Predicting native speaker choice: the role of corpus-based frequency metrics in morpho-syntactic alternations

2.1 Which metrics are the cognitively most realistic?

The issue of which frequency metric is the cognitively most realistic one has received increasing attention in corpus-based cognitive linguistics (Blumenthal-Dramé 2016b; McConnell and Blumenthal-Dramé 2022). In these studies, the benchmark for cognitive realism has been the extent to which a metric predicts behavioural data in language processing experiments. In line with these studies, we see cognitive realism a relative notion, and as an ideal that can only be approximated: A metric cannot be cognitively real in an absolute sense, but it can outperform competing metrics in terms of cognitive realism, and this can be tested by comparing how well competing metrics align with experimental data. Moreover, any winning metric can only be realistic in the rather loose sense of imposing restrictions on the neural and psychological levels at which the human mind can be modelled (for a more in-depth discussion, see Blumenthal-Dramé 2016b).

Indeed, a vast panoply of competing metrics is available in the literature, with different primary uses in mind. For example, metrics inspired by information theory (like entropy reduction or surprisal) have served as the gold standard in psycho- and neurolinguistics (Frank et al. 2015; Hale 2016; Lowder et al. 2018; Shetreet et al. 2016; Smith and Levy 2013), different types of association scores between words (e.g. LogLikelihood, Dice, T-Score) have been widely used in corpus-linguistics (McConnell and Blumenthal-Dramé 2022), and still other researchers have strongly relied on raw frequencies (Arnon and Christiansen 2017; Arnon and Cohen Priva 2013). While all of those metrics have been insightful in different ways, most of them have not been designed to model cognition; their cognitive realism should therefore not be taken for granted. For example, corpus-derived association scores have first and foremost been developed for the purposes of computational linguistics, lexicography, and applied linguistics (e.g. for corpus-based dictionaries, translation resources, or automated language processing systems), and Evert (2009) rightly cautions that they merely capture “a statistical attraction between certain events and must not be confused with psychological association”. Brysbaert et al. (2018) point out that word knowledge and word frequency are not directly related; they introduce word prevalence measure instead.

On the other hand, corpus-derived metrics inherently capture collective behavioural patterns; it therefore stands to reason that they must, however indirectly, be related to cognition. It does not come as a surprise that many empirical studies have identified significant correlations between corpus-derived association measures and different aspects of language processing (for review, see Blumenthal-Dramé 2016a). However, as cautioned by several researchers, the fact that collective patterns are describable in terms of some statistical generalisation does not entail that the individuals giving rise to these patterns unconsciously apply or follow this specific generalisation (cf. Frank and Bod 2011; Kronenfeld 2006; Perruchet and Peereman 2004: 116). This clearly calls for studies adjudicating between competing corpus-based metrics with the aim of identifying the cognitively most insightful one.

2.2 Different levels, different metrics?

The second open question at the interface of corpus and cognitive research concerns the range of linguistic levels at which established metrics can successfully predict aspects of language processing. Most cognitive linguistic research so far has looked into attraction strengths between two or more content items, or between content items and abstract grammatical constructions (Blumenthal-Dramé 2016a; Gries 2012; Gries et al. 2005). By contrast, much less research has focused on attraction strengths between content and grammatical items (like inflectional suffixes or adpositions). This is conspicuous given the fact that cognitive linguists generally subscribe to the tenet that language is “constructions all the way down” (Goldberg 2006: 18). This widely cited slogan captures the view that constructions at different levels of language description from pragmatics down to morphology are not qualitatively different but vary only in terms of complexity (in terms of syntagmatic length) and abstractness (i.e. the number of different realisations that a constructional schema can subsume). A strict interpretation of this claim suggests that it should also be “the same metrics all the way down”. In other words, metrics that have proven to be predictive in collocational or collostructional research should also work at the morphological level or between content and function words. Conversely, if it was found that the same metrics are not equally insightful across different levels of language, this would arguably weaken the cognitive plausibility of the constructionist uniformity claim. Also note that the it’s-the-same-metrics-all-the-way-down hypothesis that we are putting forward is less trivial, or more contentious, than might seem at first glance. Recent research suggests that humans do not possess one unified statistical learning and processing ability; their statistical sensitivity rather varies as a function of modality (e.g. auditory vs. visual signals) and information domain (e.g. musical vs. linguistic acoustic sequences) (Conway and Christiansen 2009; Frost et al. 2015; Siegelman 2020; Siegelman et al. 2017; Thiessen 2017). Against the light of these findings, it makes sense to ask whether the same statistics apply at least across different levels of (one) language.

2.3 Different languages, different metrics?

Related to the question of whether different levels of language are affected by different frequency metrics is the third question of whether the same corpus metrics are equally cognitively realistic across different languages. Over the last years, an increasing number of studies have suggested that language users’ sensitivity to statistics might be modulated by typological features of their language and its spelling system (Potter et al. 2017; Saksida et al. 2017). This could mean that statistical metrics are computed over different units, that different statistical metrics are used, or that different thresholds are set depending on the language under consideration. This highlights that it would be fruitful to explore whether metrics that are well-established for quantifying attraction strength in English can and should be extended to typologically different languages without further qualification.

2.4 The epistemological status of frequency

The fourth and final frequency-related question that this research seeks to address is the epistemological status of frequency-based cognitive explanations. More specifically, to what extent does frequency in itself have a driving influence on language processing, and to what extent should it rather be considered as a spurious by-product of other underlying variables? This question is notoriously difficult to answer, but given that the usage-based paradigm understands itself as an offshoot of American functionalism (Bybee 2010: 195) that is officially committed to explaining “aspects of grammar through reference to meaning and discourse function” (Bybee 2010: 10), the idea that non-frequency related predictors might be the actual driving forces of language processing should not be discarded too quickly (Baayen 2010; Bley-Vroman 2002; Caldwell-Harris 2021; Jolsvai et al. 2020; Newmeyer 2003). Bybee (2007) has stressed that the complex question of whether frequency is a cause or an effect can be answered in both ways. As an observable pattern in text, it is an effect, at the same time “repetition of experiences has an impact on cognitive representations and in this way becomes a cause” (Bybee 2007: 17–18).

One way of achieving a more nuanced understanding of the relation between frequency and other, well-established functional predictors is to gauge their respective roles when combined into the same statistical model. Do they account for the same share of variance (in which either frequency or the other predictors could rightly be deemed epiphenomenal), do they show independent effects (in which case frequency could be assigned the role of one functional predictor among others), or does frequency supersede the other predictors (in which case the share of variance explained by the other predictors would be a proper subset of the share of variance accounted for by frequency, which would point to frequency being the actual driving force)? We will investigate these questions by taking a morphosyntactic alternation in Estonian as our case study.

4.1 Forced choice task

4.1.1 Materials

The experiment consisted of 60 corpus sentences with a blank for the original construction followed by the two constructional alternatives. There were 30 experimental items and 30 filler items in the experiment. The list of filler items were alternations of a different type, e.g. pro-drop alternation oletan ∼ ma oletan “I assume”. The experimental items were randomly sampled from five probability bins defined by the logistic regression model fitted by Klavan (2012: 176–181). The sentences therefore represent the full probability scale – the stimuli ranged from sentences where one construction was very probable (near-categorical preferences) to sentences where both constructions were equally probable (approximately equal probability estimates for both choices). Six sentences were randomly sampled from each probability bin, adding up to a total of 30 (5 × 6) experimental items.

We decided to sample items across the corpus model predicted probabilities rather than the actual frequency measures used in our study because we did not know which frequency metric would turn out to be the winning metric. We did not want to prejudge the analysis by sampling our data according to only one metric which might have reduced variability according to the other metrics. This study is part of a larger scope research which was not only focused on frequency metrics. Using Spearman’s correlation coefficients, we looked at the relationship between the different frequency metrics and the corpus probabilities among the experimental stimuli (see Appendix A for an overview of the results). These findings indicate that frequency and probability are associated.

Figure 1 presents the item probabilities estimated by the original logistic regression corpus model. For each sentence an alternative paraphrase was constructed for the original construction and both alternatives were presented together with the original sentence context. Items were pseudo-randomized so that no two sentences from the same probability bin followed each other. There were four versions of the questionnaire to diminish potential order effects. The order of the construction choices was alternated between the versions. For example, in versions 1 and 3 the peal construction was given first, in versions 2 and 4 the adessive construction was presented first for the same item.

Figure 1:

Estimated probabilities for experimental items based on the binary regression model fitted in Klavan (2012).

4.2 Annotation: semantic and morphosyntactic variables

The 30 experimental items were annotated for 5 variables based on the results of the previous studies on the adessive and peal alternation in Estonian (Klavan 2012; Klavan 2020; Klavan and Veismann 2017; Klavan et al. 2015): complexity, length, mobility and the syntactic function of the Landmark phrase and the word class of the Trajector. Following is a very brief overview of these variables; for more details on the annotation scheme, see Klavan (2012). Appendix B lists all the experimental items together with the annotated categories.

The variable COMPLEXITY takes the reference value “simple” if the Landmark lemma is not a compound (e.g. laud ‘table’) and the value “compound” if it is (e.g. kirjutuslaud ‘writing desk’). LENGTH is the number of syllables in the Landmark phrase (log-transformed). The variable MOBILITY has two levels: mobile and static. Following de Vega et al. (2002), mobile Landmarks are those that do not have a fixed position in the environment, either because they move by themselves (e.g. humans, animals) or can be moved by an external agent (e.g. a table, a car). Static Landmarks have a fixed position in the environment (e.g. street, market).

The variable SYNTACTIC FUNCTION denotes the grammatical function of the Landmark phrase: adverbial means that the Landmark is a sentence constituent in its own right, while modifier means that the Landmark modifies the Trajector within the same constituent. For example, both the adessive and the postpositional construction with peal can fulfil two syntactic functions in a clause – that of an adverbial, as Vaas on laua peal/laual ‘The vase is on the table’, or an adverbial modifier as vaas laual/laua peal ‘the vase on the table’ in Vaas laual/laua peal on ilus ‘The vase on the table is pretty”. The variable has two levels: ‘adverbial’ and ‘modifier’.

The word class of Trajector (TRWC) has two levels – it can either be a noun or it can belong to other word classes (verb phrases and pronouns). For example, in the sentence Vaas on laua peal ‘The vase is on the table’, the Trajector is vaas ‘vase’ and it is labelled a ‘noun’. However, in the sentence Istusin voodi peal ‘I was sitting on the bed’, the Trajector is istusin ‘I was sitting’, a verb phrase and it is labelled as ‘other’.

Among the 30 experimental items, there are 4 compound and 26 simple Landmarks, 13 mobile and 17 static Landmarks with a mean length of 2, corresponding to 4 syllables. 27 of the Landmark phrases were used in the adverbial syntactic function and 3 in the modifier function. 15 of the Trajector phrases were nouns and 15 from other word classes. For the subsequent modelling of the results, it should be noted that in the previous studies the importance of these variables was assessed in a corpus-based dataset where the distribution of the variable levels is very much the same, but the entire number of lexemes included in the studies extends drastically beyond the 30 items included in the present study (cf. Klavan 2021 whose corpus sample covered 438 different lemmas). It is thus expected that these semantic and morpho-syntactic predictors may not yield models with the same predictive power as models built on corpus data provide.

4.3 Corpus-based frequency measures

This section gives an overview of the corpus-based frequency metrics used in the present study. All of the frequency metrics of interest were extracted from the Estonian National Corpus (ENC 2017; size 1.1 billion words in total; Kallas and Koppel 2017). ENC is one of the most comprehensive corpora of present-day written Estonian. It consists of the Estonian Reference Corpus (2013), the Estonian Web corpus (2013 and 2017) and Estonian Wikipedia (2017). The corpus was accessed using the Sketch Engine interface which provides access to the morphologically annotated corpus.^[1] The automatic morphological annotation for ENC 2017 has been done using the tagging tool Filosoft,^[2] which assigns the correct morphological tag to a word in 97 % of cases (Kaalep 1998: 23). We used the special Corpus Query Language (CQL) of Sketch Engine for extracting the necessary frequency metrics. The corpus data was extracted in March 2019.

Table 1 gives a quick overview of the different frequency metrics used in the study; the details of how each metric was calculated are discussed below. Appendix C lists all of the relevant values of the frequency metrics for the experimental items. Appendix D lists all of the relevant measures from collostructional analysis.

In total, we looked at seven frequency measures for the present study: (1) word lemma frequency, (2) the frequency of the particular word in the adessive case, (3) the frequency of the genitive word form in combination with the postposition peal, (4) the ratio of the particular word in combination with the adessive construction divided by the word in combination with the peal construction, (5) collostructional strength of the word in the adessive construction, (6) collostructional strength of the word in the peal construction, and (7) the ratio of the two collostructional measures. Following is a description how we operationalized the frequency measures.

lemma. The frequency of the word lemma was calculated by logarithmically transforming the raw frequency count of the lemma extracted from the ENC 2017 corpus. For example, from Appendix C we can see that the raw frequency count of the lemma for the word aken “window” is 139,805; for our modelling we used the logarithmically transformed value of 5.15.

ade_form. The frequency of the particular word in the adessive case was extracted from the ENC 2017 corpus by searching for the exact word form used in the experiment. If the word was used in the singular adessive form in the experiment, then we looked up only the frequency of this form in the corpus, e.g. aknal (window-ADE.SG) and ignored the plural forms, e.g. akendel, aknatel (window-ADE.PL). We used the same procedure when the word was used in the plural adessive form in the experiment, e.g. searching only the plural adessive form piltidel (photo-ADE.PL) and ignoring the singular form pildil (photo-ADE.SG).

peal_form. The frequency of the particular word used together with the postposition peal was extracted from the ENC 2017 corpus by searching for the exact combination used in the experiment. If the word was used in the experiment in singular genitive form together with peal, then we looked up only the frequency of this combination, e.g. akna peal (window-GEN.SG on) and ignored the plural forms, e.g. akende peal, aknate peal (window-GEN.PL on). We used the same procedure when the word was used in the plural genitive form, e.g. searching only the plural genitive form together for piltide peal (photo-GEN.PL on) and ignoring the singular form pildi peal (photo-GEN.SG on).

ade_peal_ratio. It is the log transformed ratio of two frequency counts: ade_form and peal_form, i.e. a relative frequency of the word in the adessive construction and the word in the peal construction. Because the two forms compete, it may be that the absolute frequency of the two constructions is not as important as the frequency of one construction relative to the frequency of the other construction with which it is competing.

ade_coll. It is the collostructional strength of a word in the adessive construction.

peal_coll. It is the collostructional strength of a word in the peal construction.

adecoll_pealcoll_ratio. It is the ratio of the two collostructional values: ade_coll and peal_coll, i.e. a relative frequency of the collostructional strength of a word in the adessive construction and the collostructional strength of a word in the peal construction. Similarly to the ade_peal_ratio measure, we decided to calculate the ratio of the two collostructional measures.

The computations for both of the collostructional metrics were done with the R-script “coll.analysis 3.5” developed by Stefan Gries (Gries 2014). The script uses an exact binomial test to quantify the association strength between the words and their use with the adessive and peal construction. The script allows us to calculate the association strength on the basis of a crosstabulation of the individual frequencies of the word and the construction in question as well as their joint frequency, using the Fisher exact test. For example, to calculate the association strength (i.e. collostruction strength) between the word aken and the adessive, the data in Table 2 are needed.

The Fischer exact p-value is calculated for all 30 words occurring with both the adessive and the peal-construction. From the script output we used the measure “collostructional strength” which is based on p-values of the Fischer exact test. The collostructional strength for the 30 words for both constructions is given in Appendix D. The measure can be interpreted as follows: collostructional strength > 3 = p < 0.001, collostructional strength 1.30103 = p < 0.005. For statistical modelling we centered and scaled the values.

5.1 Model with the past predictors

To test whether the predictors that have been found to be significant in the previous studies conducted on the Estonian morphosyntactic alternation between the adessive case and the postposition peal construction, we built a mixed-effects model using five linguistic features as fixed effects (length, mobility, complexity, syntactic function, Trajector word class) and subject as a random effect. The goodness of fit measures for this model are given in Table 3. According to Hosmer et al. (2013), a C value of 0.75 is acceptable discrimination. We can see that both the accuracy score (70 %) and the C value have increased compared to the intercept only model (accuracy = 67 %, C value = 0.69). Still, the model is not doing a particularly good job in predicting the choice. We may conclude that the model may be missing some important predictors and we now proceed to reporting models that take into account one frequency metric at the time. In the last section of our results, we combine the winning frequency metric(s) with the linguistic metrics to see if we are able to build a model with higher predictive power. We need to keep in mind that we are modelling the results of an experiment that has only 29 lexical items for which 103 native speakers of Estonian have provided a forced choice. Overall, the dataset is sufficiently large (2,987 datapoints), but it does not represent the linguistic diversity of lexemes we would find in a corpus-based study.

5.2 Models with one frequency metric per model

The aim of the second analysis of the data was to see to what extent the different frequency measures calculated for the experimental items can be used to explain the proportion of peal choices in the forced choice task. We decided not to fit all of the frequency metrics to the data in one single model and instead we ran 8 different models, one of them the so-called zero model with intercept only. Figure 2 is a correlation matrix of the seven frequency measures studied.

Figure 2:

Correlation of frequency measures.

We used the specific frequency metric as the fixed effect, and subject as random effect. The overall model formula for all the seven models is the following:

• choice_cx ∼ frequency metric + (1|subject)

The overall accuracy and the C measure for the eight models are given in Table 4. We also include among our models the so-called null model or the intercept only model. In terms of the model fit, the models are not performing at a very good level of accuracy. The accuracy scores range from 67 % (the intercept only model) to 71 % (models that include frequency measures for the word in combination with the peal construction). The C values tell us that the fit of the models is on the borderline between poor discrimination and acceptable discrimination (range: 0.69–0.76).

As for assessing the importance of specific frequency measures, we take decrease in the Akaike information criterion (AIC; Hosmer et al. 2013: 120) as an indicator of the importance of a particular predictor (Baayen et al. 2013: 264). AIC is used to compare the fit of models with different numbers of parameters – a smaller value is taken as an indication of a better model fit. Individual parameter estimates were tested by the likelihood ratio test, a test based on the difference in deviances. All models except Model 5 (ade_coll) were significantly better than the null model, based on the likelihood ratio tests (DF = 1; see Appendix E for details). The frequency measures were added to the intercept only model (the null model). The last column in Table 4 lists the reduction in AIC – the larger the reduction in AIC compared to the null model, the more important we consider the frequency measure to be. Based on the different model accuracy measures we have selected the following two models as our winning models: Model 2 (peal_form) and Model 6 (peal_coll).

As to the specific predictions made by the models, we may inspect the estimated coefficients of the main effects (Figure 3). We discuss here only the coefficients of our two winning models significant (Model 2 and Model 6; see Appendix E for the coefficients of the other models). The coefficients in Figure 3 indicate that for Model 2 (peal_form) the higher the frequency of the peal form with the lemma in the corpus data, the higher probability of peal construction being chosen in the forced choice task. This result is mirrored in Model 6 (peal_coll), where a higher collostructional strength of the peal construction for a lemma indicates the higher probability of the peal construction being chosen in the forced choice task. It seems that the frequency of the lemma in the adessive form (Model 1), the collostructional strength of the lemma in the adessive construction (Model 5), and the overall frequency of the lemma (Model 4) make a less prominent contribution towards explaining the choices made by the participants in the forced choice task. We return to the discussion of why frequency measures related to the peal construction play a more prominent role in explaining the variation in the forced choice data in the next section.

Figure 3:

Effect plots for Model 2 (peal_form) and Model 6 (peal_coll).

5.3 Models with frequency in combination with previous predictors

The aim of the third set of statistical analysis is to combine the winning frequency metrics with the five previous predictors from the mixed-effects regression model presented in Section 4.1. For the purposes of comparison, Table 5 also includes the model evaluation measures for the model with only the previous five predictors. We can see from the table that both frequency measures add some explanatory power – there is an increase in both accuracy and the C value and a drop in AIC when we compare the two models with a frequency measure in combination with the previous predictors to the model that only has the previous predictors (Model 0).

In order to evaluate the contribution of each predictor in each of the models likelihood ratio tests were conducted. It should be kept in mind that our focus was to test whether the frequency measures add explanatory power to the models or not. Hence, all of the predictors were kept in the final models irrespective of whether they significantly improved the model fit or not. The results of the likelihood ratio tests for each model are reported in Appendix F. From these tables it can be seen that the results of our mixed-effects logistic regression models highlight the importance of the frequency metrics in explaining the part of the variability in the choice between the adessive and peal construction in our data.

6.1 Does frequency correlate with alternant choice?

A positive answer can be provided for whether frequency affects the choice between the adessive and peal constructions in Estonian. The following two frequency metrics were particularly significant: frequency of the genitive form in combination with the postposition peal (peal_form, Model 2) and the collostructional strength of the word in the peal construction (peal_coll, Model 6).

By contrast, the frequency of the adessive form (Model 1 and Model 5) does not seem to play a role in explaining the variation in the forced choice data. Neither was the overall frequency of the lemma used with the constructions particularly significant (Model 4). In one of the few previous studies that have reported the effects of lemma and whole-word frequencies in Estonian, Lõo et al. (2018: 90) show that the variable importance of lemma frequency was inferior to that of whole-word frequency as predictors for both production latencies and articulation durations. They conclude that, at least for Estonian, lemma frequency does not afford the predictive precision that comes with whole-word frequency (Lõo et al. 2018: 90). Our study is in line with this finding concerning word lemma frequency, however, we did find word frequency to be predictive with the peal construction.

Lõo et al. (2018: 91) claim that the support for a whole-word frequency effect for Estonian case-inflected nouns fits well with the findings that in English, sequences of words also show frequency effects; the same applies to our study. Follow-up research will be needed to explore whether our winning metrics are equally predictive of subjects’ offline choices at other levels of language processing (e.g. at the level of fully abstract syntactic constructions).

6.2 Which frequency metrics yield the most accurate predictions?

In looking for an answer to our second research question – which out of several frequency metrics yields the most accurate predictions – we retain that frequency effects reflect a complex interaction between the various frequency measures of different grain-sizes. We found two frequency measures to play a particularly significant role in explaining the choices made by native speakers in the forced choice task. Such an assumption about different frequency metrics being pivotal aligns with the exemplar models of language (Bod 1998, 2006]; Pierrehumbert 2001, 2006]; Snider 2008). What matters in the context of the present study is the frequency of the words used with the peal construction and the relative frequency of the word in the adessive construction and the word in the peal construction. Our analysis shows that the more frequent the word form is used with the peal construction, the more probable the choice of peal construction in the forced choice task. The same is echoed by the collostructional measure – the stronger the collostructional strength for a particular word lemma with the peal construction, the more probable the choice of the peal construction in the forced choice task. For the relative frequency metric we found that if the adessive and peal ratio is small, i.e. the smaller the difference in the overall frequency of the word with the adessive and peal, the more probable the choice of the peal construction in the forced choice task.

Our study shows that out of the different frequency metrics that we tested, the collostructional model performed the best when looking at the frequency metrics individually. A similar finding was observed by Colleman and Bernolet (2012) who offer a detailed comparison of the results of the Dutch dative alternation based on a corpus analysis and a series of picture description experiments. As with the Estonian data, where one of the constructions, the adessive construction, is much more frequent overall compared to the other, postpositional, construction, the Dutch data demonstrates that the double object construction is by far the most frequently realized option in natural language. Yet, the experimental data display a distinct bias toward the prepositional dative. One of the explanatory factors offered by the authors is the alternation biases of the individual dative alternating verbs as measured not simply in terms of the raw observed frequencies, but in collostructional terms. Furthermore, Colleman and Bernolet (2012) also take into account the fact that different senses of a polysemous verb may well exhibit different argument structure preferences.

We set out to test, inter alia, whether collostructional analysis is extendable to Estonian, a morphologically rich language. We conclude that, at least for the peal construction, collostructional analysis provided meaningful insights into the alternation between the two Estonian morphosyntactic alternations. As to whether the collostructional analysis is extendable to morphology, we are left in the dark. We only found an effect of the collostructional strength for the peal construction, which in terms of collexeme analysis plays out nicely since the construction with peal has more to do with measuring the attraction of two separate words rather than attraction between a word and its inflection. Further research needs to be done with Estonian or other morphologically rich languages to test the applicability of the collostructional analysis to their morphology. Our models show that the collostructional strength between the word and the adessive case affix did not predict the choices in our forced choice task.

6.3 Does the inclusion of frequency significantly improve model fit?

For our third research question, we pitted the models with frequency metrics against a model with five other relevant variables identified in the previous research (length, mobility, complexity, syntactic function, trajector word class). Out of this “competition”, the model with the collostructional measure for the peal construction was the “winner”. The model with the collostructional measure in the formula on its own does a better job at predicting the choices in the task than a model with all the five variables combined. Yet, adding the collostructional frequency metric to the model in addition to the other variables yields an even better model. We, therefore, infer that frequency on its own is a particularly strong predictor, but does not explain away the other predictors since adding them into the picture is still significant.

In light of the result that the ratio variable is also significant, we may entertain the idea of whether the token frequency of the peal construction is actually an artifact of the relative frequency measure, i.e. the ratio variable. Hay (2001) discusses the role of relative versus absolute frequency effects in the decomposability of morphologically complex words and whether the two measures are correlated. Hay (2001: 1053) concludes that absolute frequency and relative frequency are not independent of one another. Our paper provides experimental evidence that a relative-frequency effect exists – participants consistently chose the peal construction more when the ratio of the adessive construction and the peal construction with the same word was small.

An important insight provided by our study is the fact that frequency affects the adessive and peal construction differently. The peal construction may not necessarily have to be very frequent for it to win out in the forced choice task; similarly, the adessive construction in combination with a particular word may be very frequent but still not affect the choices. What we observe in our data is the phenomenon referred to as the inverse frequency effect. This effect explains the observation that less frequent structures tend to prime more strongly (Bock 1986; Colleman and Bernolet 2012; Ferreira 2003; Jaeger and Snider 2013). For example, Bock (1986) showed syntactic persistence for passives, which are overall rather infrequent in her production experiments; the much more frequent active structure were not found to prime. In our study, we have also observed stronger frequency effects for the less frequent construction (the postpositional variant, in our case). Klavan and Schützler (2023) report the frequencies of the Estonian exterior locative cases (including the adessive) and the corresponding postpositions (including the postposition peal) from the Estonian National Corpus (ENC 2017; 1.1 billion words). These figures show that the adessive construction is 127 times more frequent in the language compared to the peal construction (30,661,120 adessive case occurrences vs. 241,263 peal construction occurrences).

Behind these drastically different raw frequency measures of adessive and peal lies the fact that the adessive case has other essential functions to fulfil in the Estonian language, for example expressing the possessor in the possessive construction ( Maril on kaks last. ‘Mari has two kids.’) and agents with finite verb forms (See asi ununes mul kiiresti. ‘I quickly forgot about that thing.’) (see Appendix G for more details about the different functions of the adessive case in Estonian). Support for the inverse frequency effect is found in Klavan (2020) who has shown that both the mixed-effects logistic regression model and the naïve discriminative learning model do a slightly better job in predicting the peal construction than the adessive construction. Since we only found an effect of the peal construction frequency to affect the choices, we can conclude that for adult native speakers of Estonian, a word in combination with peal is a unit of representation while a word in combination of the adessive is not. This conclusion needs further study, especially psycholinguistic experiments; particularly interesting would be to look at the role of these two constructions in language learning.

There are, of course, other frequency metrics which we can further investigate that are currently missing from our study. One such metric relates to the semantics of the adessive construction. In Estonian, the adessive case expresses other relations besides the locative (cf. Appendix G). Importantly, the adessive case is used to temporal relations and agents with finite verb forms, including the possessor in the possessive construction. For these functions of the adessive, postpositional construction is not an attested alternative construction. The polysemy of the adessive case creates a possible ambiguity for language users since it is not clear from the adessive form itself whether location or other more abstract functions are intended. In a pilot study, we calculated the adessive function ratio by counting the usages of the word in the adessive expressing the locative functions versus other functions. We found that if the ratio is negative, i.e. the adessive form with a particular word has a low proportion of uses where a locative function is expressed, the more probable the choice of the peal construction in the forced choice task. Further studies will hopefully shed light on whether the semasiological frequency of the adessive form influences the onomasiological frequency (expressing the locative function of the same word with adessive or postpositional construction). This can be tied in with research carried out by Mehl (2021: 228) who has shown that semasiological approaches are useful for research questions related to exposure rates while onomasiological approaches are useful for research questions related to selection preferences.

Similarly to the potential effect of the adessive case polysemy, verb sense has been suggested as an important predictor that needs to be considered in argument structure studies. Hare et al. (2003) suggest that the empirical discrepancies found in studies that have looked at how comprehension is affected by subcategorization biases of the verb arise because verb-sense has not been taken into account. For example, Gahl (2000), Garnsey et al. (1997), MacDonald (1994), and Trueswell et al. (1994) found an effect of subcategorization biases of the verb, while Ferreira and Henderson (1990), Kennison (2001), and Pickering et al. (2000) did not. The conclusion to be drawn from these studies is that different senses of a polysemous verb may exhibit different argument structure preferences. The results of our preliminary follow-up study about the polysemy of the adessive case clearly demonstrate that it is important to take into account the relative frequency of the locative adessive function versus other functions – an additional frequency measure of a different grain-size.