Grammatical functions in transitive sentences are commonly assumed to express or to be associated with role-semantic properties of NP arguments (e.g., Foley 2011; Hörberg 2016: 9–10), henceforth referred to as participant roles. Participant roles concern the kind of involvement that the NP argument referents have in the event expressed by the sentence. Traditionally, notions such as agent and patient have been used to define participant roles (e.g., Fillmore 1968; Chomsky 1981). However, the degree to which participant roles are associated with agent or patient properties ultimately depend on the kind of event that is expressed by the main verb of the sentence. It is therefore not always possible to clearly and unambiguously assign a particular role to an argument (see Dowty 1991). Participant roles can instead be conceived as cluster or prototypicality concepts. In transitive sentences, the main verb may assign both agent and patient properties to the NP arguments, but the arguments will always be differentiated from each other in terms of possessing the most agent or patient properties, respectively. Participant roles can therefore be subsumed under two general categories, the Actor and the Undergoer role (see, e.g., Foley & Van Valin 1984; Dowty 1991; Van Valin & LaPolla 1997; Van Valin 2005; Primus 2006; Bornkessel-Schlesewsky & Schlesewsky 2009; Bickel 2010), the former being expressed by the subject, and the latter by the object. As such, the subject NP refers to the participant that is responsible for the activity expressed by the verb, and the object NP to the participant or entity that is affected by that activity.

Many verbs require an Actor that is volitional and/or sentient, and therefore human or animate, but do not put the same constraints on the Undergoer role (Van Valin & LaPolla 1997: 305; Dahl 2008). Animacy is therefore strongly correlated with grammatical functions in natural discourse. As a multitude of corpus studies show, subjects are much more frequently animate than objects (see Dahl & Fraurud 1996 and Dahl 2000 for Swedish; Kempen & Harbusch 2004 for German; Øvrelid 2004 for Norwegian; Bouma 2008 for Dutch). For instance, in his corpus of spoken Swedish, Dahl (2000) found that 93.2% of the subjects and 9.9% of the objects were animate. According to Dahl (2008: 145), animacy is “important for determining what can be said about an entity” in the sense that many predicates require an animate Actor argument. Animacy can therefore function as a semantic cue to grammatical functions, and thereby render overt formal marking of grammatical functions (such as case marking or other morphosyntactic information) redundant. The present study will test this assumption by investigating whether writers less frequently use formal markers of grammatical functions when grammatical functions can be determined on the basis of an animacy difference between two NPs.

Grammatical functions are also related to the information structure of individual sentences in a discourse. Information structure concerns how writers and speakers structure their sentences to their addressees, with respect to the information provided in the previous discourse and in the situational context. Most (but not all) declarative sentences are used to express an assertion about some referent that is assumed to be known by the addressee. This referent is expressed by the sentence topic (e.g., Reinhart 1981; Lambrecht 1994; Erteschik-Shir 2007). On some views, topics are seen as a means of storing information about the discourse model (Reinhart 1981; Erteschik-Shir 2007). Accordingly, topics serve as referential entries under which propositions are stored. Crucially, what constitutes the topic of a sentence depends on the discourse that the sentence occurs in. Two equivalent sentences may have different topics depending on the context. For example, as pointed out by Reinhart (1981), in a sentence such as

Max constitutes the topic when (2) is answered to the question Who did Max see yesterday?, but Rosa is the topic when it is the answer to Did anybody see Rosa yesterday?. The information of a declarative sentence that the writer or speaker assumes to be new to the addressee, on the other hand, is called the sentence focus (Lambrecht 1994; Gundel & Fretheim 2004; Erteschik-Shir 2007). The focus can be seen as new information that is added to the discourse model by the sentence at hand. What is considered to be focus of a sentence is also context-dependent. When (2) is answered to Who did Max see yesterday? then Rosa is focus, and when (2) is the answer to Did anybody see Rosa yesterday?, Max is focus. Focus is often conceived of as highlighting of information that stands in opposition with, or is contrasted against, a (possibly open-ended) set of information1 that potentially could constitute the focused information (e.g., Rooth 1992; Gundel & Fretheim 2004; Molnár 2006; Erteschik-Shir 2007; Krifka 2007; Molnár & Winkler 2010). In (2), for example, the focused referent (Max or Rosa) can be seen to be contrasted against a possibly open-ended set of other individuals. However, there are good reasons to differentiate the notion of focus from the notion of contrast, which involves contrasting some information against a limited number of contextually given alternatives (Molnár 2002; 2006; Molnár & Winkler 2010). Both topics and foci can be contrastive, and, as illustrated in Examples (3) and (4) from Prince (1984), English topicalisation is in some situations only valid when the topicalised information is contrasted against a limited number of contextually determined alternatives:

Whereas the positioning of the NP Annie Hall in the sentence-initial position is inappropriate in (3), in (4), where the NP Annie Hall is contrasted against all other Woody Allen movies, the NP may be positioned sentence-initially. Since Annie Hall is contrastive in (4), it may serve as the topic of the sentence. Contrast therefore appears to be an information structural notion that is orthogonal to topic and focus, but that shares features with both. As with focus, it involves the highlighting of information that it is in opposition to, and as with topic, it facilitates discourse coherence in that it relates information to the previous discourse context (Molnár 2002; 2006; Molnár & Winkler 2010).

Topical NPs typically refer to entities that are presupposed, either by virtue of being introduced earlier on in the discourse, or by being known by the interlocutors. Topical referents tend to be maintained through longer stretches of discourse through the use of anaphoric expressions (i.e., topic-chaining or topic-continuity, see Givón 1983; Erteschik-Shir 2007: 44–45; Engdahl & Lindahl 2014). NPs that are part of the sentence focus, on the other hand, more frequently refer to entities that are unknown by the addressee and that are new in the discourse. NP referents which are presupposed are sometimes called given, and referents which are discourse new and assumed to be unknown, are called new. Some suggest that topical arguments always are given (Erteschik-Shir 2007: 20), others that they at least have to be familiar to the addressee (Lambrecht 1994: 262). Focused arguments, on the other hand, tend to be new since they often introduce new referents into the discourse (Lambrecht 1994: 262).

This givenness distinction is generally assumed to concern the cognitive status or accessibility of NP referents in the discourse model, and therefore only to be indirectly related to linguistic expressions (Ariel 1990; Gundel et al. 1993; Gundel & Fretheim 2004). However, the givenness status of a referent is often reflected in the form of NP arguments (Ariel 1990: 69; Gundel et al. 1993; Lambrecht 1994: 105). For example, whereas NPs consisting of personal pronouns or definite nouns refer to entities that are active in the discourse, or assumed to be known by the addressees, indefinite NPs refer to new entities which have not been previously introduced into the discourse. The linguistic features definiteness, pronominality and person are therefore highly correlated with givenness. Definiteness is used to signal whether a referent is uniquely identifiable in the discourse, either by virtue of being anaphoric or through association (Lyons 1999). Definite pronouns tend to be anaphoric and refer to highly discourse prominent referents that have already been introduced into discourse. First- and second-person pronouns are inherently high in givenness in the discourse by virtue of referring to the speaker/writer, and the listener/reader, respectively.

Although there is no one-to-one mapping between subjects and topics (as Example (1) illustrates), sentence-initial subject NPs constitute the topic in most sentences, and object NPs are more frequently part of the sentence focus (Reinhart 1981; Lambrecht 1994; Erteschik-Shir 2007). There is therefore a correlation between grammatical functions and the givenness of NP referents, which in turn are reflected in their forms. Several corpus studies have therefore found that subjects are much more frequently definite, first/second person, and pronominal than objects (see Dahl 2000 for Swedish; Øvrelid 2004 for Norwegian; Bouma 2008 for Dutch; and Du Bois 2003 for a review of studies on differences in referentiality and givenness between subjects and objects). For example, Dahl (2000) found that 60.7% of the subjects but only 2% of the objects were first- or second-person pronouns.

The object-initial construction is generally assumed to be used when it is the object referent, that is, the Undergoer, rather than the subject referent, that is, the Actor, that is topical. Since topical NPs also tend to be highly discourse prominent, several corpus studies therefore show that sentence-initial objects frequently are high in discourse prominence in terms of prominality and definiteness, in comparison to objects positioned in their canonical position (See Weber & Müller 2004 for German; Øvrelid 2004 for Norwegian; Snider & Zaenen 2006 for English; Bouma 2008 for Dutch). The present study will investigate whether this also holds in written Swedish. More specifically, I will investigate whether topicalisation is the only motivation for positioning the object in the sentence-initial position, as implied by earlier investigations of object-fronting in written Swedish (Rahkonen 2006; Bohnacker & Rosen 2008; 2009; Bohnacker 2010). Before presenting these previous findings on object fronting in Swedish, I first give a summary of studies investigating whether speakers and writers adapt their productions in order to avoid ambiguities, and provide a short presentation of relevant aspects of Swedish grammar.

As illustrated in Section 1, object-initial sentences in Swedish are potentially ambiguous with respect to grammatical functions. A question that arises, therefore, is whether speakers and writers are more inclined to provide unambiguous information regarding grammatical functions in OVS sentences in comparison to SVO sentences or other unambiguous alternatives, such as the passive. The passive is similar to the object-initial construction in that it, too, can be used when it is the Undergoer rather than the Actor that is the topical participant (e.g., Givón 1979; Pitz 2006; Foley 2011). In the passive, the Undergoer is realised as an intransitive subject NP, and the Actor argument is optionally realised as the argument of a prepositional av (‘by’) phrase. The passive therefore differs from the object-initial construction in Swedish with respect to potential ambiguity. Whereas OVS sentences potentially are ambiguous with respect to grammatical functions, in the passive, the initial NP can only function as an (intransitive) subject which expresses the Undergoer role. It is therefore of interest to investigate whether writer’s choice between the object-initial construction and the passive is influenced by whether the grammatical functions of the NPs can be determined on other morphosyntactic or semantic information.

Results from previous studies regarding whether speakers and writers tend to avoid ambiguities during speaking and writing are mixed. On the one hand, several studies have either directly or indirectly provided evidence for ambiguity avoidance. For instance, some studies have found that the animacy difference between subjects and objects is more pronounced in object-initial sentences in both written and spoken language (Øvrelid 2004; Snider & Zaenen 2006; Bouma 2008; Bader & Häussler 2010). That is, speakers and writers more frequently use the potentially ambiguous OVS word order when the grammatical functions of the NPs can be determined on the basis of an argument animacy difference. It has also been found that speakers more frequently use morphological information regarding the argument functions in situations where they cannot readily be inferred on the basis of referential, semantic or plausibility information. In Korean and Japanese, overt realization of morphological case marking is optional in colloquial speech (Lee 2006; Kurumada & Jaeger 2015). In a corpus study of spoken Korean, Lee (2006) found overt nominative case marking to occur more frequently on subjects that are inanimate or low in discourse prominence in terms of person and definiteness. Overt accusative case marking was, on the other hand, more frequent for animate or discourse prominent objects (Lee 2006). In a sentence recall study, Kurumada and Jaeger (2015) showed that Japanese speakers more frequently use overt object case marking in transitive sentences with two animate NP arguments, and in semantically implausible sentences in which the assignment of participant roles is unexpected (e.g., The criminal arrested the police officer). These results suggest that speakers are inclined to provide additional morphological information regarding grammatical functions in situations where they are harder to determine on the basis of other information types. In a corpus study of written English, Temperley (2003) found the use of optional relative pronouns and complementisers to be more frequent in situations where their omission potentially would result in a syntactic ambiguity.2 Haywood et al. (2005) found participants engaged in a dialogue task to more frequently avoid preposition phrase attachment ambiguities in situations where such ambiguities would potentially confuse their addressees.3 Taken together, these results indicate that both speakers and writers actively avoid ambiguities in order to accommodate the understanding of their addressees.

Several other studies have, however, failed to find any evidence for ambiguity avoidance (see, e.g., Ferreira & Dell 2000; Arnold et al. 2004; Jaeger 2006; 2010; Roland et al. 2006; and Ferreira 2008 for a review). For instance, in a spoken sentence production experiment, Ferreira and Dell (2000) did not find the optional complementiser that to be used more frequently in cases where its omission would lead to a direct object-subject ambiguity.4 Similar findings were done by Jaeger (2010) on the basis of spoken corpus data. In contrast to the findings of Haywood et al. (2005), Arnold et al. (2004) did not find any tendency for participants in an on-line production experiment to avoid prepositional phrase attachment ambiguities. It has been suggested that ambiguities that appear to be problematic in isolated sentences might in most cases be unproblematic at the discourse level (Rahkonen 2006; Ferreira 2008). There is also some evidence suggesting that sentence level ambiguities in fact can be beneficial for communication when they can be resolved in context (Piantadosi et al. 2012).

Because of these contradicting findings, it is still unclear whether people do actively avoid sentence level ambiguities. This study therefore investigates whether there is a tendency for writers to avoid potentially ambiguous object-initial sentences such as (1a).

The sentence materials used in the study were collected from the treebank Svensk trädbank, a syntactically annotated version of the Stockholm-Umeå Corpus (SUC) (Gustafson-Capková & Hartmann 2006) with materials from the Talbanken (TB) corpus (Einarsson 1976a; b). The SUC corpus, which also was used by Rahkonen (2006), consists of 500 published texts from various genres that were published in between 1990 to 1993. The TB corpus consist of 85 professional prose texts from four different genres, written in between 1960 to 1971, originally compiled for the aforementioned study of Westman (1974). Most of the texts in these two corpora have undergone editing and proofreading. A list of the main genres of the corpora as well as the number of texts and words of each genre is shown in Table 2. The treebank is morphologically and syntactically annotated in Tiger-XML format.

Searches were conducted with TIGER search 2.1. (König et al. 2003), using the TIGER search query language (König & Lezius 2003). Data collection was done with the aim to minimise the constraints on the structural variation of the sentences. Following Jaeger (2011), search patterns were constructed to avoid false exclusions rather than false inclusions. False hits were excluded from the initial data at a later stage. The target sentences were SVO sentences (Example (5)), OVS sentences (Example (6)) and passives (Example (7)), including both main and subordinate clauses. Search strings allowed for NPs of any length and sentences with up to three auxiliary verbs in addition to the main verb (Teleman et al. 1999 (3): 278), one verb particle, and one sentential adverbial phrase.

Data exclusion was guided by finding sentences that 1) are declarative or passive, 2) contain argument NPs that refer to the participants involved in the event denoted by the sentence predicate, 3) contain argument NPs that serve as arguments of that predicate only,11 and 4) do not contain any idiomatic or lexicalised constructions.

A more detailed description of the search queries and the exclusion procedure, together with a list of excluded sentence types can be found in Appendix 1. The actual search strings are provided in Appendix 2.

A total of 14,500 sentences remained after data exclusion, out of which 12,949 sentences were SVO sentences, 869 were OVS sentences, and 682 were passives. In other words, only about 5.6% of all transitive sentences are object-initial. The rather low number of passives is due to the constraint that each passive must contain a prepositional av (‘by’) phrase.

Table 2 also shows the percentage of each sentence type of each genre. It is worth noting that there is quite a bit of variation in the proportion of sentence types across genres. However, genre differences is beyond the scope of the present paper and will not be discussed further. A more thorough investigation and discussion about genre differences in the frequency of use of subject-versus object-initial sentences can be found in Hörberg (2016: 105–107).

The NP properties under investigation in this study are givenness, animacy, definiteness, pronominality, and case marking (i.e., whether or not the NP at hand consists of a personal pronoun). The annotation of these properties were, to the extent that it was possible, based upon the morphosyntactic annotation available in Svensk Trädbank. In the following, I give brief descriptions and definitions of these properties as operationalised in the present study. A more detailed description of the annotation procedure of these properties can be found in Hörberg (2016: 79–88).

Givenness. As discussed in Section 2.1, the form of an NP argument is commonly assumed to correspond to a specific level of givenness or accessibility (e.g., Gundel et al. 1993). NPs were therefore annotated for givenness on the basis of their form. I used the four-level givenness scale shown in Table 3. The scale aims to capture the givenness/accessibility distinctions suggested by Prince (1981), Ariel (1990), and Gundel et al. (1993). It differentiates between new/type identifiable NPs, token identifiable NPs, familiar NPs and given NPs, as defined in Gundel et al. (1993). The categories largely correspond to the categories of Prince (1981), and to the NP forms of the Accessibility Hierarchy of Ariel (1990).

Conceptually, new/type identifiable NPs (e.g., indefinite NPs and a few grammatically definite NPs with generic reference) designate referents of a new type that are introduced in the discourse under the assumption that the referent in question is unknown to the addressee. The addressee needs to “create” and determine the type of referent without being able to determine which specific referent it pertains to.12 Token identifiable NPs (e.g., full names with modifiers and definite descriptions) designate a specific referent which is assumed to be unknown to the addressee. The addressee needs to be able to either “construct” a unique referent with no previous knowledge of it, or to identify the referent on the basis of the expression, either through inference or by the information provided in the expression. Familiar NPs (e.g., names, kinship terms and short definite NPs) designate a specific referent which is assumed to be known by the addressee. Given NPs (e.g., personal pronouns or short NPs with a proximal demonstrative) finally, designate a specific referent which is active in the discourse. The speaker therefore assumes that it is represented in the short term memory of the addressee, possibly at the center of attention.

In most of the analyses below, I use the dichotomous distinction between discourse new NPs, consisting of New/type identifiable NPs and token identifiable NPs, and discourse given NPs, consisting of Familiar and Given NPs. I will also differentiate between arguments as being high or low in discourse prominence, using discourse prominence as an umbrella term for the extent to which arguments possess properties associated with discourse givenness (e.g., such as being definite, pronominal and first- or second-person).

Animacy. A dichotomous distinction between animate and inanimate was used. NPs referring to humans and non-human animate beings, groups of animate beings, as well as organizations and politically organised geographic regions that depend on human organization in some sense (e.g., rock bands, political parties, municipals, countries etc) were annotated as animate. All other NPs were annotated as inanimate.

Definiteness. Definite and indefinite NPs were classified on the basis of the grammatical or semantic definiteness of either the head or the prephrasal element (e.g., the article, determiner or initial pronoun, Teleman et al. 1999 (3): 15), as annotated in the corpus. Indefinite NPs also included NPs without a grammatically definite/indefinite head or prephrasal element (i.e., indefinites with weak reference, see Teleman et al. 1999 (4): 175–176).

Pronominality. A two-way distinction between pronominal NPs and lexical NPs, consisting of proper and common noun NPs, was done on the basis of the word class of the NP head.

Case/person. As discussed in Section 2.3.2, the case marking system of Swedish is limited to personal and generic pronouns (see Table 1). As such, the case distinction coincides with a distinction between personal pronouns and other NPs. Swedish case distinguishes subjects from direct objects and obliques. All NPs that consisted of or was headed by the pronouns shown in Table 1 were classified as case marked/person – person here referring to personal pronouns generally. All others were classified as unmarked/non-person, non-person referring to all NPs not consisting of or being headed by a personal pronoun.

Text deixis. As discussed in Section 2.4.1, object-initial word order is especially common in Swedish when the object is text deictic, consisting of a third-person neuter pronoun or a neuter demonstrative (i.e., det and detta, Rahkonen 2006; Bohnacker & Rosén 2008), and therefore often refers to an earlier or an upcoming proposition of the discourse (Levinson 2004). The present study also investigates the distribution of text deictic NPs. NPs in which the head of the NP either consisted of a third-person singular neuter pronoun (i.e., det), a singular neuter demonstrative pronoun (i.e., detta, det här or det där) or the neuter relational pronoun detsamma were categorised as text deictic (see Teleman et al. 1999 (2): 5). All other NPs were classified as non-text deictic.

I conducted all statistical analyses that are presented below with the statistical software R (R Core Team 2016).13 The significance of distributional differences of some property between two or more categories (e.g., the difference in the proportion of Discourse Given NPs between subjects and objects in SVO sentences, see Table 4), were tested on the basis of χ²-tests, testing the null hypothesis that the property at hand has the same distribution across the categories (e.g., the proportion of Discourse Given NPs being the same for subjects and objects in SVO sentences). Multiple χ²-tests were therefore conducted on the same data set. Since the risk for a false positive (i.e., falsely rejecting a null hypothesis in favour of the research hypothesis) goes up with the number of tests that are conducted, all p-values from these χ²-tests were corrected for multiple comparisons on the basis of the method suggested by Holm (1979).

Logistic mixed effects modelling was done with the glmer() function in the lme4 package (Bates et al. 2015). All of the analyses, apart from the logistic mixed effects model analysis, are conducted across all genres of both corpora without taking any potential differences between genres into account.

As discussed in Section 2.1, several corpus studies have found that subjects are more frequently animate and high in discourse prominence than objects. This asymmetry stems from the functional difference between subjects and objects. Whereas subjects express the Actor role and in the prototypical case are topical, objects refer to Undergoers that commonly are focused and therefore new in the discourse. Differences in the distribution of animacy and referential properties between subjects and objects in prototypical SVO sentences was also investigated in the present study. The results are shown in Table 4.14 The table shows that subjects of SVO-sentences are animate (rather than inanimate) to a much greater extent than objects. Subjects are also much more frequently high in discourse prominence in terms of givenness, definiteness, pronominality and person than objects. These findings are rather uncontroversial and confirm the findings of previous studies (Dahl & Fraurud 1996; Dahl 2000; Du Bois 2003; Kempen & Harbusch 2004; Øvrelid 2004; Bouma 2008).

As discussed in Section 2.4.1, object fronting in Swedish is commonly assumed to be used when the object is topical and therefore high in discourse prominence (Teleman & Wieselgren 1970; Teleman et al. 1999 (4): 432; Rahkonen 2006; Bohnacker & Rosen 2008; 2009; Bohnacker 2010). In line with this, Rahkonen (2006) found sentence-initial objects to more frequently be high in discourse prominence than post-verbal objects. Both Rahkonen (2006) and Bohnacker & Rosen (2008) further found object fronting of text deictic objects to be particularly frequent. The findings of the present study in part confirm these results. The distribution of referential properties of both sentence-initial and post-verbal object NPs is shown in Table 5. Objects are more frequently high in discourse prominence when positioned in the sentence-initial position in comparison to when they occur in post-verbal position in terms of discourse givenness, definiteness and pronominality. Sentence-initial objects are also much more frequently text deictic in comparison to objects positioned in their canonical, post-verbal position.

These findings indicate that objects more often are positioned sentence-initially when they are highly discourse prominent, and therefore refer back to either a referent that has been previously introduced into the discourse, or a previously mentioned proposition. In particular, text deictic objects are almost exclusively positioned sentence-initially and very seldom occur in post-verbal position.

However, object fronting does not only take place when the object is high in discourse prominence. Sentence-initial objects are less frequently discourse prominent than sentence-initial subjects. This is illustrated in Table 6, which compares the distribution of referential properties of initial objects to that of initial subjects. Initial objects are less frequently discourse given, definite and new than initial subjects. If the only motivation for object fronting was to emphasise that it is the object, rather than the subject, that refers to the most discourse prominent referent of the two NPs, one would expect initial objects and subjects to be discourse prominent to an equal extent. In fact, as will be discussed below in Section 4.4, subjects are more frequently high in discourse prominence when positioned in the post-verbal position than when they are located in their canonical, sentence-initial position (see Table 8). Consequently, it is not the relative prominence difference between the NPs (independent of their functions) that determine their positioning either, such that it always is the most discourse prominent NP that is positioned sentence-initially. This is illustrated in Table 7. The table shows the percentage of SVO- and OVS- sentences, respectively, in which the initial NP either outranks, is equally ranked or is outranked by the second NP in terms of the prominence feature at hand. In SVO-sentences, the sentence-initial subject (NP1) most frequently outranks or is equally ranked to the post-verbal object (NP2) in terms of all prominence properties (givenness, definiteness and pronominality). In OVS-sentences, on the other hand, the post-verbal subject (NP2) more frequently outranks or is equally ranked to the sentence-initial object (NP1) in terms of discourse prominence. These findings show that object fronting is not only used when the object is high in discourse prominence.

In order to shed further light on functional differences between object- and subject-initial sentences, I also investigated differences in the distribution of NP properties between post-verbal and sentence-initial subjects. Table 8 shows the results. Subjects of transitive sentences are more frequently high in discourse prominence in terms of pronominality and person when positioned in the post-verbal position. Transitive subjects are also more frequently animate when positioned in the post-verbal position. An important contribution to these distributional differences is that post-verbal subjects more commonly occur as personal pronouns (such as in, e.g., Example (12)). Subjects therefore more often refer to highly discourse prominent first-, second- or third-person participants when occurring in the post-verbal position. This finding is consistent with the suggestion that, at least in object-initial sentences with discourse new objects, the remainder of the sentence tend to be “informationally light” and predictable. As a result, post-verbal subjects can be expected to be encoded with personal pronouns, which commonly refer to discourse participants that are highly given and therefore predictable in the discourse.

However, it is not the case that the predictability of the post-verbal subject (in terms of being realised as a personal pronoun) depends on the givenness status of the sentence-initial object. Table 9 shows the proportion between marked/person and unmarked/non-person post-verbal subjects, differentiated on the basis of the givenness of the sentence-initial object NP. Thus, Table 9 only includes data from OVS sentences. The table illustrates that the distribution of a personal pronoun and non-personal pronoun post-verbal subject NPs is roughly the same over the givenness levels of the sentence-initial object. This observation was confirmed by a χ²-test of independence which found no support for a relationship between the NP2 case marking/person and NP1 givenness in object fronted sentences, χ² (3) = 5.45, p = 0.28. The strong tendency to use a personal pronoun – and therefore highly predictable – subject NP in object fronted sentences is therefore independent of the givenness of the sentence-initial object. This suggests that it is not only when the fronted object is discourse new that writers aim to keep the remainder of the sentence predictable and therefore easier for the reader. This rather seems to be a more general tendency of all object-initial sentences.

I now turn to the issue of whether writers are inclined to avoid potentially ambiguous object-initial sentences, first with respect to whether other formal, morphosyntactic information about the argument functions is available, and then with regard to whether the grammatical functions can be determined on the basis of an argument animacy difference. This is done by investigating the distribution of formal, morphosyntactic, means of disambiguation, on the one hand, and the distribution of animacy-based means of disambiguation, on the other, in terms of how they differ between object-initial sentences, subject-initial sentences and passives.

As discussed in Section 2.1, word order variation is generally assumed to be motivated by information structural considerations. In particular, direct object fronting in Swedish is usually considered to be used when the object (rather than the subject) is topical and therefore high in discourse prominence (Teleman & Wieselgren 1970; Teleman et al. 1999 (4): 432; Rahkonen 2006; Bohnacker & Rosen 2008; 2009; Bohnacker 2010). Rahkonen (2006: 45) claimed that OVS sentences in which the object is low in discourse prominence do not “represent the typical information structure of OVS sentences”. Although the findings of the present study show that objects more frequently are high in discourse prominence – and in particular text deictic – when positioned sentence-initially, sentence-initial objects that are discourse new are quite common. In fact, sentence-initial objects are significantly less frequently discourse given than sentence-initial subjects. These findings show that the object-initial word order not only is used when the object is topical and therefore discourse prominent. They also to some extent call into question Bohnacker & Rosén’s (2008) claim that the preference to start a sentence with highly discourse prominent, given or thematic information is stronger in Swedish than in German.

In line with Molnár & Winkler (2010), I instead suggest that object fronting often is used when the object is contrastive, that is, that it is contrasted against a set of alternatives that have been introduced in the previous discourse. Following Molnár & Winkler (2010), I assume that contrast is orthogonal to topic and focus, in the sense that both topical and focused NPs can be contrastive.

I also suggest that object fronting is used to introduce new topics into the discourse, that is, to mark a discourse topic shift (James 1995). In such cases, a discourse new object NP is positioned in the sentence-initial position in order to emphasise that it contrasts with the previous discourse topic. Object fronting is in such cases used to establish discourse coherence between the old and the new discourse topic (cf., Molnár & Winkler 2010). Apart from the “informationally heavy” sentence-initial object NP that is discourse new, such constructions tend to be “informationally light” in that they often contain a semantically weak main verb, such as a possessive verb, and a post-verbal subject that is given in the discourse and therefore high in discourse prominence. Further indirect evidence for this account comes from a study by Qian and Jaeger (2011). They found the information content of full sentences, that is, the in-context predictability of whole sentences, to be negatively correlated with the strength of the discourse topic shift in those sentences, as estimated on the basis of topic modelling. In other words, speakers appear to avoid unpredictable information in sentences with new discourse topics, in order to keep the level of information flow as constant as possible throughout the continuation of the sentence (Shannon 1948; Fenk-Oczlon 2001; Jaeger 2010). This account is also corroborated in the Grammar of the Swedish Academy. According to it, the meaning of an OVS sentence following a sentence-initial object that is discourse new and focused is likely to be predictable, so that the comprehender can infer it upon encountering the object (see Teleman et al. 1999 (4): 432).

In this study, I investigated whether writers tend to avoid potentially ambiguous OVS sentences which either lack formal information about grammatical functions – case marked personal pronouns, auxiliary verbs, and verb particles – or in which those functions cannot be determined on the basis of animacy.

All three types of formal information regarding grammatical functions were found to occur most frequently in OVS sentences, followed by SVO sentences, and least frequently in passives. This is exactly the kind of pattern that can be expected if writers actively adapt their language in order to avoid potential ambiguities. Sentences with OVS word order almost always contain one or several formal markers that can be used to unambiguously assign grammatical functions. In OVS sentences, readers cannot rely on word order, and other formal means are often required. SVO-sentences less frequently occur with formal markers, although they quite commonly are used when the argument functions cannot be determined on the basis of animacy. In SVO sentences, readers can resort to word order when no other means are available. Formal markers are therefore not crucial for interpretation. Here, additional formal markers serve to assist the interpretation process, which is of particular importance when grammatical functions cannot be determined on the basis of animacy. In passives, on the other hand, additional formal markers are in fact quite rare. Even in passive sentences with two animate arguments, they occur in less than 50% of the occurrences. This is because the passive, which is syntactically intransitive, contains several unambiguous morphosyntactic markers to grammatical functions (i.e., the passive marker and the av preposition). Additional means will therefore most likely not facilitate interpretation further, independent of whether the argument functions can be determined on the basis of an argument animacy difference.

These results are in line with those of Rahkonen (2006) who found OVS sentences to occur more frequently with case marked personal pronoun subjects than both SVO sentences and OSV object relative clauses. Although Rahkonen’s findings indicate that writers tend to avoid OVS sentences without a case marked subject, Rahkonen (2006) instead suggested that they stem from a dispreference for using passive sentences in which the Actor argument is realised as a personal pronoun (i.e., as the argument of a prepositional av (‘by’) phrase). Accordingly, writers tend to resort to the object-initial construction whenever the Actor needs to be realised as a personal pronoun, in order to avoid using a prepositional av (‘by’) phrase with a personal pronoun argument. The results of the present study speak against this interpretation, since they show that it is not only case marked personal pronoun Actors that occur less frequent in passive sentences, but also auxiliary verbs and verb particles. Since all of these information types serve as formal cues to grammatical functions in Swedish, the most plausible interpretation is instead that writers tend to resort to the passive construction – which is structurally unambiguous with respect to grammatical functions – whenever the target sentence at hand will lack any of these formal cues to grammatical functions (see Pitz 2006 for similar ideas regarding the use of the passive in Norwegian).

It should be stressed, that the use of the passive is not only motivated by a need to topicalise or foreground the Undergoer. As shown in Section 4.5.2, the Actor argument of passives is much more often low in discourse prominence than in both SVO and OVS sentences. This is possibly because the passive also serves to background the Actor argument (Haspelmath 1990; Foley 2011). This is likely to more frequently be done with Actors that are low in discourse prominence. The passive construction have also been claimed to express inactivization of the event or situation, in the sense that the Actor no longer is responsible for the activity (Haspelmath 1990), as well as to modify the aspect of the event (Landén & Molnár 2003). Thus, the choice between the passive and the object-initial construction is obviously not only a choice that is driven by a motivation to avoid potential ambiguities. Also other information structural and semantic considerations come into play.

The present study also found that OVS sentences less frequently contain two animate NP arguments, than both SVO sentences and passives. This finding indicates that writers tend to avoid using the OVS word order when the argument functions cannot be determined on the basis of animacy. It was further found that all three types of formal markers are used more frequently in sentences with two animate arguments than in sentences with at most one animate argument. This is not only the case in potentially ambiguous OVS sentences, but also in both SVO sentences as well as in passives. Writers therefore seem to be more inclined to provide formal markers to grammatical functions in cases where they might not be able to be determined on the basis of animacy.

Similar findings were done by Rahkonen (2006), who found case marked subjects in OVS sentences to be somewhat more frequent in sentences with two animate arguments in comparison to sentences with one animate argument. Rahkonen (2006) did not, however, interpret these findings as being driven by a motivation to avoid ambiguities. He argued that although OVS sentences with two animate arguments might be ambiguous with respect to their argument functions in isolation, this is very seldom the case in context. Semantically ambiguous sentences are therefore likely to be unproblematic to interpret in their discourse contexts, and there is no need to provide additional morpho syntactic information regarding the argument functions. However, Rahkonen (2006) failed to appreciate that language comprehension most likely involves interpretation at two or several different stages, and that the discourse context is taken into account only at the final stage of interpretation. More specifically, several models of language comprehension assume that interpretation of local structure on the sentence level – at which point participant role assignment occurs – temporarily precedes interpretation of global structure on the discourse level (e.g., Bornkessel-Schlesewsky & Schlesewsky 2006; Friederici 2011). Although the argument functions might be able to be determined on the basis contextual information during the final stage of language comprehension, the comprehension process is still likely to be hampered at the earlier stage of local structure building. Contrary to Rahkonen’s claims then, the use of formal markers in potentially ambiguous OVS sentences is therefore motivated by the need to facilitate language comprehension during local structure building, even if those sentences are unambiguous in their discourse context.

Overall then, the findings of the present study provides rather compelling evidence for the idea that writers are inclined to actively avoid using object-initial sentences when this might result in an ambiguity. As such, these findings provide additional evidence for the hypothesis that language producers – in this case writers – are inclined to actively avoid ambiguities. It should be noted that this tendency, which is probabilistic rather than absolute, most likely is driven by a trade-off between a motivation to avoid using redundant information, on the one hand, and a motivation to make the message informative enough for the reader, on the other. As mentioned in Section 2.2, Kurumada & Jaeger (2015) found Japanese speakers to be more prone to use overt object case marking in subject-initial transitive sentences where the function of the object argument was harder to infer on the basis of animacy or plausibility information. However, in all sentences in their study, the sentence-initial subject was unambiguously case marked, so that the argument functions could readily be assigned already upon encountering the subject. The results of their study therefore seem to reflect a tendency for speakers to balance their production efforts between avoiding redundant information, on the one hand, and providing enough information in order for the listener to effectively comprehend the message, on the other. There is no reason not to assume that the production of written language is influenced by similar principles. That is, writers of Swedish might be more inclined to avoid the object-initial construction when no other means for identifying grammatical functions – such as case marking or animacy information – will be available in the target sentence.