2. Object of study

It is a well-established fact of Icelandic that the subject status of a verbal argument is not necessarily associated with nominative case marking (Andrews Reference Andrews1976, Thráinsson Reference Thráinsson1979, Zaenen, Maling & Thráinsson Reference Zaenen, Maling and Thráinsson1985, Sigurðsson Reference Sigurðsson1989, Jónsson Reference Jónsson1996, Barðdal Reference Barðdal2001, inter alia). For these so-called oblique subjects, at least the following nine subjecthood diagnostics have been identified (Andrews Reference Andrews1976, Thráinsson Reference Thráinsson1979, Zaenen, Maling & Thráinsson Reference Zaenen, Maling and Thráinsson1985, Sigurðsson Reference Sigurðsson1989, Jónsson Reference Jónsson1996, Barðdal Reference Barðdal2001, Reference Barðdal2006, Reference Barðdal2023, Barðdal, Eythórsson & Dewey Reference Barðdal, Eythórsson and Kim Dewey2019, inter alia):

• first position in declarative clauses

• subject–verb inversion

• first position in subordinate clauses

• subject-to-object raising

• subject-to-subject raising

• long-distance reflexivisation

• clause-bound reflexivisation

• conjunction reduction

• control infinitives

It has been demonstrated that Icelandic oblique subjects pass all of the aforementioned tests, usually referred to in the literature as behavioural tests, as opposed to coding tests (cf. Keenan Reference Keenan and Li1976). Note that the coding test involving subject–verb agreement is not applicable to oblique subjects, as is well known in the literature (Sigurðsson Reference Sigurðsson1990–1991, Reference Sigurðsson, Bhaskararao and Subbarao2004, inter alia), since agreement is only found with nominative arguments in Icelandic, Germanic, and the Indo-European languages in general (cf. Barðdal Reference Barðdal2023:97–98). Moreover, the tests in the bulleted list above confirm the status of oblique subjects as behavioural subjects in Icelandic. In this article we intend to lend corpus-based support to the first and the third behavioural test, i.e. word order distribution in main and subordinate clauses, applying them to Dat-Nom and Dat-Nom/Nom-Dat verbs in Icelandic.

It has already been mentioned above that Dat-Nom verbs come in two different guises: non-alternating Dat-Nom verbs, and alternating Dat-Nom/Nom-Dat verbs. The latter class, which allows for two diametrically opposed case frames, was first discovered by Bernódusson (Reference Bernódusson1982), and it has since been the subject of several studies (Jónsson Reference Jónsson1997–1998, Barðdal Reference Barðdal1999, Reference Barðdal2001, Platzack Reference Platzack1999, Sigurðsson Reference Sigurðsson2006a, Rott Reference Rott2013, Reference Rott2016, Barðdal, Eythórsson & Dewey Reference Barðdal, Eythórsson and Kim Dewey2014, Reference Barðdal, Eythórsson and Kim Dewey2019, Wood & Sigurðsson Reference Wood and Ármann Sigurðsson2014, Somers & Barðdal Reference Somers and Barðdal2022, Reference Somers and Barðdal2023, inter alia). In this article, we refer to them either as ‘alternating Dat-Nom/Nom-Dat verbs’, or as nægja-verbs.

Verbs of the nægja type allow both the dative as well as the nominative to take on the role of subject, yet not at the same time. This is manifested in the fact that each of the aforementioned arguments independently passes the subject tests mentioned above, so that, when the dative behaves as the subject, the nominative takes on the role of object, and vice versa (cf. Barðdal Reference Barðdal1999, Reference Barðdal2001, Reference Barðdal2023:Ch. 3, Barðdal, Dewey & Eythórsson Reference Barðdal, Eythórsson and Kim Dewey2014, Reference Barðdal, Eythórsson and Kim Dewey2019, where it is shown that either argument passes all the subject tests in Icelandic). Examples (1a–b), here repeated as (3a–b), illustrate this phenomenon, in that they show that both arguments may take initial position in declarative clauses without there being a change in meaning or focus.

It is clear, however, that the two word orders reflect two different construals of the same event, namely that an experiencer directs his or her attention to a stimulus in (3a) and that a stimulus affects an experiencer in (3b) (cf. Barðdal Reference Barðdal2001, Reference Barðdal2023:Ch. 3, Barðdal, Eythórsson & Dewey Reference Barðdal, Eythórsson and Kim Dewey2019). We take this to be a consequence of the fact that the relevant verbs are force-dynamically neutral in the sense of Talmy (Reference Talmy and Shibatani1976), meaning that there is no causal chain found in the event structure of these verbs, as neither of the participants acts upon the other. In other words, there is no causation involved (see also Croft Reference Croft2012:233, Barðdal Reference Barðdal2023:Ch. 3).

Because of their dyadic nature, Barðdal (Reference Barðdal2001, Reference Barðdal2023) and Barðdal, Eythórsson & Dewey (Reference Barðdal, Eythórsson and Kim Dewey2019) have suggested that alternating verbs of this type in fact instantiate two different argument structure constructions: a Nom-Dat construction that licenses a nominative subject and a dative object, i.e. the stimulus affects the experiencer construal, and a Dat-Nom construction that licenses a dative subject and a nominative object, i.e. the experiencer directs his/her attention towards a stimulus construal. Our approach is fully in line with this analysis, as we subscribe to the view that the subject is the first argument of the argument structure. This is a theory-neutral definition of subject, as all theoretical frameworks employ argument structure or subcategorisation frames in their machinery.

Returning to the examples in (3a–b) above, what speaks against a simple topicalisation analysis is the positioning of the verbal arguments relative to the conjugated verb hafði ‘had’. In Icelandic the subject must be adjacent to the conjugated verb (unless it is either indefinite or heavy): that is, it must either precede or follow the verb. This is because of the so-called verb-second constraint, which also operates on other Germanic languages (cf. Eythórsson Reference Eythórsson1995, Axel Reference Axel2007:27–67, Harbert Reference Harbert2007:398–415, Thráinsson Reference Thráinsson2007:40–45, inter alia). Had either (3a) or (3b) been a topicalisation of the other, then the nominative in (3a) and the dative in (3b) had been realised in between the conjugated verb hafði ‘had’ and the past participle nægt ‘sufficed’. This is not the case, though, since both the nominative in (3a) and the dative in (3b) are realised after the non-finite verb, which is an object position.

Consider now the examples in (4a–b) below, which show that attempts at topicalising the object with alternating verbs result in ungrammatical structures in Icelandic (for one exception to this, see (17a–c) below). As already stated above, this involves an inversion of the subject and the verb. Hence, the intended subject argument immediately follows the verb in these examples and the intended object argument occurs in first position.

Interestingly, not all Dat-Nom verbs allow for the type of alternation shown in (3a–b), as is already mentioned above. Some, such as líka ‘like’, only license dative subjects; their nominative argument invariably behaves as an object with regard to word order distribution. The fact that, for these verbs, subject status is unequivocally associated with the dative case is illustrated by examples (5a–b).

Recall that (5b) is ungrammatical because the subject barninu ‘the child’ and the conjugated verb hafði ‘had’ have been separated from one another by the past participle líkað ‘liked’. If the nominative is realised preverbally for information-structural reasons, the dative, being the syntactic subject, breaks open the verbal group and is once again reunited with the conjugated verb, as shown in (6).

Hence, the example in (6) represents topicalisation and not neutral word order; that is, it is an example of topicalisation that fronts a non-subject constituent to initial position for emphasis (Thráinsson Reference Thráinsson2007:342). Since the dative subject and the conjugated verb have now been reunited, the example is grammatical. Verbs, like líka, that only allow their dative argument to pass the aforementioned subject tests are henceforth called ‘non-alternating Dat-Nom verbs’, but we will also refer to them as líka-verbs in the remainder of this article. Thus, the default argument structure construction that líka-verbs instantiate is the Dat-Nom construction. The linear nominative-first order with líka-verbs is only used for information-structural purposes (Barðdal, Eythórsson & Dewey 2019). In essence, this means that líka-verbs have lexicalised only one of the two construals mentioned above, namely the construal where the experiencer directs his/her attention to a stimulus.Footnote ²

Both alternating Dat-Nom/Nom-Dat verbs, as well as non-alternating Dat-Nom verbs, should be distinguished from ordinary Nom-Dat verbs, or – as we will also be calling them – hjálpa-verbs. These are also two-place predicates requiring a nominative and a dative argument, but, crucially, it is the nominative argument that behaves as the syntactic subject and the dative as the object (Barðdal, Eythórsson & Dewey Reference Barðdal, Eythórsson and Kim Dewey2019:158), as is evident by the grammaticality of (7a) and the ungrammaticality of (7b).

Thus, hjálpa-verbs constitute the mirror counterpart of the aforementioned líka-verbs, in that they exclusively occur in the Nom-Dat argument structure construction, which is the opposite of the Dat-Nom argument structure construction. Also, hjálpa-verbs only allow for preposed datives in cases where the dative is topicalised, as is shown in (7c).

In this study we lend corpus-based statistical support to the analysis that the dative and the nominative arguments of nægja-verbs are indeed syntactic subjects. This we do by comparing the frequency of topicalised arguments in first position to the frequency of subjects in first position. In other words, if an oblique argument behaves as a subject, it can be expected to be strongly associated with first position in declarative clauses (diagnostic test 1) and first position in subordinate clauses (diagnostic test 3), while topicalised objects would not show the same association. Moreover, since word order in Icelandic is understood to be quite rigid (Thráinsson Reference Thráinsson2007:342), topicalisation can be expected to be relatively rare, and even less common in subordinate clauses than in main clauses. This is confirmed by Angantýsson’s (Reference Angantýsson, Woods and Wolfe2020:261) study, although his study is based on acceptability judgements and not corpus frequencies. Nevertheless, empirical studies on how frequent topicalisation actually is are quite scarce for Icelandic.

One study that does include frequency counts is that of Callegari & Ingason (Reference Callegari and Karl Ingason2021). In their diachronic investigation of matrix-clause ditransitive constructions, they explore object topicalisation in Icelandic texts from the twelfth to twenty-first centuries, drawing their data from the IcePaHC corpus (Wallenberg et al. Reference Wallenberg, Karl Ingason, Freyr Sigurðsson and Rögnvaldsson2011). Callegari & Ingason include both pronominal and nominal objects in their study, i.e. objects realised as both pronouns and full NPs. Out of a total of 1,100 hits, they find 128 instances of object topicalisation, of which 89 have the direct object topicalised (8%), and 39 the indirect object (3.5%). Thus, topicalisation affects approximately 11.5% of the tokens under study, and direct object topicalisation turns out to be more than twice as common as indirect object topicalisation. Callegari & Ingason do not include an unambiguous overview of object topicalisation per century, but a summary graph seems to reveal that, for the twenty-first-century data, both direct objects as well as indirect objects are each topicalised approximately 6% of the time.

Another study worth mentioning in this respect is that of Barðdal & Eythórsson (Reference Barðdal and Eythórsson2012), who map out the word order patterns for monotransitive verbs licensing nominative subjects. Barðdal & Eythórsson’s data have also been drawn from the IcePaHC corpus, although an earlier version than that of Callegari & Ingason. The main difference between the two versions is the size of the corpus, with the more recent version containing nearly 60% more texts (1,002,390 vs. 632,000, Einar Freyr Sigurðsson p.c.). With all else being equal, and on the assumption that the smaller version of the corpus is large enough, there is no reason to assume that a comparison involving the frequency of topicalisations between these two studies is not justified. Thus, zooming in on Barðdal & Eythórsson’s (Reference Barðdal and Eythórsson2012) results for verb-second clauses (i.e. SVO vs. OVS structures), it turns out that nominative subjects occur 2,327 times (or 80%) in initial position, and 578 times (or 20%) in postverbal position.Footnote ³ Therefore, object topicalisation is clearly much more frequent with monotransitives than with ditransitives, at least diachronically. See also Booth & Beck’s (Reference Booth and Beck2021) study where it is statistically documented that the clausal initial position in Modern Icelandic is a topic position.

It is unclear if the predicates in our study are equally permissive of topicalisation as Barðdal & Eythórsson’s (Reference Barðdal and Eythórsson2012) monotransitives and Callegari & Ingason’s (Reference Callegari and Karl Ingason2021) ditransitives. For that reason, we map out word order preferences for both the hjálpa and líka classes and use these counts as a baseline against which word order preferences for the nægja class will be measured. We now turn to a description of our methodology, before we present our findings in Sections 4–5 below.

3. Methodology

This study is based on 15 simple verbs that fall into one of three categories: (i) ordinary Nom-Dat verbs (the hjálpa type), (ii) non-alternating Dat-Nom verbs (the líka type), and (iii) alternating Dat-Nom/Nom-Dat verbs (the nægja type). Our aim was to follow Rott (Reference Rott2013) in our selection of verbs, but some of the verbs he used were too infrequent in the corpus to yield enough eligible tokens. Thus, we complemented the dataset with additional known non-alternating Dat-Nom and alternating Dat-Nom/Nom-Dat verbs (cf. Jónsson Reference Jónsson1997–1998, Barðdal Reference Barðdal1999:89, Reference Barðdal2001:53–58, Reference Barðdal2023:81–83). Each category contains five verbs:

1. (i) Ordinary Nom-Dat verbs: hjálpa ‘help’, líkjast ‘resemble’, mótmæla ‘contradict’, treysta ‘trust’, and þakka ‘thank’.

2. (ii) Non-alternating Dat-Nom verbs: áskotnast ‘receive’, blöskra ‘be shocked, be horrified’, leiðast ‘be bored’, líka ‘like’, and þykja gott/slæmt/… ‘think, find, seem good/bad/…’.

3. (iii) Alternating Dat-Nom/Nom-Dat verbs: duga ‘suffice, be enough’, dyljast ‘be hidden to somebody, be aware’, endast ‘last’, henta ‘suit, befit’, and nægja ‘be enough, be sufficient’.

We follow Rott (Reference Rott2013:103) in using blöskra ‘be shocked, be horrified’, leiðast ‘be bored’ and líka ‘like’ in the class of non-alternating Dat-Nom verbs, and henta ‘suit, befit’ and dyljast ‘be hidden to somebody, be aware’ in the alternating Dat-Nom/Nom-Dat class.

The analysis is based on a data collection from the Icelandic Web 2020 corpus (isTenTen20, Jakubíček et al. Reference Jakubíček, Adam Kilgarriff, Rychlý and Suchomel2013), which consists of approximately 520 million words. The corpus itself has been accessed through the Sketch Engine interface. For each of the aforementioned verbs, a lemmatised search query has been carried out targeting the verb’s bare infinitival form. That is also true for the etymologically reflexive -st-verbs, as the search engine considers -st-forms to be instantiations of the non-suffigated base form. Thus, líkjast, áskotnast, leiðast, dyljast, and endast were run as líkja, áskotna, leiða, dylja, and enda, respectively.

The material has subsequently been extracted in one or more files containing 10,000 randomised tokens per verb type, depending on how abundant the data were. In contrast to Rott, who also includes middle field tokens, we only focus on tokens in which the main verb is flanked by either a nominal or a pronominal element. Thus, only instances of the type [Nom-V-Dat] or [Dat-V-Nom] have been taken into account, regardless of clause type. As a consequence, there are no tokens in our dataset of any other kinds of topicalised elements, which in turn excludes, for instance, adverbials.

Contrary to the Mainland Scandinavian languages, Icelandic is a so-called symmetric V2-language, which means that the conjugated verb takes second position both in main clauses as well as in subordinate clauses (Thráinsson Reference Thráinsson2007:41, Angantýsson Reference Angantýsson, Woods and Wolfe2020:243). Eligible tokens are therefore not restricted to main clauses only but also include subordinate structures. Per verb type, the first 200 tokens have been withheld for study. Hence, the total number of collected tokens equals 3,000, and the number of collected tokens per verb class equals 1,000.

Per token, all arguments, dative and nominative, have been manually annotated for the following variables: case, (pro)nominality, pronoun type (if applicable), referentiality, person, number, definiteness, animacy, and length. The choice of variables is motivated by two considerations, namely that each of these (a) are well known in the field for affecting word order, (b) may serve as a proxy for discourse prominence or topicality (cf. the discussion in Sections 4.3 and 5.3 below). For each variable in boldface, the relevant values are rendered in small caps, followed by examples in brackets. The nine variables are discussed below.

• Case: nominative (þessi sími ‘this phone’ nom.sg, mín eigin föt ‘my own clothes’ nom.pl) or dative (hundinum ‘the dog’ dat.sg, unglingunum ‘the youngsters’ dat.pl).

• (Pro)nominality: pronoun (þú ‘you’ sg, ykkur ‘you’ pl, einhverjum ‘some’) or full NP (Ísland ‘Iceland’, ýmsir þingmenn ‘some congressmen’ nom.pl, bókin ‘the book’ nom.sg).

• Pronoun type: personal (ég ‘I’, hann ‘he’, þeir ‘they’ 3.m), demonstrative (þessi ‘this’, hinum ‘the other’, slíkur ‘such’), indefinite (öllum ‘all’, engum ‘no-one’, báðum ‘both’), or reciprocal (hvert öðru ‘each other’ sg.n, hver annarri ‘each other’ sg.f). Reflexives are excluded from study, as they are hypothesised to prefer the postverbal slot. In line with Heylen (Reference Heylen2005:103), conjoined pronouns are also excluded, as they arguably lose their pronominal status.

• Referentiality: referential or correlative. Icelandic allows for the third person personal pronoun það ‘it’ to be used as an expletive or a correlate. Expletives are wholly absent from our dataset, but correlates, which have a clause-anticipating function, show up approximately 300 times. It is hypothesised that such placeholders, given their impoverished semantic status, are inclined to follow the verb, rather than precede it. All remaining arguments, either nominal or pronominal, have been annotated as referential.

• Person: first person (mér ‘me’ dat.sg, við ‘we’, okkur feðgum ‘us, father and son’), second person (þú ‘you’ sg, ykkur ‘you’ pl, yður ‘you’ pl.hon), or third person (þeim ‘them’ dat.pl, henni ‘her’ dat.sg.f, augum manna ‘people’s eyes’ dat.pl).

• Number: singular (rigningarvatn ‘rainwater’ nom.sg, stúlkunni ‘the girl’ dat.sg) or plural (stúlkum ‘girls’ dat.pl, Hauknum og frú ‘Haukur and his wife’ dat.pl).

• Definiteness: definite or indefinite. Icelandic pronouns are always definite, except for indefinite pronouns (báðar ‘both’ pl.f, manni ‘one’ dat.sg) and indefinite demonstratives (slíkt ‘such a thing’ sg.n). NPs are considered to be definite if they are preceded by a definite demonstrative pronoun (þessi hey ‘this hay’ nom.pl) or a possessive pronoun (þitt fyrirtæki ‘your company’ nom.sg). Constituents that are followed by a cliticised definite article (blaðinu ‘the paper’ dat.sg), a possessive pronoun (hlátur hans ‘his laughter’ nom.sg), or a postposed genitiveFootnote ⁴ (stjórn félagsins ‘the company board’ nom.sg) also receive a definite reading. Names of people (Þorsteini dat.sg), institutions (Fjölmiðlanefnd ‘the Media Committee’ nom.sg), places (Keflavík nom.sg), and population groups (Reykvíkingum ‘the people of Reykjavík’ dat.pl) are inherently definite. If a conjoined constituent exhibits conflicting definiteness, in that one conjunct is definite but the other is indefinite, the string is coded for the first conjunct. Thus, the string 4x4 klúbburinn og allir sem elska hálendið í sinni hráustu mynd ‘the 4x4 club and all who love the highlands in their rawest form’ is coded as definite, because of the definite status of the first conjunct (4x4 klúbburinn ‘the 4x4 club’).

• Animacy: individual, collective, inanimate, non-inferable or NA. The label ‘individual’ is used to index constituents referring to humans (bræðurnir ‘the brothers’ nom.pl), animals (fuglum ‘birds’ dat.pl), and what Bresnan & Ford (Reference Bresnan and Ford2010:175) call ‘humanoid beings’ (Guð ‘God’ nom.sg, heilagur andi ‘the Holy Spirit’ nom.sg). This includes cells (krabbameinsfrumur ‘cancer cells’ nom.pl). Groups of individuals are annotated as collective (fólkið ‘the people’ nom.sg, Landsbankanum ‘National Bank’ dat.sg). All other constituents, including plants (þessi jurt ‘this plant’ nom.sg), fungiFootnote ⁵ (myglusveppum ‘mould’ dat.pl), and dead animals, are labelled ‘inanimate’. When animacy cannot unequivocally be determined, we have resorted to the label ‘non-inferable’. This, for instance, applies to lántakanda ‘borrower’ dat.sg, which, in the given context, can both refer to an individual as well as to a corporate borrower. Pronouns that serve as placeholders for a subclause are annotated as ‘NA’, since their referent is linguistic, and not extra-linguistic. If a conjoined constituent exhibits conflicting animacy, in that one conjunct complies with one label but the other complies with another label, the string is coded for the first conjunct. Thus, the string bæði einstaklingum og fyrirtækjum ‘both individuals as well as companies’ is coded as ‘individual’, because that is the label that captures the animacy status of the first conjunct (bæði einstaklingum ‘both individuals’).

• Length: constituent weight measured in words.

We now turn to our findings and a discussion thereof.

5. Statistical modelling

In what follows, we investigate the factors guiding the word order variation in the set of alternating Dat-Nom/Nom-Dat verbs by means of two logistic regression models. This involves a comparison across all configurations, on the one hand, and across double NPs, on the other. The reason why we single out double NPs is because pronouns are well known to skew word order preferences (Du Bois Reference Du Bois1987, Croft Reference Croft2012, Cristofaro Reference Cristofaro, Bakker and Haspelmath2013, Reference Cristofaro, Karsten Schmidtke-Bode, Maria Michaelis and Seržant2019, Booth & Beck Reference Booth and Beck2021, inter alia). Importantly, since henta behaves as a clear outlier, it is excluded from the remainder of the analysis. As such, the results presented in the current section are only based on our findings for the verbs duga, dyljast, endast, and nægja.

The purpose of the logistic regression analyses is to identify, and quantify, nuanced empirical interactions between the factors that we hypothesise are involved in the alternation. These include the variables discussed in Section 3, for which the dataset is annotated, namely case marking, pronominality, pronoun type (if applicable), referentiality, person, number, definiteness, animacy, and length. The dependent variable is argument position, either first or second. Binary logistic regression is a probabilistic algorithm that models the outcome as a probability, conditional on the value of the predictor variables (Harrell Reference Harrell2015:219). Although binary logistic regression makes relatively few assumptions, as with any regression model, collinearity (i.e. correlation) among the predictors can be a concern (Harrell Reference Harrell2015:255).

We have chosen to use ordinary logistic regression rather than mixed-effect/multilevel models. The reason for this is simple: the natural group-level (or random effect) in such a model would be verbs, but a much larger number of verbs would need to be included to defend the added complexity of mixed-effect models (Gelman & Hill Reference Gelman and Hill2007:247). When the random effect variable has few levels, mixed-effects logistic models reduce to ordinary logistic models with only fixed effects, in the absence of meaningful group-level variation.

For evaluating the logistic regression models, we rely on a combination of inspecting model residuals (Gelman & Hill Reference Gelman and Hill2007:97–101) and measures of predictive capability, in particular the c-index, since formal tests of fit are often inappropriate for logistic regression (Harrell Reference Harrell2015:236). The c-index measures the proportion of correctly classified responses when comparing the predictions of the model with the observed values in the dataset (Harrell Reference Harrell2015:257). While several other measures exist (Harrell Reference Harrell2015:256–257), we have chosen to report the c-index since it has a reasonably intuitive interpretation. A c-index value of .5 indicates random choice, 1.0 indicates perfect prediction, and .8 and above is often taken as an indication of good predictive capability (Baayen Reference Baayen2008:284, Harrell Reference Harrell2015:257). However, it is worth noting that there is some arbitrariness in these thresholds and in medicine, for example, the threshold for an acceptable model is usually taken to be .7 (Hartman et al. Reference Hartman, Kim, He and Kalbfleisch2023, White et al. Reference White, Parsons, Collins and Barnett2023). Importantly, over-reliance on a single measure can be detrimental, as noted by White et al. (Reference White, Parsons, Collins and Barnett2023), which is why we use the c-index alongside inspection of model residuals, bearing in mind that although a higher c-index is better than a lower one, it does not tell the whole story.

The logistic regression models have all been fitted in R using the rms package. A close inspection of the binned model residuals shows no signs of structural problems. Due to the skewed (i.e. non-symmetrical) distribution in the underlying data set, the length variable has been transformed by taking the natural logarithm of the observed data, which reshapes the data by adjusting the scale, resulting in a more symmetrical distribution. Although other logarithm bases would work equally well for the data transformation, the natural logarithm benefits from being directly interpretable in the model as proportional differences (Gelman & Hill Reference Gelman and Hill2007:60–61). A positive regression coefficient indicates that a variable is associated with the first argument position, while a negative value signals association with the second argument position.

Section 5.1 presents the results of the logistic regression analysis modelling all the data obtained across configurations, whereas Section 5.2 singles out the tokens instantiating the double-NP configuration. As such, the first model is based on all 800 observations and the second on 131 observations. In Section 5.3 we discuss how our findings tie in with the concept of topicality.

5.1 Across configurations

The output of the first logistic regression model, which builds on all 200 observations per verb type, is presented in Table 9. Recall that henta, the outlier, has been excluded. The c-index of the model is .794, a value that is only decimals away from what is commonly taken as a good predictive capability (Baayen Reference Baayen2008:204, Harrell Reference Harrell2015:257).

Table 9.

Results of the logistic regression model for alternating verbs across configurations excluding henta (N = 800). Significant p-values are in boldface

Table 9 shows the logistic regression coefficient (𝛽), standard error (SE), z-score (Z) and p-value (p) for eight variables, seven of which exert a significant influence on the alternation. Only animacy (value: inanimate) does not have any predictive power, presumably because it strongly correlates with nominative case. Positive regression coefficients indicate an association with the first argument slot, while negative regression coefficients indicate an association with the second slot. As such, animacy (value: individual), animacy (value: NA), case (value: nominative), and person (modelled numerically, see below) are associated with the first argument position, whereas definiteness (value: indefinite), length, and number (value: singular) are tied to the second argument position. Importantly, factors generating a significant effect do not necessarily correlate with one other. As an example, the second argument is very often either indefinite or long, but it is not necessarily associated with both properties at the same time. All the logistic regression coefficients represent the log-odds ratio of switching from second to first position.

Starting with the variables tied to the first argument position, the model attributes a major effect to nominative case marking and animacy (value: NA), which generate a coefficient of 2.26 and 1.33, respectively (corresponding to a 56.5% and 33.2% increase in the likelihood of switching from second to first argument slot). The fact that the former, nominative case, accounts for such a large portion of the variation is hardly surprising, as approximately two thirds of all tokens (i.e. 547 out of 800) with duga, dyljast, endast, and nægja instantiate the Nom-Dat order (cf. Section 4.2.1 above). Nevertheless, the question remains what exactly this means. That is, are alternating verbs, in fact, more strongly drawn to the Nom-Dat order than they are to the reverse Dat-Nom order? Or is the Nom-Dat bias in our sample merely the result of chance? We return to this point in Section 5.2 below.

The second variable closely tied to the first argument position, i.e. animacy (value: NA), captures all instantiations of correlative það ‘it’, which is a third person personal pronoun anticipating a subclause. Correlative það, exactly like expletive and existential það, is indeed well known to occur clause-initially in Icelandic (Rögnvaldsson Reference Rögnvaldsson2002, Thráinsson Reference Thráinsson2007:366–367, inter alia). Still, the fact that this correlate so willingly takes initial position is remarkable, as it goes against the expectation that semantically impoverished units should take a less prominent position than arguments referring to extralinguistic entities (cf. Siewierska Reference Siewierska1993:831). Somers & Barðdal (Reference Somers and Barðdal2023:19–21), for instance, have shown that the strong inclination of Icelandic correlates towards the nominative-before-dative order is something that sets these apart from their German counterparts, since the latter are much more permissive of alternation.

The third variable whose connection with the first argument position is significant is person. We decided to encode this variable as numeric, not categorical, and every unit increase in person (i.e. from first to second, and from second to third) yields an increase in association with the first slot. The effect is quantified by the coefficient as .74. This means that the likelihood of an argument being realised preverbally increases by approximately 18.5% for every one-unit increase along the scale. Thus, constituents are overall more likely to be allotted the first slot if their referent is non-local, third person, as opposed to local, first and second person. Interestingly, this constitutes a violation of the person hierarchy, which stipulates that local pronouns should take precedence over both non-local pronouns and full NPs (Silverstein Reference Silverstein and Dixon1976, Siewierska Reference Siewierska1993:831, Croft Reference Croft2003:130, Haude & Witzlack-Makarevich Reference Haude and Witzlack-Makarevich2016, inter alia).

The reason, we believe, that non-local referents, i.e. third person, are more likely to occur in the first slot, despite the person hierarchy, is due to the effect of correlates, which nearly always occur preverbally (140 out of 148 cases, or 95%), as opposed to postverbally (eight out of 148 cases, or 5%). The number of local pronouns in our dataset is quite limited (a mere 250 across argument positions). One reason why local pronouns are so rare is because these are never nominative in our dataset (however, see below). Local pronouns are also more strongly tied to the second slot (197 out of 250 cases, or 79%). Interestingly, 118/197 postverbal (dative) local pronouns compete with a nominative pronoun, and nominative pronouns show a strong tendency to take first position in any case.

Moreover, in accordance with the person hierarchy, if a nominative is a first or second person pronoun, i.e. a local pronoun, only the Nom-Dat word order is acceptable with alternating predicates in Icelandic. One such attested example, cited from Barðdal & Eythórsson (Reference Barðdal and Eythórsson2003), is given in (17a) with the nominative við ‘we’ in first position. As is shown in (17b), if the nominative is a local pronoun, the Dat-Nom construction is ungrammatical, since the nominative við ‘we’ cannot occur in the object position immediately following the non-finite verb. This analysis is further confirmed by the example in (17c), which shows that if the dative occurs in first position, the local pronoun must invert with the finite verb, which is a clear-cut subject property. For obvious pragmatic reasons, the constructed examples in (17b–c) render the dative fólki ‘people’ as a definite NP instead of an indefinite one.

Recall that at the beginning of this article (see examples 3–4), we provided evidence for the analysis that alternating verbs may instantiate two diametrically opposite argument structure constructions, Dat-Nom and Nom-Dat, and that neither structure is a topicalisation of the other. This is invariably true in all cases except when the nominative is a local pronoun, as in (17) above.

Returning now to the logistic regression analysis, another variable that shows a mild preference for the preverbal slot is animacy (value: individual), with a coefficient of .77. Evidently, this is fully in line with the expectation that animate beings should take precedence over both collectives and inanimates (Allan Reference Allan1987, Siewierska Reference Siewierska1993, Dahl & Fraurud Reference Dahl, Fraurud, Fretheim and Gundel1996, inter alia). Rott (Reference Rott2013) also found an effect for animacy in his study of four Icelandic alternating verbs. More specifically, he has shown that the nominative is hardly ever animate, but that it invariably precedes the dative when it is. Our data are indicative of a similar trend: out of 800 nominative constituents, a mere 14 are animate. Of these, 11 (or 79%) are attested in the Nom-Dat order, and three (or 21%) in the reverse Dat-Nom order. The same holds, mutatis mutandis, for the dative: it is hardly ever inanimate (19 out of 800 tokens), but when it is, it shows a very strong preference for the Nom-Dat order (15 tokens, or 79%) and not the Dat-Nom order (four tokens, or 21%).Footnote ⁶ Observe that we hereby dispel the myth that the dative of Dat-Nom verbs is animate by definition (see Kutscher Reference Kutscher2009:24, Verhoeven Reference Verhoeven2009, Reference Verhoeven2015, Rott Reference Rott2013:93). The tendency for the dative of such verbs to be animate is indeed very strong, but it is by no means an absolute.

Two variables that associate with the postverbal slot are length and definiteness (value: indefinite) (coefficients −.83 and −1.08). Again, these facts rhyme well with what is known in the literature, namely that indefinite and heavy constituents tend to occur later in the clause (see Behaghel Reference Behaghel1909/10, Allan Reference Allan1987, Siewierska Reference Siewierska1993, Arnold et al. Reference Arnold, Losongco, Wasow and Ginstrom2000, inter alia). Specifically with regard to the Dat-Nom/Nom-Dat alternation, Rott (Reference Rott2013) has also found length to exert an effect, but his evidence for this factor only relates to the distribution of clausal nominatives, which greatly prefer the Dat-Nom order (82 out of 87 tokens, or 94%) to the Nom-Dat order (five out of 87 tokens, or 6%). Our own study conclusively confirms that length is a factor even when both arguments are (pro)nominal.

Rott also finds an effect for definiteness, but the results are again clouded by the high number of clausal nominatives in his dataset. Starting with the Nom-Dat order, he reports that 70 out of 73 nominatives (or 96%) are definite, compared to 55 out of 78 datives (or 71%). Thus, the Nom-Dat order clearly correlates with definite nominatives. As for the Dat-Nom order, 48 out of 94 datives (or 51%) are definite, compared to ten out of 12 nominatives (or 83%). However, since the Dat-Nom order so strongly correlates with clausal nominatives already (cf. supra), the speaker’s choice for this alternant has already been accounted for.

Additionally, our model also connects singular constituents with the second argument position, but their effect is considerably weaker (coefficient −.30) than for the remaining variables.

5.2 Double-NP configuration

The results of the second logistic regression model, which is solely based on the 131 tokens containing double NPs, are shown in Table 10. The c-index is .7, indicating a lower predictive power than the previous model (whose c-index is .794) although still in a range that would be considered acceptable in some fields. A lower c-index value is not unexpected given that the dataset on which the model is trained is smaller, and as such we consider it worth reporting on, especially since the model residuals do not indicate any particular issues.

Table 10.

Results of the logistic regression model for alternating verbs in the double-NP configuration excluding henta (N = 131). Significant p-values are in boldface

A comparison with the first model yields fascinating results. First, the current model singles out three variables that were also identified as strong predictors by the first model, i.e. nominative case marking, indefiniteness, and length (coefficients 1.16, −1.15, and −.54, respectively, corresponding to a 29% increase, a 28.8% decrease, and a 13.5% decrease in the likelihood of switching from second to first argument slot). The effects of case marking and length in particular seem to be somewhat mitigated compared to the first model, but both still generate highly significant p-values. Second, the weaker predictors pertaining to both animacy and number no longer appear to have any predictive power, and third, the person variable has been levelled out, as full NPs are self-evidently always third person.

The second logistic regression analysis essentially reveals two tendencies. As with the first analysis, it demonstrates the importance of nominative case marking for the first argument slot. Recall that the four alternating verbs under study yield a total of 71 Nom-Dat tokens and 60 Dat-Nom tokens. The difference between both these subsets is taken to be statistically significant, but, again, it remains to be investigated (i) whether a different sample would equally single out nominative case marking as a significant predictor, and (ii) whether other alternating verbs are equally sensitive to the effect of nominative case marking.

Second, the analysis shows that both indefinite and lengthy constituents have a proclivity for the second argument slot. The model neither reveals whether these variables are interrelated, nor whether they correlate with case marking. We discuss these issues further in the following subsection.

5.3 Interim conclusions

The current section has provided an in-depth statistical analysis of the alternating verbs duga, dyljast, endast, and nægja by means of two logistic regression models, one scrutinising the results across configurations (n = 800), another exploring the results for the double-NP configuration (n = 131). We have identified several well-known predictors from studies of word order as playing a role here, most notably animacy, indefiniteness, and length. Both models also single out nominative case marking as a predictor for the first argument slot.

Starting with the first three predictors, we have found animate (dative) constituents to prefer the first argument slot and indefinite and long constituents to prefer the second argument slot. These findings are evidently interesting in themselves, but the key question is whether there is a greater generalisation to be made. In fact, as is already mentioned in Sections 3 and 4.3 above, these three factors, animacy, indefiniteness, and length, are all proxies for topicality. That is, topical constituents tend to be animate rather than inanimate, and non-topical constituents tend to be indefinite and long rather than definite and short (Givón Reference Givón and Li1976:152, Croft Reference Croft2003:178–179, Arnold et al. Reference Arnold, Kaiser, Kahn and Kim.2013:406, Cristofaro Reference Cristofaro, Bakker and Haspelmath2013:74, Booth & Beck Reference Booth and Beck2021:11, inter alia). Thus, apart from their relevance as individual predictors, the factors in question also suggest that word order with alternating verbs is a derivative of discourse prominence (cf. Barðdal Reference Barðdal2001, Barðdal, Eythórsson & Dewey Reference Barðdal, Eythórsson and Kim Dewey2014, Reference Barðdal, Eythórsson and Kim Dewey2019).

Turning now to the last predictor correlating with the first argument slot, nominative case marking, the question arises whether the nominative is a factor in itself or whether the case marking is an epiphenomenon of other factors, such as, for instance, pronominality. Out of 331 nominative pronouns in first position, 42%, or 140 instances, are correlates, which show a clear preference for first position in Icelandic in any case (Rögnvaldsson Reference Rögnvaldsson2002, Thráinsson Reference Thráinsson2007:366–367, inter alia). A closer inspection of the [Pro-V-Pro] configuration, including correlates, reveals that nominative pronouns in first position show an average length of 1.1 words, while dative pronouns in second position have an average word length of 1.9. This suggests that the nominative-first effect with two pronouns is a consequence of length.

However, such a length effect with nominatives in first position is not found for the double-NP configuration. Instead, nominatives in first position turn out to be definite in 43 out of 71 instances (61%), which clearly makes them topical. What is more, a gauge at the 28 indefinite examples of nominatives in first position reveals that they are either specific or simply more topical than the dative, as is evident from example (18) below, despite being indefinite.

Observe that the indefinite nominative, rigningarvatn ‘rainwater’, in the Nom-Dat example above reiterates information mentioned earlier in the discourse, with both the former (accusative) rigningarvatn ‘rainwater’ and rignir ‘rains’ rendering the latter (nominative) rigningarvatn ‘rainwater’ highly topical, despite it being indefinite. This not only shows that topicality may not simply be reduced to one (or more) of its proxies but also that it has considerable explanatory power of its own. At least for the [NP-V-NP] configuration, it seems promising to explicitly factor in topicality as a predictor variable, because the effect of nominative case uncovered in the present study appears to be an epiphenomenon of a topic-first effect rather than a veritable nominative-first effect in itself.

6. Summary and conclusions

In this article we have succeeded in lending empirical support to the claim that behavioural subjects in Modern Icelandic are strongly tied to clause-initial position, regardless of whether these are marked in the nominative or the dative case. For this purpose, we have extracted 200 examples of 15 verbs each from the Icelandic Web 2020 corpus, thus amounting to a total of 3,000 tokens, all occurring with a dative and a nominative. The first class consists of five ordinary Nom-Dat verbs like hjálpa ‘help’, the second consists of five classical Dat-Nom verbs like líka ‘like’, and the third one of five alternating Dat-Nom/Nom-Dat verbs like nægja ‘find/be sufficient’.

The dataset has been annotated for nine variables: case marking, (pro)nominality, type of pronoun, referentiality, person, number, definiteness, animacy, and length. Our goal has been to provide statistical evidence of our alternating analysis for nægja-verbs, namely that these verbs alternate between two word orders, dative-before-nominative and nominative-before-dative, due to the fact that they instantiate two diametrically opposite argument structures, i.e. Dat-Nom and Nom-Dat.

We first establish a baseline with the help of ordinary Nom-Dat verbs, or hjálpa-verbs, and non-alternating Dat-Nom verbs, or líka-verbs, in configurations with two full NPs. It turns out that both these verb classes, hjálpa- and líka-verbs, realise the syntactic subject clause-initially 99.5% of the time. In contrast, for alternating Dat-Nom/Nom-Dat verbs, i.e. nægja-verbs, our findings generally confirm that the subject is the first argument of the argument structure, be that the dative or the nominative.

When nægja-verbs occur with two full NPs, their distribution is considerably less skewed towards one of the two argument structure constructions than with either hjálpa- or líka-verbs. There are, however, substantial differences found across verbs, with the Nom-Dat case frame attested more frequently than the Dat-Nom case frame, or in 72% vs. 28% of the cases. This number of 28% Dat-Nom is considerably higher than the 0.5% baseline for topicalisation documented with hjálpa- and líka-verbs above, and it is also noticeably higher than the 8% topicalisation documented by Callegari & Ingason (2021). This, in turn, rules out a topicalisation analysis of dative-before-nominative word orders with nægja-verbs. As a matter of fact, there is one particular verb, henta, that behaves unexpectedly in that it occurs consistently with the Nom-Dat linear order, irrespective of whether the two arguments are realised as full NPs or as pronouns. Thus, when recalculating the numbers for full NPs without the outlier, henta, the distribution amounts to 54% Nom-Dat vs. 46% Dat-Nom. Again, this rules out a topicalisation analysis of the dative-before-nominative order altogether.

Our analysis of nægja-verbs has also shown that their word order distributions are considerably more prone to pronominal influence than the ones attested for either hjálpa- or líka-verbs. More specifically, in contexts where the nominative is pronominal, nægja-verbs strongly prefer the nominative to precede the dative. However, contexts in which a dative pronoun enters into competition with a nominative full NP show the same word order distributions as the [NP-V-NP] configuration. These findings confirm the status of alternating Dat-Nom/Nom-Dat verbs as a syntactic class in their own right, distinct from non-alternating Dat-Nom verbs.

Finally, we have modelled the word order variation of nægja-verbs statistically. Recall that we removed henta from our dataset, as its frequencies were unexpectedly skewed. Across configurations, word order patterns are prone to a host of factors, including nominative case marking, indefiniteness, length, animacy, and person. For the double-NP configuration, the logistic regression analysis has identified nominative case marking, indefiniteness, and length as the most important predictors.

As it turns out, the factors underlying the variation in word order, both across configurations as well as in the double-NP configuration, converge nicely in that all these appear to reflect topicality in one way or another. After all, topicality is highly interwoven with animate, pronominal, definite, and short constituents. The only two exceptions to the topical-first trend that we have uncovered involve nominative case marking and person. Starting with person, third person arguments are generally relatively equally divided across the two positions, except for correlates, which occur in first position 95% of the time. Thus, we believe that the third person effect, detected in the logistic regression analysis for first position, is an epiphenomenon of this.

Turning to nominative case marking, we have also shown that in the double-pronoun configuration, which favours the Nom-Dat word order, the preverbal nominative pronoun is considerably shorter than the postverbal dative pronoun, indeed suggesting that the real issue here is length rather than case marking. Regarding the configuration with double NPs, 61% of the nominatives in first position are definite, again confirming the role of topicality. The remaining 39% of the preverbal nominatives are indefinite, yet an initial inspection of these instances shows that the majority are topical, although some are specific. This again validates the role of topicality, also for double NPs, confirming that the strongly observed nominative-first effect is an artefact of topicality.

To conclude, comparing nægja- and líka-verbs, we have shown that the former, but not the latter, have a choice between two alternating constructions, Dat-Nom and Nom-Dat. It turns out that well-worn pragmatic factors such as topicality govern the choice between the two diametrically opposite constructions with nægja-verbs. In contrast, with líka-verbs, the grammar does not provide this option to begin with, meaning that this verb class is confined to the Dat-Nom argument structure construction.

Regarding future research, the most pressing issue at this point is a comparison of the behaviour of alternating Dat-Nom/Nom-Dat verbs across the languages where such a class has been shown to exist, for instance Russian, Lithuanian, Romanian, Latin, and Ancient Greek (cf. Barðdal Reference Barðdal2023:Ch. 3 and the references therein). A particularly promising comparison is one between Modern Icelandic and Modern German, due to their close kinship. For a first attempt at such a venture, see Somers & Barðdal (Reference Somers and Barðdal2023), although a more fine-grained analysis of the relevant data is needed to improve our understanding of the factors at play.