1.1 How to infer unknown forms

Linguists typically test speakers’ productive use of lexical patterns with nonce word studies, in which speakers are asked to inflect a made-up word; since they cannot have stored associations between these words and the patterns they take, they must fall back on broader generalizations. When tested on lexically variable patterns through nonce word studies (e.g. Albright & Hayes 2003; Ernestus & Baayen 2003; Gouskova et al. 2015) and artificial language studies (e.g. Hudson Kam & Newport 2005), adults usually follow what Hayes et al. (2009) call the Law of Frequency Matching:

For example, Hayes & Londe (2006) found that Hungarian speakers obeyed this “law” when assigning suffixes with either front or back harmony (see Section 2.1) to nonce words with ambiguous harmony, like [haːdeːl], which contains one back vowel and one front vowel. Their participants showed probabilistic behavior: they were more or less likely to assign a harmony class to a noun stem depending on its vowels, and the rate at which speakers placed nouns with given vowels into each harmony class corresponded to the distribution of real noun stems with those vowels. While there are exceptions to this “law” (e.g. Pertsova 2004; Becker et al. 2011), it describes experimental results across a wide range of languages and phenomena. To date, these experiments have generally studied how speakers stochastically generate novel forms of nonce words according to their phonological characteristics.

Formal theories require a means of grammatically encoding these phonological generalizations, as they constitute part of speakers’ knowledge about the morphophonology of their language. One such method builds on a common approach to lexically arbitrary allomorphy in which words following a certain pattern are marked with diacritic features: for example, nouns like [ɡaːt] ‘dam’ that take possessive -jV would be marked with a feature [+j]. Likewise, [fok] ‘degree’ and other words that take -V would be marked with [–j]. Speakers can then extract and store generalizations over nouns that share a feature, as proposed by the sublexicon model of phonological learning (Allen & Becker 2015; Gouskova et al. 2015; Becker & Gouskova 2016) described in Section 6. Thus, the preference of words like [ɡaːt] for -jV can be expressed as a cooccurrence relation between phonological features and [+j]:

The phonological properties of a Hungarian noun are not the only source of information about its possessive, as mentioned above. Wurzel (2001) notes that a word’s form in one cell of its morphological paradigm can be predictive of its behavior in other cells. This is clearest in the case of typical inflection classes: knowing the nominative singular suffix of the Russian noun /sobák-a/ ‘dog’ allows speakers to infer that this noun has the suffix -i in the genitive singular, -e in the dative singular, and so on; all of these suffixes are characteristic of what is commonly called inflectional class II (cf. Corbett 1982: 204–5). While Ackerman et al. (2009), Ackerman & Malouf (2013), Parker & Sims (2020), and others have studied the information contained within morphological paradigms, they have not tested whether and how speakers actually use this information. Thus, Bonami & Beniamine (2016: 178) write: “It should be stressed that this paper only established that speakers are exposed to relevant information and that this information is helpful; the next step, of course, is to establish experimentally that speakers do indeed rely on [certain correlations between paradigm cells] when predicting the form of unknown words.” This paper shows that Hungarian speakers do, in fact, learn a correlation between paradigm cells (the plural and possessive) from their lexicon.

As I show in Section 6, the sublexicon model’s cooccurrence relations between features can also account for morphological dependencies. For example, if nouns like [ɟaːr] with plural -ɒk are marked with the diacritic feature [+lower], then the dependency between plural and possessive can be expressed as follows (Halle & Marantz (2008) make a similar proposal for Polish):

Thus, I propose that speakers learn phonological and morphological dependencies in the same way and apply them together when productively generating unknown forms of new words. This approach captures morphological dependencies with independently proposed theoretical mechanisms and correctly accounts for the fact that speakers weigh generalizations of different kinds against one another when choosing novel forms of words.

1.2 Road map

In Section 2, I provide a detailed background on the Hungarian plural and possessive, and Section 3 contains a formal analysis. In Section 4, I present a corpus data showing the phonological and morphological factors predicting the distribution of possessive -V and -jV in the Hungarian lexicon. In Section 5, I present a nonce word study showing that Hungarian speakers productively extend these phonological and morphological generalizations to novel forms. Section 6 describes a theory of phonological and morphological generalization based on Gouskova et al. (2015). Section 7 concludes with considerations of the theoretical import of my proposal.

2.1 Vowel harmony

As mentioned briefly in the introduction, Hungarian words have either back or front harmony, and suffix vowels alternate accordingly. The mid suffixes also show rounding harmony: front-harmonizing suffixes with mid vowels have rounded and unrounded variants to match the last vowel of the stem. These alternations, for short vowels, are shown in Table 2; see Siptár & Törkenczy (2000: 63–73) for more details. One striking asymmetry is that low and mid vowels are not differentiated for all harmony classes: the same vowel, [ɛ], is the low counterpart of mid [ø] for words with front rounded harmony, and both the low and mid vowel for words with front unrounded harmony. There is no distinct front low vowel like [æ].¹ Examples in this paper have back harmony. Table 2 can be used to find the front-harmonizing version of each suffix. Thus, the front-harmonizing equivalents of possessive -ɒ and -jɒ are -ɛ and -jɛ. Regular-stem plural -ok (from the mid vowel set) has two front-harmonizing variants, depending on rounding, -øk (e.g. [ʃyl-øk] ‘porcupines’) and -ɛk, while the lowering stem plural -ɒk (from the low vowel set) only has one front-harmonizing variant, -ɛk (e.g. [fyl-ɛk] ‘ears’). Words with front unrounded harmony can only have plural -ɛk and thus cannot be distinguished on the surface as lowering stems. Siptár & Törkenczy (2000: 225) mark some nouns with front unrounded harmony as lowering stems on the basis of other properties that correlate (more or less reliably) with lowering stem status. Since this difference is not marked in Papp (1969), which I use as a corpus (see Section 4.1) and cannot be reliably inferred, I assume that all words with front unrounded harmony are undetermined for stem class. In the nonce word experiment (Section 5), I treat stimuli with front unrounded harmony as fillers.

Table 2

Hungarian suffix vowel harmony alternations (from Siptár & Törkenczy 2000: 65).

A stem’s harmony class is usually but not always predictable from its vowels (Siptár & Törkenczy 2000; Hayes & Londe 2006; Hayes et al. 2009; Rebrus et al. 2012; 2020)—thus, at least some nouns must be explicitly marked for harmony class. However, I assume that vowel harmony is handled in the phonology proper, unlike the distinction between possessive -V and -jV: -ɒ and -ɛ are surface variants of one underlying form, while -jɒ and -jɛ are surface variants of a different underlying form.

2.2 The possessive

Table 3 shows the full paradigm of possessives for the four words in Table 1 (see Rounds 2008: 135–137). Hungarian distinguishes between the person and number of possessors, as well as the number of the possessed noun, so [dɒl-om] means ‘my song’, while [dɒl-ɒ-i-m] means ‘my songs’, and so on. There are two main points of variation among these paradigms, which I discuss in this and the next section. The first is the variable presence of [j] (bolded in Table 3) in singular nouns with 3sg and 3pl possessors and plural nouns with all possessors.² This is the possessive morpheme, with allomorphs -V and -jV. Its vowel deletes before 3pl -uk.

Under the standard syntactic analysis (cf. Bartos 1999; É. Kiss 2002; Dékány 2018), -V and -jV are realizations of a Poss head, which has a zero allomorph when adjacent to a first- or second-person possessor marker. (The marker for third singular possessors is null.) Thus, I gloss -V and -jV as poss (not 3sg), while -(V)m, -(V)d, etc. mark 1sg, 2sg, and so on.

2.3 Linking vowels

The second point of variation in the possessive paradigms in Table 3 is the alternation between [o] and [ɒ] (underlined in Table 3) in the 1sg, 2sg, and 2pl singular. This is the same lowering stem alternation described for the plural in Section 1.1. Table 4 shows these forms again for the regular stem [dɒl] ‘song’ and the lowering stem [vaːlː] ‘shoulder’, as well as forms of [kɒpu] ‘gate’. The non-possessed plural and the three possessive markers show the same pattern: the suffix is a bare consonant (or -tok) after a vowel, as in [kɒpu] (as well as the plural -i that marks the plural of possessed nouns, as shown in Table 3). Otherwise, this consonant is preceded by a “linking vowel” (cf. Siptár & Törkenczy 2000: 219), which is mid after [dɒl] and low after [vaːlː]. This suggests an analysis of lowering stems in which the different linking vowel suffixes are handled in a unified way. I present this analysis, alongside an account of possessive allomorphy, in Section 3.

3.1 Possessive allomorphy

In the introduction, I suggested that possessive allomorphy is indexed by a diacritic feature: nouns that take -jV are marked with [+j] and, analogously, nouns that take -V with [–j]. In (4), we see how these features govern allomorphy selection: by appearing in rules of realization that spell out the possessive (ignoring vowel harmony).

In the assumptions of Distributed Morphology (Halle & Marantz 1993; Embick & Marantz 2008) with root-outward vocabulary insertion (Bobaljik 2000), the noun root is inserted first. Its underlying form bears a lexical feature indexing its possessive: for example, the underlying form of [dɒl] ‘song’ is /dɒl_[–j]/.³ This [–j] feature is then visible during spellout of the possessive, matching the context in (4b). Thus, the possessive is spelled out as -ɒ, yielding [dɒl-ɒ].

Following Gouskova et al. (2015), whose framework I adopt in Section 6, I assume that both of the rules in (4) are marked with a diacritic feature; that is, there is no default rule, and every noun root must have a [±j] feature in order to get a possessive form. Rácz & Rebrus (2012), in contrast, argue that -jV is a productive default for most words (meaning that (4a) should be unmarked). Although I adopt the symmetrical, default-free analysis for theoretical reasons, this decision is borne out empirically by my experimental results, in which participants do not show evidence of default behavior, but rather use -V and -jV for the same nonce words.

3.2 Lowering stems and the plural

In Section 2.3, I showed that the plural and some possessive suffixes show a three-way distinction between -C (after vowel-final stems), -oC (after regular stems), and -ɒC (after lowering stems). In (5), I assume that these suffixes lack the linking vowel underlyingly, but are marked with a feature, [LV],⁴ indicating that they undergo linking vowel alternations:

Readjustment rules can then insert the appropriate linking vowel after consonants. In particular, lowering stems like [vaːlː] ‘shoulder’, marked with [+lower] in addition to their possessive feature (/vaːlː_[+lower,_-j]/), trigger insertion of the low linking vowel [ɒ], while most consonant-final nouns—marked with a complementary [–lower] feature, as in /dal_{[–lower,–j]}/ ‘song’—trigger insertion of the linking vowel [o].

This analysis correctly predicts that a noun that has a low linking vowel in one suffix (e.g. the plural) will have a low linking vowel in all suffixes. It also enables speakers to learn the morphological dependency between lowering stems and the possessive as a correlation between features [+lower] and [–j], as shown in (3) above (see Section 6.3).

3.3 The representation of lowering stems

The analysis in the previous section assumes that the lowering alternation is encoded morphologically: lowering stems are marked with [+lower], and suffixes with linking vowels have an [LV] feature. Siptár & Törkenczy (2000: 231–244) instead propose an abstract phonological analysis: lowering stems have a floating low feature [+open₁] and linking vowel suffixes have an underlying vowel unspecified for height. This vowel surfaces as low in the presence of [+open₁], otherwise it surfaces as mid after consonants and deletes after vowels.

These analyses represent two approaches to morphophonologically exceptional morphemes. In my analysis in (4), (5), and (6), exceptional lexical items are marked with a diacritic indexing a morpheme-specific rule or constraint (e.g. Pater 2010; Gouskova 2012; Rysling 2016). Siptár & Törkenczy (2000) instead use defective segments and subsegmental units that cannot surface faithfully and behave differently from full segments (e.g. Lightner 1965; Rubach 2013; Trommer 2021).

The two approaches are not mutually exclusive—for example, Chomsky & Halle (1968) use both—and the choice between them is often one of elegance and coverage. Moreover, both have been criticized on similar grounds: Pater (2006) and Gouskova (2012) argue that accounts with abstract underlying forms can overgenerate and be hard to learn, while Bermúdez-Otero (2012; 2013), Haugen (2016), and Caha (2021) argue that arbitrary lexical marking and readjustment rules are unrestrained and weaken our theory of grammar. In this case, the two analyses are largely equivalent: for Siptár & Törkenczy (2000), the floating feature has no phonological effect beyond producing a low linking vowel, making it akin to what Kiparsky (1982) calls “purely diacritic use of phonological features”.⁵

This paper argues that Hungarian speakers learn generalizations over [+lower], a feature marking lowering stems. For Siptár & Törkenczy (2000), the floating [+open₁] feature is similarly unique to lowering stems. Both analyses are thus compatible with my main hypothesis that speakers learn generalizations over features that index unpredictable morphophonological behavior.

With this empirical and formal background, we can turn to quantitative study of the distribution of possessive allomorphs and lowering stems and begin to test the correlation between the two.

4.1 Representing the Hungarian lexicon

In this section I discuss my representation of the Hungarian lexicon.⁶ My source of data is Papp (1969), a printed morphological dictionary of Hungarian which I digitized manually. I use Papp (1969) for its comprehensive tagging of derivational morphology, but it has potential disadvantages: it is over 50 years old and reflects lexicographic work rather than pure corpus data. In Section 4.2.3, I compare my corpus with that of Rácz & Rebrus (2012), who use a web corpus. The distribution of possessives in the two sources is quite similar, so I conclude that Papp (1969) is a relatively accurate representation of contemporary Hungarian.

Under standard assumptions in Distributed Morphology, lexical information like allomorph selection is stored for roots and affixes, not complex stems (Embick & Marantz 2008). Thus, if speakers are generalizing over the frequency of types in the lexicon (cf. Bybee 1995; 2001; Pierrehumbert 2001; Albright & Hayes 2003; Hayes & Wilson 2008; Hayes et al. 2009), derived words and compounds with the same head (rightmost affix or root) should not count as separate types. Root-based storage predicts that words ending in the same suffix should take the same possessive, which is largely true in Hungarian (Rácz & Rebrus 2012). I adopt the assumption of root-based storage by limiting my corpus to monomorphemic nouns. In Section 5.7.3, I suggest that this corpus more accurately reflects the behavior of participants in the nonce word study than a corpus including complex nouns.

Although adjectives can also take possessive suffixes, I limit my corpus to nouns. Unlike nouns, most adjectives are lowering stems (Siptár & Törkenczy 2000: 229–230), so including adjectives would complicate the relationship between lowering stems and possessive allomorphy (see also Rebrus & Szigetvári 2018). Of 35,130 nominals in my corpus, 5,055 are monomorphemic, and 4,443 of those are listed exclusively as nouns. I excluded 1,768 vowel-final words, since these categorically take -jV and would be undefined for many of the factors in my regression. I also removed the 27 words ending in orthographic h, which is phonologically complicated (Siptár & Törkenczy 2000: 274–276). Finally, I excluded 216 nouns listed in Papp (1969) with variable or unknown possessive to allow for binary coding of the possessive variable (-V vs. -jV). This leaves 2,432 noun types.

4.2 Corpus study: the distribution of -V and -jV in the lexicon

In this section, I present the results of a corpus study showing phonological and morphological factors predicting a noun’s possessive allomorph in the Hungarian lexicon. Like other cases of lexically specific variation (e.g. Hayes et al. 2009; Becker et al. 2011; Gouskova et al. 2015), the distribution of -V and -jV shows gradient tendencies; these are conditioned upon both phonological properties and the morphological dependency of stem class. In Section 5, I test whether speakers productively apply this morphological effect to new forms.

5.1 Predictions

I hypothesize that speakers form the possessives of novel words by matching the distribution of -V and -jV in the lexicon (described in Section 4) according to both phonological factors and stem class, following the Law of Frequency Matching of Hayes et al. (2009) (see Section 1.1). Since my primary concern is the morphological dependency, I focus on phonological frequency matching in the aggregate rather than individual phonological effects.

Rácz & Rebrus (2012) argue that -jV is the productive default for most words (see Section 2.2). If this is true, speakers should instead categorically assign -jV to most words.

5.2 Participants

Participants were recruited through Prolific (https://app.prolific.co/) and were born in Hungary and raised as monolingual Hungarian speakers. I recruited 30 participants for the stimulus norming study and 91 for the stimulus testing study. One participant was rejected for poor quality (see Section 5.4.2), and an additional 48 participants were recruited for earlier versions of the stimulus testing study; because the experimental task was substantially different (fill-in-the-blank rather than forced-choice, or missing the attention check for the plural), their data are not presented here.

5.3 Stimuli

I trained the UCLA Phonotactic Learner (Hayes & Wilson 2008) on the corpus of Hungarian nouns used in Section 4.1. The program produces a “sample salad” of 1,968 nonce words randomly generated from the probability distribution defined by the phonotactic grammar it learned. Many of the generated words included long strings of consonants like [ʒuːɡkfkb]. These ridiculous words reflect the limitations of the learned grammar: in particular, its constraints were limited to bigram sequences of segments or natural classes (that is, two segments/classes long) plus a word boundary, so the learner could not learn restrictions against long strings of consonants. The learned grammar was much less restrictive than actual Hungarian, so words like [ʒuːɡkfkb] were not used as stimuli. However, the grammar was good enough that many of its generated words looked like reasonable Hungarian words (like [tuːs]), and some were actually existing Hungarian words, like [iːɲ] ‘gum’.

I selected the 498 generated nonce words with the shape (C)VC(C) or (CV)CVC(C) and removed 162 disyllabic disharmonic words (which had one front vowel and one back vowel). Finally, I removed 19 generated stimuli that were headwords of any part of speech in Papp (1969).¹³ This left a final set of 317 nonce word stimuli, some of which were still, impressionistically, phonologically questionable (e.g. [ɲɒsm]). Each word was presented in the singular and plural, in the latter case with either a regular or lowering plural suffix.

5.4 Procedure

This experiment was split into two studies. First, participants rated all stimuli for plausibility as Hungarian words. The ratings obtained in this study were used to select a smaller set of stimuli for the main experiment, in which participants selected possessive forms.

5.4.1 Stimulus norming and selection

The goal of this study was to confirm the intuitions of the author (not a native Hungarian speaker) that the stimuli were broadly plausible as Hungarian words, including when presented as lowering stems. Participants completed 50 trials, each with a different stimulus. Stimuli were chosen randomly: five words never appeared at all, and the remaining 312 words were used in 1–10 trials each (mean 4.8 trials, standard deviation 2.1). An example trial is shown in Figure 1. Each trial had a frame sentence containing the target stimulus twice, presented in written form with regular Hungarian orthography (lufan, lufanok). In its first occurrence, the stimulus appeared in bare nominative form; the second time, the stimulus had a plural suffix (and sometimes subsequent suffixes). Most stimuli were shown with regular plurals (e.g. -ok), but 8 randomly chosen trials instead showed stimuli as lowering stems (e.g. with plural -ɒk).¹⁴ Participants rated each stimulus on a five-point scale according to the question Could the underlined words be Hungarian?, with the ends of the scale labelled no (1) and yes (5). Frame sentences, stimuli, and the sample trial in the original Hungarian are included in the supplemental materials.

Figure 1: Trial for Hungarian stimulus norming study, with forms annotated for harmony and stem class.

Three participants were discarded because their ratings disagreed substantially from the average ratings for a given word (an average deviation of 1.5 from the mean for stimuli that were rated by at least 3 participants). The ratings of the remaining 27 participants were used as inputs to a Python script (available in the supplemental materials) that generated potential subsets of stimuli to be used in the stimulus testing study. These sets were evaluated on three criteria: distance from a desired phonological distribution similar to that of the base corpus, a high average overall rating, and a consistent average rating for stimuli falling into each phonological category. The distribution was chosen to allow for a sizable number of trials for stimuli with each phonological property while still ensuring that the aggregate distribution of all trials can be compared to the aggregate lexicon. The goal of maximizing the average rating was to make the experiment as naturalistic as possible: while I would not expect any individual low-rated stimuli to be treated differently from high-rated stimuli, participants are supposed to pretend that the stimuli are real words of Hungarian, and too great a presence of phonotactically deviant words might make the experiment more obviously artificial. I examined high-ranked sets manually and selected a set with 81 stimuli to use for the main testing phase.

5.4.2 Morphological dependency testing

Participants each completed 35–50 trials,¹⁵ which had the format shown in Figure 2. First, the stimuli were presented in written form (lufan, lufanok, etc.) in the same frame sentences as in the stimulus norming experiment. As before, stimuli were chosen randomly: each stimulus appeared in 44–62 trials (mean 53.1, standard deviation 4.2). In Figure 2, the nonce word [lufɒn] has a regular plural -ok, but in 8–12 trials, the stimuli were presented as lowering stems, e.g. plural [lufɒn-ɒk]. As an attention check, participants had to correctly select the plural form appearing in the first sentence. Next, a second frame sentence appeared, in which participants had to select 1sg (or, for two sentences, 2sg) and possessive forms. The 1sg and 2sg suffixes have the same regular and lowering stem variants as the plural (see Section 2.3), so the linking vowel should match that of the plural: in this case, [lufɒn-om].¹⁶ The choices included both back and front variants; the possessive should have the same vowel harmony as the plural (in this case, [lufɒn-ɒ] or [lufɒn-jɒ]).

Figure 2: Trial for Hungarian stimulus testing study, with forms annotated for harmony and stem class and acceptable answers bolded.

Trials in which participants chose a discordant 1sg or antiharmonic possessive were discarded. Of 4,305 total trials, 141 had a harmony mismatch in one or both selected forms and another 565 had a stem class mismatch. Setting aside the trials with a harmony mismatch, most stem class mismatches came in lowering stem trials: in 1,003 trials in which the stimulus was presented as a lowering stem, participants only selected a lowering stem suffix for the 1sg form in 525 (52.3%). By contrast, participants matched non-lowering stems 96.2% of the time (1,890 of 1,964 trials). These results indicate that participants often resisted treating nonce words as lowering stems. Individuals varied in their matching of stem class: 10 successfully matched all lowering stem trials, while two matched none and another 11 matched 20% or fewer (the standard deviation for individual match rate was 28.7%; I compare rate rather than number of matched trials because participants saw different numbers of trials). The results presented here exclude one participant who was particularly bad (matching 44% of trials, and only matching the harmony in 62% of trials, suggesting overall lack of attention); an analysis excluding more poor performers (14 additional participants who matched in 60% or less of all trials or 20% or less of lowering stem trials) yields similar results.

5.5 Analysis

After removing discordant trials and trials of the one excluded speaker, 3,577 trials remained. Of these, 1,179 had filler stimuli with front unrounded harmony, which do not exhibit lowering stem alternations (see Section 2.3); removing these leaves 2,398 trials with 57 stimuli.

This experiment tests the hypothesis that speakers apply patterns from their lexicon in producing novel possessive forms—in particular, the morphological correlation between stem class and possessive. The clearest confirmation of the effect of stem class would be to show that the odds assigned to nonce words by the model of the lexicon in Table 5, which includes phonological factors as well as stem class (which I call phon_morph_odds), predicts the experimental results better than the odds assigned to nonce words by the lexicon model in Table 6, which includes phonological factors but not stem class (phon_odds). However, the model with phon_morph_odds performs worse than a model with phon_odds alone (χ² = 18.77). In Section 5.7.1, I show that this mismatch is, in fact, expected and that it is methodologically justified to separate stem class out from the phonological factors, as done in Section 5.6. The model presented below shows that taking the morphological factor of stem class into account leads to a better fit: participants’ choice of possessive for nonce words is influenced by whether it is presented as a lowering stem. I still represent the effect of a word’s phonology as a single aggregate factor (phon_odds), reflecting the assumption that speakers apply all of the phonological effects from the lexicon together. While this may paper over some differences between the effects of individual phonological factors, it is sufficient for this study, whose main purpose is to test the effect of stem class. (A model using individual phonological factors yields the same effect size for stem class.)

I present a mixed logistic regression model of the experimental results whose dependent variable is the possessive suffix selected in a trial (-V, coded as 0, vs. -jV, coded as 1) with random intercepts for participant and item. The model includes a fixed factor of phon_odds (the log odds that a nonce word takes -jV according to the phonological model of the lexicon in Table 6; see Section 4.2.2) and a by-participant random slope for phon_odds. It also includes a factor of stem class. I represent this as a dummy-coded categorical variable, rather than as a morph_odds factor derived from a lexicon model trained on stem class alone, for ease of presentation: given that this is a factor with only two levels (regular and lowering), all observations would have one of two values for morph_odds anyway, meaning that the two representations behave similarly. (The equivalent model using morph_odds yields identical results.) The model does not include a by-participant random intercept for stem class, as adding this factor does not significantly improve the model (χ² = 3.05, p = .383) and makes the model worse according to the Akaike Information Criterion (AIC), which penalizes model complexity. The final model is possessive ∼ phon_odds + stem class + (phon_odds | participant) + (1 | item). The model was fitted with the glmer function in version 1.1-34 of R’s lme4 package (Bates et al. 2015) using the bobyqa optimizer, which was better at finding converging models than the default Nelder_Mead optimizer.

5.6 Results

In Table 9, we see the effects of the mixed logistic regression predicting participant responses by stem class and phonology, using the phon_odds calculated from the phonological model of the lexicon in Table 6 as described in Section 4.2.2.

This model shows an overall bias towards -jV (since the intercept is positive): there were 1,031 responses of -V and 1,367 of -jV. The results also show evidence of frequency matching—that is, a correspondence between actual rates and rates predicted from a word’s phonology: the β coefficient for phon_odds is positive. However, this coefficient is smaller than 1. This means that the overall range of likelihood predicted by the experimental model is narrower than that predicted by the model of the lexicon: a change of 1 in a word’s predicted log odds of taking -jV according to its phonology corresponds to a change of only .38 in its predicted log odds in the experiment.

Even with this narrowed range, phonology (together with stem class) does not fully predict the experimental results—the marginal R² of the model Table 9, which is the proportion of the variance described by the fixed effects, is .367. The random intercept for item likewise shows that different words were given fairly divergent rates of -jV even accounting for the fixed effects—and, indeed, the conditional R², which is the proportion of the variance described by both fixed and random effects, is much higher, .556. One example of the imprecision in the model’s fit is [jøs], which has the lowest phon_odds of all the stimuli (–9.70) and is thus predicted by the phonological model of the lexicon in Table 6 to be least likely to take -jV (0.0006%). By a similar process to that described in Section 4.2.2, we can use the phon_odds to calculate the experimental likelihood of -jV predicted by the model in Table 9: 5.2% for non-lowering stem trials. In fact, participants assigned -jV to this word in 6 out of 36 non-lowering stem trials, or 16.7%, a much higher rate. The random intercept for [jøs] makes up some of this difference, adjusting the predicted likelihood of -jV in non-lowering stem trials to 10.1%.

The random intercept for participant similarly shows that different participants had different baseline rates of -jV, but the by-participant random slope for phon_odds has a low variance, suggesting that participants treated a nonce word’s phonology in similar ways. Including this random slope significantly improves the model (χ² = 15.99, p < .001).

Stimuli (e.g. [huːʃɒkː]) presented as regular stems (with plural [huːʃɒkː-ok]) were assigned -jV ([huːʃɒkː-jɒ]) 58.1% of the time (1,090 out of 1,876 trials), while participants assigned -jV to stimuli slightly less often when they were presented as lowering stems (with plural [huːʃɒkː-ɒk]), 53.1% of the time (277 of 522 trials). This difference may seem small, but it is significant in the model in Table 9. An equivalent model without stem class performs significantly worse (χ² = 9.96, p = .002). The effect sizes of the two models are otherwise very similar, suggesting that the effect of stem class is orthogonal to that of phonology—that is, that speakers are taking both phonological properties and stem class into consideration separately when providing the possessive of a nonce word, as predicted.

Figure 3 shows the relationship between the predicted likelihood (according to the phon_odds and stem class) of each nonce word taking -jV and its experimental rate of -jV. Both axes are shown in terms of log odds (that is, coefficients), making the relationship linear. Rates of 0 and 1 correspond to log odds of negative infinity and infinity, respectively—so, for example, the nonce word fátyúsz [faːcuːs], which speakers assigned -V in every trial, should be at negative infinity. It is included at the bottom edge of the graph in Figure 3. Likewise, words always assigned -jV in one condition are at the top edge of this graph. The graph shows each nonce word twice, representing trials when it was presented as a regular stem (in black), and a lowering stem (in gray). The size of each word represents the number of trials in which it appears. The lowering stem words are always smaller than the regular words because each word appeared in fewer lowering stem trials than regular stem trials. A line connects the two conditions for each word, going leftward from the regular word to the lowering stem word because the model predicts a lower likelihood of -jV for lowering stems. Figure 4 shows the same data plotted on scales of raw probability. The graphs show lines corresponding to the fit of the model in Table 9 for regular (black) and lowering stem (gray) conditions.

Figure 3: The relationship between predicted and experimental log odds of possessive -jV for individual nonce words presented with regular (black) and lowering (gray) plurals, sized according to number of trials, with a line showing the fits of the experimental model by stem class in Table 9.

Figure 4: The relationship between predicted likelihood and experimental rate of possessive -jV for individual nonce words presented with regular (black) and lowering (gray) plurals, sized according to number of trials, with a line showing the fits of the experimental model by stem class in Table 9.

The main findings of Table 9 can be seen in Figure 3. First, the log odds of -jV predicted from the lexicon map well onto the experimental log odds: the relationship is a tight linear fit. Second, the graph shows the effect of stem class: most of the lines in these figures slope downward and to the left, indicating that most nonce words had a lower experimental rate of -jV when presented as a lowering stem. Figure 4 shows that the experimental results are less extreme than the predicted likelihood, especially on the low end: nouns ending in palatals and sibilants, which nearly categorically take -V in the lexicon and thus had a near-zero predicted likelihood of -jV, were assigned -jV in the experiment up to nearly 50% of the time, or even higher in the case of kuty [kuc].¹⁷ These stimuli also show a less reliable effect of stem class than those with higher predicted rates of -jV.

5.7 Discussion

I interpret the results presented in Section 5.6 as supporting the main hypothesis of this study: participants show frequency matching effects both for phonological properties of the stimuli—as in previous nonce word studies—and for stem class. In this section, I discuss some of my analytical choices and other outstanding issues, and argue that this interpretation is empirically and methodologically justified and robust across different representations of the Hungarian lexicon.

6.1 The basic sublexicon model

The sublexicon model (Allen & Becker 2015; Gouskova et al. 2015; Becker & Gouskova 2016) encodes phonological generalizations in lexically specific variation. This allows learners to pick up on the partial phonological predictability determining a given lexical item’s choice of allomorph. As such, it follows in the path of previous models, like the Minimal Generalization Learner (Albright & Hayes 2003), that use phonological analogy to determine the set of lexical items to which a given morphophonological rule applies (see Guzmán Naranjo (2019) for an overview).

In the sublexicon model, the learner divides the lexicon into sublexicons that pattern together. These sublexicons correspond with the morphological features described in Section 3: [+lower] for lowering stems, [–lower] for other consonant-final nouns, [–j] for nouns that take possessive -V, and [+j] for nouns that take -jV. These groupings are shown below:

6.2 Sublexical phonotactic grammars

Hayes & Wilson (2008) present a model of phonotactic learning in which a learner captures generalizations over a language’s surface forms through a constraint-based phonotactic grammar. In their proposal, the learner keeps track of sounds or sequences of sounds (defined in terms of features) that are rare or absent in the lexicon and proposes constraints against them, weighting them in accordance with the strength of the generalization. For example, geminate consonants in Hungarian generally do not appear in clusters, especially within a morpheme, so the phonotactic learner trained on my corpus of monomorphemic nouns generates strong phonotactic constraints penalizing geminate consonants adjacent to other consonants: *[–syllabic,+long][–syllabic] and *[–syllabic][–syllabic,+long].²¹ This is how the speaker knows that clusters with geminates are unlikely in Hungarian.

The sublexicon model extends the notion of phonotactic learning to capture generalizations over subsets of the lexicon that pattern together—that is, sublexicons. The learner induces a phonotactic grammar for each sublexicon, capturing patterns specific to that sublexicon. As discussed in Section 4.2.3, Hungarian nouns ending in sibilants and palatals categorically take possessive -V, while nouns ending in vowels always take -jV. The sublexical grammar for the [+j] sublexicon should include heavily weighted constraints penalizing final sibilants and palatals, and the [–j] sublexicon should penalize word-final vowels.

These sublexical grammars are then reflected in speakers’ behavior. When a speaker wishes to form the possessive of a novel word, they evaluate the stem against each sublexicon’s grammar, where each sublexical grammar yields a score for that word. The better a word fares on the [+j] sublexicon relative to the [–j] sublexicon, the more likely it is to be placed into this sublexicon, and thus take -jV.

In Figure 5 and Figure 6, we see two nonce words from the experiment in Section 5, runyasz [ruɲɒs] and fúzát [fuːzaːt], tested on toy sublexical grammars with the constraints described above. Here, [ruɲɒs] is penalized by *[+strident]#, penalizing word-final sibilants, in the [+j] sublexicon, but not by the constraint against word-final vowels in the [–j] sublexicon; [fuːzaːt] accrues no penalities. I assume that all three constraints have a weight of 5.

Figure 5: Evaluation of nonce words runyasz and fúzát on the [+j] sublexical grammar.

Figure 6: Evaluation of nonce words runyasz and fúzát on the [–j] sublexical grammar.

Since [ruɲɒs] has a better score on the [–j] sublexical grammar than the [+j] sublexical grammar, it is much more likely to be placed into the former and form its possessive with -V. Specifically, this is a maximum entropy model (Hayes & Wilson 2008): a word’s likelihood of being placed into a sublexicon is proportional to its (negative) score raised to the power of e. Here, the probability of [ruɲɒs] being assigned to the [+j] sublexicon is e^–5/(e⁰+e^–5) = .0067 = .67%. On the other hand, since [fuːzaːt] has the same score on both sublexicons, it has a 50% chance of being assigned to each.

The sublexicon model is designed to capture generalizations over the phonological shape of each sublexicon’s members. In Section 5, I showed that speakers also observe a morphological generalization: lowering stems are more likely to have possessive -V. In the feature-based analysis of Section 3, this means that [+lower] and [–j] are likely to cooccur on lexical items, as stated in (3). In the next section, I extend the sublexicon model to accommodate these relations.

6.3 A sublexicon model with morphology

In my proposal, each sublexicon’s grammar has constraints penalizing diacritic features alongside those penalizing phonological features. For example, every member of the [+j] sublexicon has [+j] (by definition), but very few have [+lower], since lowering stems rarely take -jV. Since [+lower] is underrepresented in the [+j] sublexicon, the [+j] sublexical grammar should contain a heavily weighted constraint *[+lower] penalizing nouns with both [+lower] and [+j]. The [–j] sublexicon, comprising words that take -V, will also have a *[+lower] constraint, since lowering stems are also uncommon among these words. However, this constraint will not be as strong, since lowering stems are better represented among -V words than -jV words.

Figure 7 and Figure 8 show the evaluation of our two nonce words on the toy grammars, now containing *[+lower]. This constraint has a heavier weight in the [+j] grammar than in the [–j] grammar (2 and 1, respectively). Here, the speaker knows that the plurals of these words are [ruɲɒs-ɒk] and [fuːzaːt-ɒk], so she has marked both with [+lower].

Figure 7: Evaluation of nonce lowering stems runyasz and fúzát on the [+j] sublexical grammar with *[+lower].

Figure 8: Evaluation of nonce lowering stems runyasz and fúzát on the [–j] sublexical grammar with *[+lower].

The *[+lower] constraint brings the likelihood of [+j] (and possessive -jV) being assigned to /ruɲɒs_[+lower] /slightly further, from .67% to .25%. For /fuːzaːt_[+lower]/, the effect is more visible: the likelihood of [+j] goes from 50% to e^–2/(e^–1+e^–2) = .269 = 26.9%. This shows how the sublexicon model can accommodate the effects found in the nonce word experiment, both phonological and morphological: nonce words ending in sibilants are less likely to be assigned -jV (that is, be placed in the [+j] sublexicon), as are words shown as lowering stems. These effects can all be assessed in a single calculation, correctly allowing them to compound or cancel out.

The sublexicon model presented in this section has two major theoretical benefits as a tool for capturing morphological dependencies. First of all, it does not require any additional theoretical mechanisms beyond those already proposed for well-established experimental effects (phonotactic constraints penalizing underrepresented feature combinations). Second, it relies on diacritic features, a common theoretical construct (though not universally accepted; see Section 3.3), making the sublexicon model compatible with theories of morphology that commonly use diacritic features—in my case, Distributed Morphology.