The derivation of adjectives from nouns is a complex issue in Russian morphology, as these lexemes display a great deal of variation in the range of suffixes employed. Consequently, they constitute quite a typical case of affixal rivalry and therefore a good testing ground for the study of the competition between rival derivational strategies for the same syntactic and semantic function (Lindsay & Aronoff 2013, Aronoff 2016, Bonami and Thuilier 2018, among others).
The competition between adjectival suffixes is determined by a complex combination of phonological, morphological and semantic factors. The goal of this paper is to explore statistical tools in order to model suffixal rivalry in the construction of denominal adjectives. The data on which our study is performed were extracted from the National corpus of Russian language. Our dataset is composed of both the highest frequent adjectives in the corpus and hapaxes. First, using both subsets we will establish the list of properties of base nouns which are the most relevant to the choice of the adjectival suffix in each set using logistic regression. Second, we will test the effectiveness of the set of properties established for highly frequent adjectives in predicting the adjectival suffix for the low frequency subset.
- 1 The notation -Ov- indicates the variation of the vowel of this suffix, which may correspond to diff (...)
- 2 Note that only -n- and -sk- are concerned by extended variants, this is why the only one form is gi (...)
The strategies used to derive adjectives from nouns in Russian are varied. For instance, Townsend (1975) and Švedova (1980) both enumerate more than 25 suffixes which have various degrees of productivity. Three suffixes are identified as being productive in synchrony: -n-, -sk- and -Ov-1 (Zemskaja 2015, Hénault and Sakhno 2015, Kustova 2018). These suffixes may be considered as the three main adjectival suffixes, while others may be interpreted as their extended variants2 (Bobkova & Montermini 2019: 2-3), as shown in Table 1.
Table 1. Adjectival suffixes and their extended variants
-N-
|
-SK-
|
-OV-
|
-n-
-Ovn-
-ičn-
-ivn-
-on(n)-
-en(n)-
-(e)stven(n)-
-ozn-
-al'n-
-onal'n-
-arn-
-in-
|
-sk-
-esk-
-česk-
-ičesk-
-ističesk-
-ijsk-
-ansk-
-ensk-
-insk-
-istsk-
-Ovsk-
|
-Ov-
|
The variation in strategies to form adjectives in Russian may result in the emergence of doublets and even triplets, where various suffixes can be attached to the same nominal base. Table 2 shows some cases of doublets formed with the main adjectival suffixes.
Table 2. Sample of doublets / triplets in adjectives
Base noun
|
Doublets / Triplets
|
slesar’
‘locksmith’
kon’
‘horse’
|
↔ slesarnyj / slesarskij / slesarevyj
↔ konnyj / konskij
|
In the present study we will limit our investigation to the main suffixes, regardless of their variants, and, in particular, we will focus on the rivalry between -n- and -sk-, as (i) they both belong to the most productive suffixes, (ii) they constitute one of the most frequent cases of rivalry (Antipina 2012: 117-121). We will also exclude doublets from our analysis and focus on cases where only one adjective can be formed given a base noun, with either -sk- or -n-.
Both suffixes are used to form relational adjectives, but the opposition between them is not regular and is hard to formalize, since the number and types of relations between these adjectives and their head nouns may be countless (Švedova 1980: 269-279, Zemskaja 2011: 37-42). However, -sk- tends to form more relational adjectives, and -n- more qualitative ones. In this paper we will exclude the semantics of the derived adjectives as well as their relations with head nouns and we will focus exclusively on the properties of base nouns which form these adjectives.
- 3 The indications that we provide are non-exhaustive, for a larger discussion cf. Švedova (1980: 269- (...)
Recent developments in derivational morphology (Plénat 2011, Roché 2011 among others) consider that various types of constraints (phonological, morphological, semantic, pragmatic, etc.) display a complex interaction, resulting in the choice of one of the rival suffixes. Which factors correlate with the choice of the suffix for denominal adjectives in Russian? Traditionally, scholars (Švedova 1980, Zemskaja 2015, among others) provide extensive lists of the properties of base nouns that combine with each suffix. These indications are often impressionistic and exclude any quantitative evidence. For instance, in Švedova (1980: 269-284) the following information is provided on -n- and -sk-3:
- 4 This phenomenon is discussed in more details in Section 4.
-n- combines mostly with non-animate common nouns (spor ‘dispute’ → spornyj); animate nouns are also possible (inžener ‘engineer’ → inženernyj). Both abstract (gnev ‘anger’ → gnevnyj) and concrete nouns (kiparis ‘cypress’ → kiparisnyj) can form adjectives with -n-. This suffix leads to consonant mutation4 (orthographically noted k-č, g-ž, c-č, as in zvuk ‘sound’ → zvučnyj; drug ‘friend’ → družnyj; konec ‘end’ → konečnyj); it combines with both native (dym ‘smoke’ → dymnyj) and foreign stems (arxitektura ‘architecture’ → arxitekturnyj). In general, the stem of base nouns is stressed (kómnata ‘room’ → komnatnyj) but inflectional suffixes can be stressed too (zimá ‘winter’ → zimnij);
-sk- may combine with common nouns, both animate (kon’ ‘horse’ → konskij) and non-animate (universitet ‘university’ → universitetskij), as well as nouns denoting humans (brat ‘brother’ → bratskij) and proper non-human nouns (Iran ‘Iran’ → iranskij). -sk- choses stems ending with alveolars and dentals (sosed ‘neighbor’ → sosedskij, šef ‘boss’ → šefskij). Consonant mutations are also provoked by this suffix (Volga ‘Volga (river)’ → volžskij; Čexija ‘Czech Republic’ → Češskij). The stem of the base noun is generally stressed.
The choice of the exponent cannot be straightforwardly explained on phonological, morphological or semantic grounds. Both suffixes seem to impose no specific restrictions and may be combined with various nominal bases. In this paper we will explore some statistical tools to highlight the properties that may allow a clearer distinction between -n- and -sk-.
Our study is based on the National corpus of Russian language,5 a corpus of modern Russian containing over 600 million words. We automatically extracted adjectival lemmas based on their suffixes and, due to noise in the data, we performed a manual cleaning to eliminate false positives. Among false positives among others we encountered proper nouns containing the sequence -sk- (Stanislavskij) or adverbs formed with -n- (mirno ‘peacefully’). After manual verification, 4,350 adjectives were left for the study. Token frequency for each adjective was extracted as well.
The data were further divided into two subcorpora: the highest frequent lexemes (>100) and hapaxes (frequency 1 lexemes). Table 3 provides the detail of high- and low-frequency forms included in the corpus for -n- and -sk-.
Table 3. Data distribution for high- and low-frequency adjectives
|
-n-
|
-sk-
|
Total
|
High
|
550
|
436
|
986
|
Low
|
325
|
349
|
674
|
|
|
|
1,660
|
The interest of studying hapaxes was highlighted by several studies. Dal & Namer (2012) show, for instance, that very low frequency lexemes, if observed on a large scale, are likely to be good indicators of the creative use speakers make of morphological constructions, since they are less likely to have undergone phenomena of lexicalization and thus to be formally and/or semantically opaque (even though not all hapaxes are nonce formations). Hapaxes are also considered as good indicators of the productivity of affixes (Baayen 1992). Comparing this set to high-frequency lexemes may give hints on the functioning of derivational morphology independent of lexicalization and similar phenomena, especially when comparing large sets of data. Since low-frequency lexemes reflect speaker creativity, we expect to observe a mismatch between the properties of base nouns that contribute to the suffix choice, compared to high-frequency data.
Lastly, high-frequency lexemes may be considered as part of the core lexicon, i.e. the lexicon shared by the majority of native speakers. We argue that high-frequency lexemes determine the morphological competence of speakers. We will attempt to evaluate to which extent this competence can be captured by logistic regression on high-frequency data and transposed to form hapaxes. Since we expect to observe a mismatch between data distributions, we hypothesize that the model built on high-frequency data may be less accurate on hapaxes.
In order to get insights about the factors that determine the choice of the adjectival suffix, we identified the base nouns and annotated their properties. However, the identification of the base noun of a derived adjective is not always obvious. In many cases a derived adjective may be motivated by different possible nouns, formally, semantically or both.
The tendency for multiple motivation has been observed in different languages (cf. Roché 2010, Booij & Audring 2018, among others). Table 4 presents such examples of many-to-one relations in Russian.
Table 4. Derived adjectives and their potential base nouns
Base1
|
Base2
|
Adjective
|
Francija
‘France’
|
francuz
‘French’
|
francuzskij
|
vospitanie
‘education’
|
vospitatel’
‘educator’
|
vospitatel’skij
|
- 6 The first example corresponds to non-native lexical elements; the last one shows that the same unce (...)
In both cases determining the base noun without the context each adjective appears in may be confusing, even impossible, since the derived adjectives could be motivated by both nouns.6 Bonami & Thuilier (2018: 9-11) address this issue in French by collecting information about morphological families.
We chose to keep only one base for further annotations in case of ambiguity – the base which undergoes fewer formal modifications to derive an adjective (in the examples in Table 4, for instance, frančuz and vospitatel’ were kept, since they are formally closer to the radicals of derived adjectives).
After all base nouns were identified for each adjective, we proceeded with their annotation according to several formal and semantic properties.
Phonological properties include information about the features of the last phoneme of the stem (labial, dental, alveolar, velar or vowel), the length of the base noun in syllables (the only continuous property in the dataset); both of these properties are highlighted as important in prediction of the suffix by Lignon (2013) and Bonami & Thuilier (2018) for French, and by Lindsay & Aronoff (2013) for English. Stress position is also taken into consideration: both from the phonological (which syllable is stressed – ultimate, penultimate, antepenultimate) and morphological point of view (if the stress is positioned on the root of the base noun, or on the derivational or inflectional suffix, if any).
Morpho-phonological phenomena typical of Russian derivation such as vowel / Ø alternation, palatalization and consonant mutation were annotated as well. Morphological properties include the inflectional class of base nouns. Semantic properties include the features [±proper], [±human], [+animate], [±concrete] in their different combinations which form different degrees of animacy; etymological properties express whether the base noun is a native Slavic word or a loanword.
In order to identify correctly the stem of each noun we should consider one of the peculiarities of the Russian language, i.e. its extensive use of inflectional morphology: the surface forms of a given lexeme are rarely equal to its stem(s).
1Consider the example in Table 5.
Table 5. Example of surface forms of Russian nouns
Lexeme
Stem
|
stena ‘wall’
sten
|
Surface forms sg
Surface forms pl
|
stenáNOM, stenýGEN, stenéDAT, sténuACC, stenójINSTR, stenéLOC
stényNOM, stenGEN, sténamDAT, stényACC, stenámiINSTR, stenáxLOC
|
- 7 The cases shown in Table 5, where graphical form is combined with stress position, represent the ex (...)
There may be up to 12 forms for Russian nouns, corresponding to 6 declensions, in the singular and plural. Since stress position is not fixed, in some cases it makes the only phonological difference between two surface forms (cf. stenýGEN.SG vs. stényNOM.PL).7
In order to perform a consistent analysis of nominal stems, we chose to keep surface forms corresponding to the nom.sg for annotation.
In cases of multiple surface forms like those listed in Table 5 it is nevertheless possible to distinguish one common stem (sten). However, for other nouns there may be stem allomorphs.
- 8 All the allomorphies in question cannot be explained by synchronically active phonological processe (...)
- 9 Long and short stems form oppositions in the inflectional paradigm, generally, nom.sg. & acc.sg or (...)
A first case of allomorphy concerns the alternation between Ø and a full vowel (fleeting vowel). This is particularly sensitive, since we consider base length in syllables in our list of properties. Vowel / Ø alternation is encountered in nouns, both in inflection and in derivation (Sims 2017: 491-497). In this case the lexical representation of a lexeme includes two stems. Note that short (vowelless) and long (with the vowel surfacing) stems correspond to different portions of the inflectional paradigm,9 as shown in Table 6.
Table 6. Vowel / Ø alternation in nominal system
Noun
|
Stem1
|
Stem2
|
Adjective
|
ugol’
‘coal’
grivna
‘hryvna (currency)’
|
ugol’
nom.sg, acc.sg
griven
gen.pl.
|
ugl’
all others
grivn
all others
|
ugol’nyj
grivennyj
|
We kept stems corresponding to the nom.sg. and introduced a distinct annotation corresponding to vowel / Ø alternation (1 if there are two stems; 0 if there is only one stem).
- 10 We provide more details on inflectional classes in Russian in the following subsections.
The second allomorphy concerns a phenomenon of palatalization (iotation or softness) which occurs for consonants appearing before front vowels (Timberlake 2004: 482). In synchrony it results in stem alternations. Examples concerning the nominal system are shown in Table 7. Note that the oppositions in inflectional paradigms are different with respect to vowel / Ø alternations, since front vowels appear in dat.sg. and loc.sg for nouns of the first inflectional class and in loc.sg – of the second inflectional class.10 If the surface form corresponding to the form of the lexeme (nom.sg) is not affected by palatalization, the latter may occur in other forms of the paradigm. Palatalization affects the annotation of the last phoneme of the stem.
Table 7. Palatalization in nominal system
Noun
|
Stem1
|
Stem2
|
Adjective
|
škola
‘school’
gora
‘mountain’
|
škol
all others
gor
all others
|
škol’
dat.sg, loc.sg
gor’
loc.sg.
|
škol’nyj
gorskij
|
As in the case of fleeting vowels, stems corresponding to nom.sg. were kept and palatalization was annotated separately (1 if there is palatalization; 0 if it does not occur).
Another semi-productive morpho-phonological alternation concerns velar mutations (cf. Sims 2017: 497). Unlike vowel / Ø alternation or palatalization, it only surfaces in noun derivation. The presence of a mutated stem is thus lexically determined, and also dependent from the derivational affix (Kapatsinski 2010, Sims 2017). We consider that in this case lexemes are stored in the lexicon with two stems. Since this modification is limited to derivation, we label the mutated stem ‘StemD’ (derivational stem, as opposed to ‘StemI’ – the inflectional stem), as illustrated in Table 8.
- 11 Note that the last two cases combine both palatalization and mutation. For the second case, mutated (...)
Table 8. Mutation in nominal system11
Noun
|
StemI
|
StemD
|
Adjective
|
ruka ‘hand’
uxo ‘ear’
zamok ‘lock’
mladenec ‘baby’
|
ruk
ux
uš
zamok
zamk
mladenec
mladenc
|
ruč
uš
zamoč
mladenč
|
ručnoj
ušnoj
zamočnij
mladenčeskij
|
For consistency, stems corresponding to the nom.sg. were kept; mutation was annotated separately (0: no mutation; 1: mutation occurs).
A morphological property that we explore in this study concerns inflectional paradigms. A canonical distinction is made in Russian between three inflectional classes: m and f nouns ending in -a/-ja in the nom.sg. for the first class; m and n nouns ending in consonant (palatalized or not) for the second; f nouns ending with a palatalized consonant for the third. Nevertheless, the assignment of inflectional classes to Russian nouns is not a trivial task, since a wide range of irregularities can be observed. There are different approaches to describe the inflectional system of Russian, with the number of inflectional classes varying accordingly and ranging from a partition of the three main classes into subclasses (Fraser & Corbett 1995: 132-137) to a distinction between 50 inflectional paradigms (Zaliznjak 2003) or even more (Sims & Parker 2015). Our goal is, however, less theoretical and more methodological, we follow the canonical distinction of inflectional classes and annotate the data according to the three inflectional classes mentioned above.
As for semantics, in order to be consistent with the annotation adopted for this study, which is mainly categorical, we annotated binary semantic properties: [±proper], [±human], [±animate], [±countable] and [±concrete]. Our definition of what a concrete noun is is relatively restrictive: only prototypically concrete nouns, i.e. referring to entities that can be clearly perceived with (at least) one of the five senses were annotated as [+concrete]. Based on this annotation, and following Thuilier (2011: 291), we grouped these properties into five classes of Animacy. Animacy is defined by Thuilier (2011) as a semantic property of entities and can generally be designed as a hierarchy or scale. The simplest distinction is made between [±human] nouns, a more elaborate one involves [±animate] nouns. For this study we present our own hierarchy as shown in Figure 1, where the five classes proper non-human; proper human; common human; common concrete; common abstract emerge.
Figure 1. Animacy and its subclasses
In what follows we will proceed with an exploratory data analysis.
Table 9 shows that -n- tends to choose 2-syllable stems more often than 3-syllable ones; -sk- combines almost equally with both, although the preferences for -n- are statistically non-significant in low frequencies (χ2=57.30; p<0.001 for high frequencies and χ2=1.75; p<1 for low frequencies).
Table 9. Data distribution by the length of base in syllables
|
|
1
|
2
|
3
|
4
|
5
|
High
|
-n-
|
85
|
263
|
144
|
35
|
19
|
|
-sk-
|
33
|
146
|
154
|
77
|
24
|
Low
|
-n-
|
24
|
126
|
119
|
46
|
8
|
|
-sk-
|
23
|
125
|
138
|
52
|
10
|
Table 10 shows some trends for the last phoneme of the stem: dental- (and to a lesser extent alveolar- and labial-)ending stems are followed by -n-; alveolar-ending stems (with vowel-ending and dental-ending stems following) occur more often with -sk-. Distributions are smoother in low frequency data: dental- and alveolar-ending stems give rise to both -n- and -sk- derivatives almost equally (high: χ2=120.72; p<0.001, low: χ2=47.96; p<0.001).
Table 10. Data distribution by the last phoneme of the stem
|
|
Vow
|
Alv
|
Dent
|
Lab
|
Velar
|
High
|
-n-
|
32
|
148
|
215
|
89
|
66
|
|
-sk-
|
89
|
199
|
74
|
45
|
29
|
Low
|
-n-
|
21
|
91
|
102
|
39
|
37
|
|
-sk-
|
23
|
125
|
138
|
52
|
10
|
According to morphological stress position (StressMorph) (Table 11), both suffixes combine in majority with underived nouns where the root (R) is stressed, -sk- favors base nouns with a stressed inflectional suffix (InfS) more often than -n-. If the base noun is already derived, -n- preferentially selects nouns with a stressed derivational suffix (DerS). In low frequencies, both suffixes choose nouns with stressed inflection more often than with stressed derivational suffixes (χ2=40.28; p<0.001 for high frequencies; χ2=2.66; p<0.5 – for low). As for phonological stress position (StressPhon), distributions are quite similar (d: last syllable stressed; ad: penultimate syllable; aad: antepenultimate syllable) (χ2=24.09; p<0.001 for high frequencies; χ2=0.50; p<1 – for low).
Table 11. Data distribution by the stress position
|
|
StressMorph
|
StressPhon
|
|
|
|
R
|
DerS
|
InfS
|
aad
|
ad
|
d
|
High
|
-n-
|
511
|
28
|
11
|
47
|
214
|
288
|
|
-sk-
|
374
|
11
|
51
|
81
|
162
|
190
|
Low
|
-n-
|
286
|
12
|
27
|
29
|
145
|
148
|
|
-sk-
|
299
|
9
|
41
|
26
|
143
|
179
|
The distribution of data by stem alternations (Table 12) indicates that there are no clear preferences in the suffix choice for palatalized stems (χ2=1.17; p<0.5 for high frequencies; χ2=0.24; p<1 – for low); however, stems with vowel alternations (χ2=18.45; p<0.001 for high frequencies; χ2=26.50; p<0.001 – for low) as well as with consonant mutations (high: χ2=46.79; p<0.001; low: χ2=51.11; p<0.001) seem to combine more frequently with -n-.
Table 12. Data distribution by allomorphies
|
|
Vowel / Ø
|
Palatalization
|
Mutation
|
|
|
0
|
1
|
0
|
1
|
0
|
1
|
High
|
-n-
|
512
|
38
|
462
|
88
|
|
473
|
77
|
|
-sk-
|
427
|
9
|
361
|
75
|
|
431
|
5
|
Low
|
-n-
|
287
|
47
|
296
|
29
|
|
264
|
61
|
|
-sk-
|
344
|
5
|
310
|
39
|
|
347
|
2
|
Table 13 suggests that -sk- is reluctant to combine with nouns of the 3rd inflectional class. Note that -sk- disfavors more the 1st inflectional class in low frequencies, comparing to high-frequency data (χ2=36.67; p<0.001 for high-frequency and χ2=32.77; p<0.001 for low-frequencies).
Table 13. Data distribution by inflectional class
|
|
1
|
2
|
3
|
High
|
-n-
|
139
|
342
|
69
|
|
-sk-
|
127
|
296
|
9
|
Low
|
-n-
|
133
|
172
|
18
|
|
-sk-
|
62
|
246
|
6
|
As for the origin of the stem (Table 14), -n- tends to combine with native bases, whereas -sk- slightly favors loanwords. In highly frequent lexemes the nouns formed with suffixes of native origin, in case their nominal base is derived, are the most frequent (voditel’ ‘driver’, čitatel’ ‘reader’, stoimost’, ‘price’). Foreign phonological material is encountered to a lesser extent (politika ‘politics’, protestant ‘protestant; protester’). In low frequencies we encounter nouns formed with a suffix of foreign origin more often, with respect to the nouns formed with suffixes of Slavic origin. Note the inverse tendencies for -n- in low frequency lexemes (high frequencies: χ2=37.20; p<0.001, low: χ2=30.30; p<0.001).
Table 14. Data distribution by etymological factor
|
|
loanwords
|
native
|
High
|
-n-
|
172
|
378
|
|
-sk-
|
227
|
209
|
Low
|
-n-
|
168
|
157
|
|
-sk-
|
254
|
95
|
- 12 The [+proper, +human] subclass was dropped out due to its low representation in both data sets.
Table 15 displays the distribution by animacy.12 Some clear tendencies emerge: -n- selects the [+common, ±concrete] (ComConc, ComAbst) animacy subclasses, whereas -sk- prefers [+common, +human] (ComHum) and [–common, –human] (PropNHum) (high: χ2=688.62; p<0.001, low: χ2=312.77; p<0.001). As for distributions by individual semantic properties, all of them are relevant to the suffix choice in high-frequency data, however in low-frequency data [+concrete] is not statistically significant (χ2=0.15; p<1).
Table 15. Data distribution by animacy
|
|
Com Hum
|
Com Conc
|
Prop NHum
|
Com Abst
|
High
|
-n-
|
19
|
135
|
1
|
395
|
|
-sk-
|
163
|
22
|
189
|
53
|
Low
|
-n-
|
17
|
142
|
3
|
160
|
|
-sk-
|
140
|
33
|
87
|
24
|
As expected, the distributions between high- and low-frequency data are not always equal. This may be explained by the evolution of language that makes new tendencies emerge. Consequently, generalizations made on high-frequency data may not work well on low frequencies.
Based on statistical significance, animacy has the strongest correlation to the choice of the suffix in both sets; the last phoneme of the stem, origin and inflectional class are also important. Base length, phonological and morphological stress position are relevant in high-frequency data but are not significant for low frequencies and thus will be dropped in the following analysis. As for palatalization, it is non-significant for both data sets and will be dropped as well.
In what follows, we will investigate if the data contain strongly correlated features among the predictors.
Multicollinearity is a problem that arises when some predictor variables (base noun properties) in the model measure the same phenomenon. The statistical significance of a variable may be undermined: the more this variable is correlated to others, the larger its standard error is. This, in turn, leads to unreliable and unstable regression coefficients, which may become statistically non-significant in regression model. Multicollinearity should be assessed in order to draw accurate conclusions about the significance of base noun properties. To test both data sets for multicollinearity, we computed pairwise correlations of variables using the Pearson correlation coefficient. Multicollinearity heatmaps for high- and low-frequency data are shown in Figure 2 and Figure 3 (with correlation coefficients ranging from -1 to 1).
Figure 2. Multicollinearity plot for high frequencies
Figure 3. Multicollinearity plot for low frequencies
Both data sets display the same tendencies: binary semantic properties strongly correlate with subclasses of animacy (correlation coefficients ranging from 0.52 to 1 in both sets); consonant mutations – with velar-ending stems (0.78 for high-frequency data and 0.73 for low-frequency ones).
In order to be able to interpret the final results we will drop highly correlated variables from both data sets: binary semantic features, as well as consonant mutation. Our final data sets have the following constitution:
- High-frequency data: source, animacy, phonological stress position, morphological stress position, length of base noun in syllables, inflectional class, vowel / Ø alternation, last phoneme of the stem;
- Low-frequency data: source, animacy, inflectional class, vowel / Ø alternation, last phoneme of the stem.
There are more base noun properties for high-frequency adjectives, compared to low-frequency ones. We hypothesize that since hapaxes reflect the creativity of speakers, the properties of base nouns become less restrictive for suffix choice.
As mentioned at the end of section 5 there are several properties of base nouns which may be correlated to the choice of the suffix. A multifactorial statistical analysis is needed to take into account these properties simultaneously and estimate if each of them is still relevant. We will use logistic regression for this purpose. This statistical model finds the probability of a dependent variable (the suffix) given a set of independent variables (the properties of base nouns). The main advantages of a logistic regression model are its clarity and interpretability of results. Consequently, it avoids the ‘black box’ problem frequently associated with classification methods based on machine and deep learning tools.
First, we will build a model on high-frequency data, then – a model for low-frequency adjectives. Finally, we will assess to which extent the model built on high-frequency data can predict the suffix for low-frequency adjectives.
Table 16 shows the output of logistic regression model for high-frequency data with p-values and regression coefficients. For illustration purposes, only variables with p-value<0.05 were kept in this table. All the constituents of animacy, the last phoneme of the stem (velar, dental) and inflectional class 3 are statistically significant for the choice of the suffix; to a lesser extent – labial ending stems and stems with stressed derivational suffix. The effects of other predictors are negligible.
Table 16. Logistic regression coefficients for high frequency data (sorted by p>|z|, where p<0.05)
|
coeff
|
std err
|
z
|
P>|z|
|
Intercept
|
-1.9012
|
0.600
|
-3.167
|
0.002
|
Anim[ComConc]
|
-1.4277
|
0.339
|
-4.207
|
0.000
|
Anim[PropNHum]
|
2.2903
|
0.309
|
7.410
|
0.000
|
Anim[ComAbstr]
|
-5.1506
|
0.855
|
-6.021
|
0.000
|
LastPhon[Velar]
|
2.3867
|
0.283
|
8.435
|
0.000
|
LastPhon[Dental]
|
1.9906
|
0.618
|
3.223
|
0.001
|
InflClass[3]
|
1.3312
|
0.449
|
2.964
|
0.003
|
LastPho[Labial]
|
2.3016
|
0.950
|
2.423
|
0.015
|
StressPos[derSuf]
|
1.0383
|
0.486
|
2.138
|
0.033
|
Note that negative coefficients are associated with probability of the suffix -sk- and positive ones – with probability of -n-. This means that the model considers [+common, ±concrete] to be strongly correlated to -n-; all other subclasses of animacy, conversely, are strongly correlated to -sk-. In general, all other predictors (labial-ending stems and stressed derivational suffix) favor -n- over -sk-.
To assess the predictive power of this model we performed a ten-fold cross-validation. This technique splits data into ten smaller sets, nine of which are used to train the model, and the tenth to validate it. This process is repeated several times for the model to be trained and tested on all ten parts. Therefore, performance, or accuracy, corresponds to the mean of all the accuracies obtained in the loop. The detailed values of the ROC curve (which measures true positive and false positive rates) is shown at Figure 4. This model gives very accurate predictions (AUC=0.90).
Figure 4. ROC on high-frequency data (with 10-fold cross validation)
We repeated the same process for low-frequency data. Fewer properties of base nouns were left for modelling compared to high frequencies. The output of the logistic regression model shown in Table 17 (which also displays only variables with p<0.05) suggests that only subclasses of animacy are statistically significant to chose between -n- and -sk- in hapax formations. All other properties become non-relevant.
Table 17. Logistic regression coefficients for low frequency data (sorted by p>|z|, where p<0.05)
|
coeff
|
std err
|
z
|
P>|z|
|
Intercept
|
-0.4593
|
0.524
|
-0.877
|
0.380
|
Animacy[ComConc]
|
-1.6180
|
0.324
|
-4.996
|
0.000
|
Animacy[PropNHum]
|
1.7247
|
0.261
|
6.603
|
0.000
|
Animacy[ComAbst]
|
-2.6784
|
0.516
|
-5.191
|
0.000
|
Animacy is thus the statistically most significant property in both data sets. The last phoneme of the stem, stress position and inflectional class, important for suffix prediction in high-frequency data, are irrelevant for hapaxes. These results are in line with our observations and our hypothesis that hapaxes reflect morphological creativity and may thus violate principles which operate on high frequency data. Base noun properties which are relevant for the suffix choice in highly frequent lexemes become less restrictive for hapaxes.
Figure 5 shows the performance of this model. Despite the fact that less properties are needed to predict suffix, the model makes also very accurate predictions (AUC=0.89).
Figure 5. ROC on low-frequency data (with 10-fold cross validation)
- 13 The properties which were previously dropped in low-frequency data due to their statistical non-sig (...)
The last model that we want to assess is the model trained on high-frequency data and tested on the low-frequency dataset.13 Our initial assumption was that highly frequent words shape the morphological competence of native speakers and thus should influence the formation of new words. However, nominal bases display different tendencies in high and low frequencies. How well the model built on the high-frequency data set can generalize to data which coming from a different distribution?
Once again, we will use a ROC curve to assess the performances of the model. Applying the high-frequency model to low-frequency data results in AUC=0.86 which is indeed a very accurate performance, although it is slightly worse with respect to cross-validation on both the data sets assessed in previous sections. Even if data do not come from the same distribution, the properties of base nouns learnt from high frequencies may be applied to predict the suffix choice for hapaxes. This result may support our claim that highly frequent words determine the linguistic behavior of speakers. Even if morphological creativity may take place in the formation od new words and violate some constraints which operate for high frequencies (the last phoneme of the stem, stress position and inflectional class), some principles are preserved in low-frequency data (animacy of base nouns).
Figure 6. Logistic regression ROC curve
The choice of the suffix in Russian denominal adjectives depends on various properties of noun bases, and these properties are different in highly frequent corpus compared to low-frequency words.
The first conclusion of this study is that fewer properties of base nouns determine the suffix choice for hapaxes. In high-frequency data set the semantic properties of the base seem to be the best predictors for the choice between -n- and -sk-; the final consonant clusters also display a high degree of relevancy; inflectional class and stress position influence the choice of the suffix to a lesser extent. As for low frequencies, only semantic properties of base nouns are statistically significant. The creativity of speakers in hapax formation may result in less restrictions imposed by base nouns to the suffix choice.
The second conclusion we make is that the evidence taken from high-frequency data set may be applied to low frequencies. Training models on high-frequency data and testing them on low frequencies brings satisfactory results, even if both sets do not have the same distributions. Even if here is a mismatch between high-frequency and low-frequency data, and less base noun constraints operate in low frequencies, animacy seems to be a very good predictor in both sets. Overall, the knowledge learnt on high-frequency lexemes may be applied to low-frequency data with good accuracy.
In this study we focused only on two suffixes, -n- and -sk-. Further on, we would like to perform the same investigation on other suffixes and include the rivalry between -n- and -Ov-, between sk- and -Ov-. As presented at the beginning of the article, the Russian language possesses extended variants of the main suffixes which can also be studied in terms of rivalry.
In order to address the problem of doublets multilabel logistic regression can be used. The study of doublets would not be complete without taking into consideration the frequency of adjectives and their distributions between different types of subcorpora (general, oral, newspapers, among others).
Lastly, the inclusion of the semantics of adjectives with rival suffixes the distributional analysis could be an interesting extension of our study (Wauquier et al., 2020).