Preferences in the use of overabundance: predictors of lexical bias in Estonian

2.1 Overabundance in Estonian nominal inflection

We set out to investigate nominal overabundance in Estonian, a Finno-Ugric language with no gender and with multiple declension classes and rich nominal morphology. Its nominal inflectional paradigm, which consists of 14 cases and two numbers, amounting to 28 cells, constitutes a valuable resource for investigating overabundance.

The system includes 26 nominal inflection classes, which are not entirely predictable based on phonological form, nor are individual word forms predictable from the reference form (see Blevins 2008). No studies have shown any semantic distinctions between these classes and they are believed not to constitute semantically cohesive groupings (Blevins 2008: 241), although the validity of this assumption is an interesting question for detailed quantitative analysis in future studies. While various classifications have been proposed with varying numbers of classes, most grammars have accepted the 26-class system offered by Viks (1992), where class membership is defined by affixes and stem types in particular cells, described by means of three dimensions. Five of the 28 cells are diagnostic in class identification: nominative singular, partitive singular and plural, genitive singular and plural. First, classes are distinguished based on the pattern by which stems vary across these five cells: some classes have the same stem across all five cells, while others may have matching stems in genitive singular and partitive plural and another stem in the other three cells. Several patterns are available, but lexemes in a single class are uniform in this dimension.

Second, gradation patterns are uniform for all lexemes within a single class. Some classes have weakening gradation, meaning the genitive stem is “weaker” (either in phonological length or type of phoneme) than nominative and partitive stems, while others have strengthening gradation, where the pattern is the other way around, with the genitive stem “stronger” than nominative and partitive. Third, formatives in partitive singular, plural and genitive plural are uniform inside a single class. Inflection classes are referred by arbitrary numbers (1–26), which we will use to indicate distinctions between classes.

With 28 nominal cells and 26 inflection classes there is a good deal of room for deviation from an entirely transparent, one-to-one system, including both cell syncretism and overabundance, the latter shown in the two forms of illative singular for two lexemes in (1).

It is justified to ask whether the inflection classes described in the reference grammars (Erelt and Metslang 2017; Viht and Habicht 2019) correspond to cognitive constructs used by speakers, or whether they are merely a descriptive tool for the linguist. We return to this question in the discussion.

According to academic grammars (Erelt and Metslang 2017; Viht and Habicht 2019), overabundance affects most plural cells in the nominal paradigm, as well as at least one singular cell (illative, see 1). Thirteen out of 26 classes include systematic overabundance for most of their plural paradigm (Kaalep 2018: 146), and another seven predict overabundance in some lexemes. The partitive plural is expected to be overabundant in seven declension classes, and illative singular in twenty classes. This would amount to 22,902 lexemes for the latter, if it were to apply to the entire class (Viht and Habicht 2019: 124–129).

In addition to inflection classes which include overabundant cells (1a), a lexeme may be overabundant through membership in multiple inflection classes at once, e.g. ‘aquarium’, in (1b). This means that the two available paradigms for that lexeme align with separate classes: the -sse illative form (akvaariumi-sse) is part of a paradigm inflected according to class 19 and the short form (akvaariumi) is part of a class 2 paradigm. This is prevalent in the nominal system, as 19 out of 26 inflection classes include some dual-membership lexemes (Kaalep 2018).

Parallel forms are accepted by normative language planning institutions, but usage is rather restricted (Aigro and Vihman 2023a; Kaalep 2010). “The language norm assumes that it is very common for a word to have multiple ways of forming a surface form per some morphosyntactic category bundle. Usage-wise, the distribution of such forms is very skewed, though” (Kaalep 2018: 147).

While there are no systematic, quantitative studies based on extensive datasets of parallel form usage, some studies have linked the preference of partitive plural form to the morphophonology derived from inflection class membership. For instance, class 19 lexemes, e.g., seminar ‘seminar’ are assumed to prefer the stem-changing partitive plural (seminare, seminar.par.pl) rather than an agglutinative option (seminari-sid, seminar-par.pl). Form preference has also been linked to derivational morphology, e.g., nouns formed with the -ik suffix (e.g., kodanik ‘citizen’) also prefer stem plural (kodanikke, citizen.par.pl) rather than the agglutinative -sid (kodanikku-sid, citizen- par.pl, Erelt et al. 1995: 205).

In addition, as mentioned, Kaalep (2010) suggests a connection between form preference in partitive plural and word frequency. Namely, speakers are said to be more likely to prefer the short, stem-changing partitive form with high frequency lexemes, while lower frequency items are expected to occur with the transparent -sid morpheme. This is because short form formation requires lexically specific knowledge of the word-final theme vowel, as the particular vowel cannot be phonologically derived from the reference form. For instance, as shown in (2), two lexemes with highly similar nominative singular forms take different theme vowels in the partitive plural.

Kaalep (2010) posits that the easier accessibility of this lexically specific information for high frequency lexemes could facilitate its dominant (entirely productive) status, and the less accessible status of low frequency lexemes would elicit uncertainty in terms of theme vowel selection, leading speakers to opt for the transparent, concatenative -sid affix instead. In addition, such form use bias would be in line with general linguistic principles, e.g., Zipf’s Law of Abbreviation of Words (1949 [2012]) and Hawkins’ Minimizing Forms principle (2014), describing the tendency of longer lexical items to be less frequent than shorter items.

However, this hypothesis appears incomplete. First, the dataset used in Kaalep’s analysis was rather small, meaning that the results on high frequency items rendering higher usage ratios of short fusional forms need to be confirmed in a larger quantitative study. Second, it does not follow that lexeme frequency alone facilitates the short form becoming dominant across all such lexemes. Other factors have been discussed for illative singular, e.g. syntactic role. The fact that illative singular formative patterns are highly analogous to partitive plural allows our study of form bias comparison to indirectly test Kaalep’s hypothesis.

Illative singular is formed by means of a number of morphological strategies, shown in (3). The pattern described in (3a) is agglutinative, while 3b, 3c and 3e are fusional, often including stem change and an increase in phonological quantity (Viht and Habicht 2019), as seen in the medial consonant lengthening in (3c):

While the affix -sse is assumed to be available to all lexemes (Viht and Habicht 2019: 93), the availability of the other patterns is defined lexically. Even though overabundance may occur in various combinations, parallel form pairs usually involve -sse as one of the two available forms (Aigro and Vihman 2023a).

Siiman (2019, née Metslang), who has investigated illative singular parallel forms, showed that form preference is informed by various lexical factors, including morphophonological as well as semantic and syntactic factors (Metslang 2015, 2016). Morphophonological factors include a number of gradation-related variables, stem-final phoneme and syllable count. Semantic and syntactic factors that were found to influence form choice include the noun’s semantic type and phrase status. Semantic type in Metslang’s study included the levels 1) proper names, 2) body parts, 3) locations, 4) states and 5) other. Phrase status codes verbal selection, following the assumption that when illative marks a verbal argument (e.g. usun sinusse ‘I believe in you’), the long form is preferred by speakers, while the short form dominates in non-selected adjunct roles. The same is claimed in the main descriptive grammar (Erelt et al. 2020: 217). Inflection class was not included, although several of the morphophonological variables underlie class membership.

There are issues with these results. First, the data in Metslang (2015, 2016 includes lexemes with available overabundance rather than only lexemes with attested overabundant usage. This is likely due to the lower availability of annotated corpora at the time of these studies, which restrict investigations of low-frequency phenomena such as overabundance. However, this means that the base assumption of her conclusions may not hold. As shown by Aigro and Vihman (2023a), both forms are used with only a handful of lexemes out of the tens of thousands with available overabundance. For the rest, one form may be so entrenched that there may no longer be any awareness of alternative forms for that particular lexeme in the speaker community.

Furthermore, Aigro and Vihman (2023a) show that Metslang’s results regarding Estonian case function in overabundant contexts do not seem to hold when examining only realised overabundance. They found no alignment of formative choice with particular case functions in a balanced sample of several cases, including illative singular. The same was hypothesised by Kaalep (2009). This further highlights the importance of drawing a distinction between dataset types. When investigating language users’ behaviour in a variation context, especially when this is done based on the indirect evidence of a corpus study, one needs to first confirm that variation truly exists for speakers, which is most credibly done by only focusing on cells for which both parallel forms are simultaneously used (at least to some degree, and even if they are used by different speakers).

In addition, Metslang’s studies (2015, 2016 do not include inflection class information or information on morphological variation patterns. Hence, there remains the possibility that many of the morphophonological and semantic patterns that she identified among nominals may be explained away by inflection class membership. An untested question remains, then, of whether inflection class has any effect on overabundant usage, and if so, whether this effect is independent from the other observed factors affecting form choice. Finally, regarding the role of entrenchment, no quantitative studies have outlined all form distributions for any of the Estonian nominal paradigm cells, and neither are there any quantitative results on the role of frequency in form type. We seek to address these gaps.

3.1 Data and method

The corpus study reported here is based on a dataset including lexemes with both parallel forms in illative singular attested in the corpus (all datasets are freely available at https://osf.io/bsp9d/). Using Python 3.8, the dataset was extracted from the Estonian National Corpus 2021, a written corpus for Estonian totalling 2.4 billion words (Koppel and Kallas 2022). Illative singular is the only cell in Estonian nominal morphology in which language technology tools annotate parallel forms with distinct labels: the long illative (-sse affix, see 4a) is parsed as illative, while the short illative (described in 4b–4d) is coded as aditive.^[1]

Out of all lexemes with at least one illative form in the corpus (n = 422,650), we extracted lexemes where both an agglutinative (-sse, see 3a) and fusional variant (3b–3d) are attested in the corpus and the more frequent form occurs at least 10 times (n = 1,907). For these 1,907 lexemes, we calculated the proportion of the most frequent -sse form and the most frequent fusional form in the grand total of their frequencies (cell frequency).^[2]

The distribution of form proportions is depicted in Figure 1. Interestingly, while there is a peak in words preferring the short form (the far left part of Figure 1), the overall variation is quite uniform. For 439 lexemes, the short form strongly dominates (being used in 90–99.9999 % of cell instances), but for the other deciles in Figure 1, the number of lexemes is more or less similar, with a range of 140–219. In fact, there are significantly more lexemes in the highly even distribution interval, where the short form makes up 50–60 % of uses (n = 189), compared to the interval where the short form makes up 70–80 % of uses (n = 140, χ² = 7.662, p = 0.006).

Figure 1:

A histogram of parallel form usage frequency ratios in 1,908 nouns which are attested in both long (-sse) and short illative forms. Nouns are plotted based on the proportion of use of their long illative form: those preferring the long form cluster on the far right, with proportions close to 1, while those favouring the short form are found on the left, close to 0. Sections A, B and C mark the three datasets used in the present study.

Out of these 1,907 lexemes, we extracted long bias lexemes, i.e., lexemes where the long illative form makes up 80 % or more of illative singular cell frequency (n = 221, group C in Figure 1), and short form bias lexemes, following the same logic (n = 600, group A in Figure 1). We chose 80 % as the threshold, because a 90 % threshold would have rendered a very short lexeme list for the long-bias dataset (76 before manual cleaning, while the 80 % threshold yielded 221 for that type). A threshold lower than 80 %, however, would yield less meaningful results in terms of bias characterisation. For lexemes in the balanced distribution dataset (n = 201, group B in Figure 1), neither form exceeds 55 % of cell frequency. Approximately half (47 %, n = 889) of the initial dataset of 1,908 lexemes had a distribution where one form occurred in 20–45 % of cell instances, expressing neither bias nor even distribution. These were excluded from the present study.

All three datasets were then manually cleaned, which involved removing: 1) a small number of errors due to homonymous reference forms (e.g., soola-sse salt-ill and soolde intestine.ill both have sool as the nominative singular form), 2) lexemes in other languages (e.g., club, subwoofer), 3) other lemmatisation errors, and 4) compound duplicates. For instance, the long bias dataset included 50 compounds with the base noun keskus ‘centre’ (e.g., noortekeskus ‘youth centre’). If the bare base noun of a compound was found in the same dataset, all compounds based on it were removed. If the base noun was missing from data, all compounds except one (with the highest cell frequency) were removed. Hence, each of the three datasets is made up of unique base nouns, which may have duplicates in one of the other two datasets. For instance, ministeerium ‘ministry’ is included in the long form dataset while põllumajandusministeerium ‘Ministry for Agriculture’ is in the balanced distribution dataset.

Cleaning the data rendered 422 short bias, 105 long bias and 130 even distribution lexemes. We randomly extracted 130 lexemes from the short bias dataset to match it more closely to the other two datasets. Overall, illative singular frequency ratios present a rather evenly distributed picture, rather than the U- or L-shape described by Guzmán Naranjo and Bonami (2021), who found very few lexemes with balanced distribution. The three datasets used in the analyses are described in Table 1.

A mixed-effects logistic regression model (R package lme4, Bates et al. 2015) was used. Dataset type (short bias, long bias, even distribution, see Table 1) constitutes the dependent variable. All 365 items were then coded for nine predictors:

1. Cell frequency: the sum total of long and short form in the corpus.

2. Concreteness: a numerical index (0–10), where 10 indicates the most concrete lexemes. These indices originate from Aedmaa (2019), who assigned concreteness ratings to more than 200 thousand Estonian lemmas by means of machine learning.

3. Animacy: a variable with the levels human, other (non-human) animate and inanimate.

4. Pattern: this variable describes the type of short illative formative used with a lexeme. The long illative is constant (-sse) in all parallel forms in this study. Variable levels therefore only express short form variation and include “gem” (marking gemination of stem consonant, see 3b), “vowel” for stem alternation based on vowel distinction and “other” for all other stem changes.

5. Lemma syllable count: number of syllables in reference form.

6. Stress: a binary variable (word-initial vs non-initial syllable stress).

7. Maximum quantity: the highest phonological quantity in the reference form, a variable with three levels (first, second and third quantity).

8. Final phoneme: a binary variable (consonant-final vs vowel-final reference forms).

A mixed-effects model was used in order to investigate the role of inflection class, which was included as a random factor in the model as it has too many levels for it to be meaningfully included as a predictor. It expresses a lexeme’s class membership, as specified in the reference dictionary (Sõnaveeb), where 26 main inflection classes are distinguished. Classes are coded without sub-specifications (classes 9i and 9 are treated as the same class). In our implementation, the variable describes a lexeme’s inflection class pattern, rather than class membership. This means that a noun with parallel forms patterning with both classes 9 and 11 is coded as “9.11”, whereas those only belonging to class 9 are coded “9”. Hence, every membership combination is a unique factor level. Combinations constitute separate subpatterns inside larger classes, meaning that class “9” lexemes can, as a group, behave rather differently from “9.11” lexemes, as was also found for Czech by Guzmán Naranjo and Bonami (2021).

Cell frequency is included to test the hypothesis that lower frequency cells have a bias toward the long (-sse) form (Kaalep 2010). Animacy and concreteness are the only lexical-semantic predictors, chosen based on the expectation that long illative dominates when the case is selected by a verb, i.e., when the case expresses a verb complement (Metslang 2015, 2016, see 2.2). As illative is a spatial case, it is expected to have concrete, inanimate referents in its canonical usage, with an increased proportion of human and more abstract referents when marking verbal arguments (Aigro 2022; Aristar 1997). Pattern is the only morphological predictor. While patterns are unlikely to vary within a single inflection class, this predictor captures the fact that for a single cell, the same pattern can conflate different classes. Finally, morphophonology was included via four variables, all automatically coded using the Python NLP package EstNLTK (v. 1.7.1). These are used to test the assumptions of Metslang (2015, 2016, that morphophonology affects illative form choice.

3.2 Results: comparing long and short form preference

Our first model compares lexemes in the long bias and short bias datasets. We used the anova function in R to select variables, building up from a model with only the interface and the random factor (inflection class). The final model (AUC = 0.979) was the following:

Under the random factor of inflection class, only illative singular cell frequency remains a significant predictor of the type of bias (β = −0.438, p = 0.012). Long bias lexemes have less frequent illative singular tokens, and they vary less (long bias lexemes, M = 2,995, SD = 5,629, short bias lexemes, M = 10,653, SD = 33,716). The same model without the random factor of inflection class shows significant effects for all variables except animacy, meaning that long form bias is predicted by lower cell frequency, non-initial syllable stress, alternation with geminating short illative and lower syllable count, consonant-final stems, and higher concreteness values (0.0003 < p < 0.02, maximum quantity was removed as its Generalized Variance Inflation Factor was 5.8, indicating high collinearity). All in all, both of our hypotheses for RQ1 hold: there are no indications of argument status related lexical-semantic distinctions and all individual morphological and morphophonological effects are explained by inflection class membership.

One such effect overlapping with inflection classes was that of pattern. A finer-grained look at alternation patterns shows that very few patterns overlap between the two datasets. First, the only patterns occurring in both datasets are “gem”, “gem_s” and “i” (Figure 2). “Gem” and “i” are more or less evenly distributed between these datasets, because these are the categories shared by most inflection classes, which is ultimately the distinguishing factor between the two plots in Figure 2 (“gem” is shared by 6 classes and “i” by 3). In the “gem” pattern, the nominative stem consonant is geminated in non-derivational lexemes (e.g., pere family.nom > pere-sse ∼ perre family.ill ‘into the family’). While “gem_s” features in both groups, it dominates in the long bias dataset (61 % of lexemes) – here the final phoneme of -us derivations is geminated (lollus stupidity.nom > lolluse-sse ∼ lollusse stupidity.ill). Unlike these, lexemes with the “gem_ne” pattern almost all demonstrate short form bias. These, in turn, include another nominalisation group, i.e., deverbal -mine nominalisations (õppimine learning.nom: õppimise learning.gen > õppimise-sse ∼ õppimisse learning.ill).

Figure 2:

Short illative formative distribution, by lexemes with short versus long form bias. All labels represent the short form, which always alternates with long forms using the agglutinative -sse affix.

Even the “other” pattern group is made up of different formatives in the two datasets, short bias lexemes including the unproductive, archaic -h- forms in illative (pea head.nom > pähe head.ill ‘into the head’), while long bias lexemes include -te/-de forms (haare grasp.nom > haarde grasp.ill ‘into grasp’) as well as other fusional forms (hooldekodu retirement-home.nom > hooldekoju retirement-home.ill ‘into the retirement home’).

3.3 Results: comparing balanced and biased distribution

In order to compare lexemes with biased and balanced distribution, two separate models were used, one for each biased distribution datasets. Both sets of variables were determined by the same process as described in Section 3.2. The model with the best fit (AUC = 0.986) for the comparison of short form bias and balanced distribution lexemes uses the following formula:

Similarly to the model in 3.2, cell frequency is the only factor which retains its effect once inflection class is introduced into the model (β = −0.957, p < 0.001). Cell frequency is much lower and less varied for balanced lexemes (balanced distribution lexemes, M = 728, SD = 2,155; short bias lexemes, M = 10,653, SD = 33,716). In an analogous glm() model without the random factor, however, four out of eight factors significantly differ between short bias and balanced distribution lexemes, the latter being more likely among lexemes with lower cell frequency and stem-final consonants (p < 0.0001), but also with non-initial syllable stress and gemination-based short forms (0.003 < p < 0.03).

The model for comparing balanced distribution to long bias lexemes was the following (AUC = 0.828):

Balanced lexemes have significantly lower cell frequency in illative singular than long bias lexemes (β = −0.532, p < 0.001). They are also shorter in terms of syllable count (β = 0.435, p = 0.003; balanced: M = 3.5, SD = 1.4; long bias: M = 2.5, SD = 1.1). A glm() model without the random factor of inflection class shows significant effects of five variables: balanced lexemes are less frequent, less concrete, longer, more likely to be vowel-final and more likely to have stem vowel based short illatives.

Up to this point (Section 3.2–3.3), we have shown that inflection class overrides morphophonological factors, but generally does not appear to affect semantic factors, which mostly do not play a role in any case, nor cell frequency, which always plays a role. Hence, at least some of the variation in terms of which form dominates the other in Estonian illative usage is guided by cell frequency and inflection class membership. When we take a closer look at class membership proportions, however, we see that long bias and balanced lexemes align much more closely to each other in terms of inflection class than to short form bias lexemes:

As shown in Figure 3, the only overlap between short bias lexemes and the other two groups is class 17, and to a lesser extent, class 22. In all other instances, lexemes in the short bias group belong to different classes than lexemes in the other two groups. In this data, class 17 is represented mostly by two-syllable lexemes with stem consonant gemination (kivi stone.nom, kivi-sse ∼ kivvi stone.ill). Such lexemes are the only group that can have either long or short bias. Class 17 nouns with balanced distribution are mostly compounds (hence the length distinction between long bias and even distribution groups, see Section 3.3). Long bias and even distribution groups have much more in common, including the fact that both are dominated by dual class membership lexemes (9 and 11, i.e., -us nominalisations). The three main classes shared by the two (11.9, 19.2, 17) dominate both datasets, describing 72 % of the long bias and 71 % of the even distribution group, meaning the two sets of lexemes mostly originate from the same classes.

Figure 3:

Inflection class distribution in the three datasets. Classes with 1–3 lexemes are grouped under “other”.

Furthermore, Section 3.2 mentioned that while three out of four vowel-based stem types (a-final: kõrva-sse ∼ kõrva ear.ill; e-final: pilve-sse ∼ pilve cloud.ill, u-final: laulu-sse ∼ laulu song.ill) are biased towards short forms, lexemes with i-final short forms (veebruari-sse ∼ veebruari February.ill) are found in both long and short form bias datasets. This finding highlights the effect of inflection class over morphological pattern as the i-final items in the two datasets belong to different classes: long bias lexemes with an -i formative belong to classes 2, 2.22 and 2.19, while short bias pairs belong to 19, 22 and 22.23.

All in all, inflection class membership appears to be the strongest lexical predictor of bias type next to cell frequency. It groups long bias and balanced distribution datasets together, while the short bias dataset is markedly different from both. The long bias and balanced groups do not even differ in terms of alternation patterns or their proportions in the datasets, with -us nominalisations making up 57–61 %, followed by non-derivational geminations (10–17 %) and -i stem vowel forms (11–12 %).

The factors which might facilitate the occurrence of balanced distribution include cell frequency and word length. Section 3.3 showed that longer lexemes with less frequent illative forms (lower cell frequency) are more likely to end up in the balanced dataset. The balanced distribution group contains a higher proportion of compound lexemes than the long bias group (42 %  vs 12 %). However, this is not the main factor differentiating between the two groups, because, even when compounds are excluded from the data, balanced distribution lexemes show lower cell frequency (t = −3.11, df = 132.98, p = 0.002) and longer lemmas (t = 2.99, df = 137.65, p = 0.003). The effect even persists inside a single alternation pattern. For instance, gemination-based -us nominalisations (e.g., süütus innocence.nom, süütuse-sse ∼ süütusse innocence.ill.sg), make up over half of the lexemes in both datasets, but the long bias group has -us lexemes with significantly more frequent ill.sg cells (M(long) = 2,631 tokens, M(balanced) = 579 tokens; t = 2.84, df = 77.6, p = 0.006) and significantly shorter lemmas (M(long) = 2,3 syllables, M(balanced) = 3.5 syllables; t = −6.58, df = 125.25, p < 0.001). This outlines a clear, strong frequency effect. Hence, we can posit that lexemes preferring the long form may end up with even usage proportion of both available forms when they are longer and have less frequent illative singular tokens, regardless of whether they are compounds or not.

In summary, the hypothesis for RQ2 holds. Rather than being entirely distinct from short and long bias datasets, lexemes with even usage ratios between illative singular variants appear to group together with lexemes preferring the long form. Balanced distribution lexemes are, however, less frequent and longer: features which in themselves are expected to correlate in Zipfian distribution. Both long bias and balanced distribution datasets are distinct from short bias lexemes, in terms of both distinct morphophonology and lower cell frequencies.