or similar (see Table 4) to check if paradigmatic structural distance corresponds to phylogenetic distance or is instead more similar to others like orthographic distance or mutual intelligibility (see Ciobanu & Liviu).

Table 4 Presence (1) or absence (0) of morphological properties across Romance languages and Latin (left), and between-language Hamming distances (right)
Full size table

A finer-grained approach to predictability reveals that some of these differences are not so categorical. Although 2SG (Z4) and 3SG present indicative (Z5), and 1PL present indicative (Z2) and the infinitive (Z12), are not perfectly interpredictable in Romanian as they are in other Romance languages, these forms are still very close to perfect predictability (i.e. conditional entropy = 0 in both directions in Table 5) but ultimately fall short of it.

Table 5 Conditional entropies (column given row) between the distillations in Fig. 2
Full size table

The average implicative entropy between Romanian cells is overall 0.1467 (0.1804 between distillations), which is slightly lower than in the other Romance languages that have been analyzed in a comparable way except Spanish: 0.28 for Latin (Pellegrini & Passarotti, 2018), 0.18 for French, and 0.17 for Portuguese (Beniamine, 2018), 0.07 for Spanish (Herce, 2023). Conditional entropies between Romanian verb distillations differ widely, between 1.172 as the highest [the uncertainty involved in predicting Z6 (ind.prez.3pl) from Z3 (the ind.prez.1sg)] and 0 as the lowest. Closest to perfect interpredictability are Z5 and Z6, i.e. 3SG and 3PL present, and the different areas within the former perfectum tenses (aka. PYTA in the literature on Romance stem alternations, see Maiden, 2001), e.g. Z9 and Z10, as well as these areas and the participle (i.e. Z14).

Table 6 illustrates some of the high-entropy (red) and low-entropy (green) alternations just mentioned, with the numbers in parentheses indicating the number of verbs (if > 1) for which a particular alternation holds. Predicting the 3PL present indicative from the 1SG of the same tense is difficult (in this direction) because there are various frequent unpredictable ways in which the latter form can be turned into the former, most importantly ⇌ e (i.e. add /e/), ⇌ ə (i.e. add /ə/), and ⇌ (i.e. leave unchanged). Predictions in the opposite direction (i.e. predicting the 1SG form from the 3PL) are easier, and the same applies to predictions between 3SG and 3PL present indicative, because most forms (e.g. ʤe ⇌ g, ʧe ⇌ k, ʃte ⇌ sk) allow a speaker to deduce one from the other. Note that, among the alternations shown, only ⇌ and e ⇌ overlap in forms, which means a ‘zero’ 3PL does not fully diagnose the 3SG, which could be formed either with ‘zero’ (the most frequent option), or by adding − e. This is the reason why, as shown in Table 5, the conditional entropy of the 3PL (Z6) given the 3SG (Z5) is 0 (no uncertainty) but that of the 3SG given the 3PL is not zero (although still very low).

Table 6 Some present indicative inflected forms and alternations in Romanian
Full size table

Table 7 shows the average predictability (average of columns in Table 5) and predictiveness (average of rows in Table 5) across all distillations. Unlike in other Romance languages, it can be observed that differences in predictiveness are similar in range to differences in predictability, ranging roughly between 0.5 and almost zero. In common with other Romance languages, however, we can see quite a sharp boundary between the most and least predictable distillations. The zones Z1, Z3, Z5, Z6 and Z7 are all quite difficult to predict from other forms (note that higher numbers indicate less predictability). This is due to unpredictable stem alternations that separate rhyzotonic forms (the so-called N-morphome cells, see Maiden, 2018) from arhyzotonic ones. Consider for example 1PL.PRS.IND.beg rugə́m > 1SG.PRS.IND.beg róg, but ‘occupy’ okupə́m > okúp and ‘calculate’ kalkulə́m > kalkuléz. Looking at predictiveness, in turn, we find that the imperfect indicatives (Z8) are clearly the forms which are least informative about the morphology of paradigm cells from other distillations. This is due to the very low allomorphic diversity of these forms compared to others, since the only lexical-class distinction that applies to them is the contrast between a suffix − a (for first conjugation verbs and a few from the fourth) and a suffix − e̯a (for all others).

Table 7 Average predictability and predictiveness of Romanian distillations
Full size table

Beyond Romance, Romanian behaves like most (or maybe all) languages concerning the association between high frequency and irregularity (Herce, 2019; Wu et al., 2019). This can be observed (see Fig. 3) both at the lexeme level, where verbs from smaller classes tend to be more frequent (see Fig. 3a) and at the paradigm cell level, where cells that belong to small or single-cell interpredictability domains (e.g. Z1, Z5, Z6, etc. in Fig. 2) tend to be more frequent than more regular cells (see e.g. the Z8 cells in Table 3), i.e. cells that can be predicted from multiple other ones in the paradigm.

Fig. 3
figure 3

a Irregularity and frequency in lemmas. b Irregularity and frequency in cells

Full size image

4 Conclusion

This paper has presented a new resource VeLeRo (Verbal Lexicon of Romanian), containing the full paradigms in phonological form of 7297 verbs, as well as the frequencies of these lemmas and of their different inflectional values in CoRoLa (Tufiș et al., 2019). The lexicon is made openly available for further research into quantitative morphology and the PCFP.

After outlining the overall interest of this resource and what role it can have within the general morphological-predictability literature in Section one, we proceeded to explain in Sect. 2 all steps and challenges that were involved into the resource’s compilation: the addition of stress (in a language whose orthography does not indicate it), the phonemization of problematic cases like palatalizations and diphthongs, etc. In Sect. 3, in turn, we used our resource and extant freely-available software [Beniamine’s (2018) Qumin], to conduct an initial assessment on morphological predictive complexity in the system. The measures and results have been contextualized within the overall Romance landscape. Romanian is widely regarded as the most divergent Romance national standard language due to its earlier phylogenetic split (Balkan Romance split off from Western Romance [Port, Sp, Fr, It] before this group of languages started to break up), the lack of geographic contiguity to the rest of the Romance world, and strong language contact with Slavic languages. Although paradigm-structural differences between Romanian and its Romance sisters might correlate with geographic and phylogenetic distances (see Table 4), our results illustrate a great degree of similarity overall to the other Romance languages, with a comparable level of complexity (similar number and pattern of distillations, and average conditional entropies). Differences are found mostly on those aspects where variation exists already within Western Romance (e.g. regarding the morphological-predictive allegiance of the 1PL and 2PL present indicative, infinitive, of 2SG and 3SG or the present indicative, etc.) and tend to be a matter of degree.

Overall, hence, the results point to the diachronic stability of morphological predictive relations within inflectional paradigms. This goes in line with the conservative nature that has been often claimed for inflectional morphology (Meillet, 1958, Nichols, 1996), which is considered to be less prone to borrowing/contact than other components of language (Matras, 2015). This should be particularly the case for structures like inflection classes and stem alternations which do not bear a direct relationship with meaning/function (Maiden, 2018). If the conservativeness of paradigmatic predictability structure is confirmed, this would open the door to the possibility of using the paradigmatic complexity and patterns themselves for phylogenetic purposes (see Herce and Bickel forthcoming), i.e. to diagnose genetic relations between languages or inflectional systems and categories. This should be the focus of future research, along with the creation of large and well-curated inflected lexicons for other languages, particularly minoritized non-WEIRD languages (see e.g. Cruz et al., 2020), which might differ importantly with respect to the larger languages we are more familiar with (see e.g. Trudgill, 2011).