1 Introduction

We examine here the distribution of morphological case (e.g., accusative, genitive, and dative) among object complements of monotransitive verbs in Classical Greek (CG). Accusative-marked objects are generally deemed to be direct objects (DO), while dative- and genitive-marked complements are sometimes treated as syntactically or semantically distinct, sometimes being considered as objects bearing exceptional/semantic/quirky case, sometimes being analyzed as indirect objects (IO), and sometimes having the status of obliques. Restricting our focus to verbs that have a single complement, we can observe that the distribution of accusative- (ACC), genitive- (GEN), and dative- (DAT) marking on these in CG is confounding. CG productively places DAT and GEN NP s alongside ACC NP s as a potential singular complement to monotransitive verbs, allowing them to occupy what would normally be thought of as the direct-object position, but for their GEN and DAT case-marking (see ex. 1–2).

(1)

καὶ	πολλ-ῶν	ἦρξε	τριήρ-ων
and	many-gen.pl	command.aor.act.3s	war.ship-gen.pl

‘and he commanded many war ships.’

Isocrates, Against Alcibiades 1.14.36

(2)

ὅτι	μετ᾽	ἀνδρ-ῶν	νῦν	ἀρίστ-ων…	τοῖς πολεμί-οις	μάχεσθε
that	with	men-gen.pl	now	best-gen.pl	the enemy-dat.pl	fight.2pl.ipfv.act

‘that you are now with the best men … fighting the enemies.’

Lysias, Against Eratosthenes 12.94

Not only is there a class of verbs in CG which productively take GEN and DAT direct objects, but there is also a class of verbs which may optionally alternate between ACC and GEN. Therefore, the attested verb-classes are as follows: i) ACC-Object verbs, ii) GEN-Object verbs, iii) DAT-Object verbs, and iv) ACC/GEN-Object verbs. The present research explores the first three, seeking to understand what is communicated in the selection of either an ACC-, GEN-, or DAT-marked DO.

Greek grammars generally tend to analyze GEN and DAT objects as traditional ‘indirect objects’ and apply these explanations across the whole of Ancient Greek (AG, 900 BCE–500 CE). However, it is clear that a more complex approach is required (and one that considers the distribution of exceptional object case in a more circumscribed time period), if we are to fully understand the range of syntagmatic roles assigned to GEN and DAT objects. Here, Grimm’s (2011) point stands out: “despite the frequency with which labels such as ‘dative’ are used, often what is termed a dative in one language only shares a few of the functions of a case in another language with the same label dative” (517). Though the AG GEN and DAT nominals certainly bear semantic roles that are similar to analogously marked nominals in other languages—e.g., GEN complements sometimes being partitive and DAT sometimes serving to mark locatives—there is more to be said about them both, particularly when they mark complements of transitive verbs. As an effort in this direction, we investigate this phenomenon through a corpus-driven analysis that seeks to reveal the precise semantic roles of DO s marked with one of these three cases, ACC, GEN, and DAT, tying the nuances of difference to the meaning of the verbs that select them.

The data collected and analyzed in the present study confirms that Luraghi’s (2010) Transitivity Hierarchy makes correct predictions regarding the case-marking of some objects, such that objects of high transitivity verbs get ACC case and those of medium transitivity verbs get GEN case. In this regard, we agree with Luraghi that high transitivity correlates with Change of State [COS] of the direct object, and that ACC objects undergo a COS, but GEN objects do not. In explaining the distribution of DAT objects, however, we find reason to believe that a Transitivity Hierarchy cannot fully account for them, and that additional factors are involved. With respect to these, it turns out that verbs’ selection of DAT objects is best explained in terms of Blume’s (1998) interaction verbs.

In accounting for the distribution of object case, we provide a more insightful analysis of the conditions under which GEN case and DAT case are assigned to the objects of monotransitive verbs than that which is typically given in traditional grammars. These grammars generally and inaccurately tend to claim that all GEN objects are partitive and mostly ignore the distribution of DAT objects (Smyth & Messing 1956, Hansen & Quinn 1992, van Emde Boas et al. 2019).

This study is intended to present an initial investigation into the questions here posed, focusing on the behavior of classes (i–iii). Given what we have determined about these classes, we propose that these findings may subsequently be taken up to explore the final, alternating, ACC/GEN class of verbs, though we do not yet tackle that subject herein. Section 2 reviews prior work on this topic, primarily that of Luraghi (2003, 2010), and situates the present research within the framework that she lays out. Section 3 then presents the data domain and methodology that figure in our study. Section 4 introduces Dowty’s (1991) conception of Proto-Roles and discusses the usefulness of these, particularly in terms of the goals of the present study. Section 5 investigates the behavior of the non-alternating classes (i.e. verbs that consistently select one case for their direct objects), presenting first some preliminary (pre-statistical) observations, followed by statistical results of our examination of the distribution of ACC, GEN, and DAT case-marking on objects. We present these first through the lens of Luraghi’s featural analysis, and then using a revised set of features, and a revised approach which we contend yields superior results. A conclusion, in section 6, provides some directions for future research, given these findings, especially regarding the final class of ACC/GEN alternating verbs, and also makes suggestions for broadening this research to include a wider range of the historical stages of Ancient Greek.

2 Review of previous research

Outside of Classical Greek (CG) pedagogical grammars, there have been few linguistic treatments of CG case assignment or argument structure, particularly in reference to the atypical behavior of dative and genitive direct objects. A largely general description of the distribution of ACC, DAT, and GEN case is provided in Luraghi 2003, with a more specific description of their distribution on direct objects in Luraghi 2010, equating this with semantic “levels of transitivity”. Lavidas (2009) also provides a comprehensive, diachronic overview of verb transitivity, from Homeric to Modern Greek.

According to Luraghi (2010: 64), ACC objects (those whose semantic role is characterized as “theme”) are “wholly affected patients, which undergo a change of state” and which canonically occur with the most highly transitive verbs. It is claimed that ACC case is nearly always assigned in highly transitive contexts, and that GEN case is exceptionally assigned to objects having a partitive interpretation. In Luraghi’s terms, GEN objects that are complements of highly transitive verbs are “partitive genitive” and thereby seen to be less affected (although still “somewhat” affected) than are ACC objects. This analysis mirrors the explanation put forth by most pedagogical grammars, claiming that GEN objects are partitives and are relics of GEN indirect objecthood (Hansen & Quinn 1992, van Emde Boas et al. 2019). Verbs subcategorized for DAT objects are, according to Luraghi, even less transitive than those subcategorized for ACC or GEN objects.

Luraghi’s proposed Transitivity Hierarchy, intended to account for the mapping of verb meaning to case, distinguishes three levels of transitivity: High transitivity (HT), medium transitivity (MT), and low transitivity (LT). These levels are distinguished on basis of (i) whether the verb’s object undergoes a Change of State (COS), and (ii) whether the object is affected.

Luraghi (2003, 2010) presents her result as an analysis of Ancient Greek (AG), a corpus spanning some 1400 years (encompassing the period from 900 BCE to 500 CE). However, because her analysis utilizes data principally from Classical Greek (~500–300 BCE), it runs the risk of trying to capture generalizations about varieties of Greek that are separated by many centuries, if not a millennium.¹ Utilizing a range of data that is as diachronically distant as Old English and Modern English can lead to a conflation of meaning-case correlations that may not be consistent across time. Accordingly, we first aim here to test Luraghi’s Hierarchy against a much smaller stretch of Ancient Greek, i.e., over only Classical Greek texts. Since our study examines a corpus spanning only 2 centuries (~500–300 BCE), it is likely to be systematically more consistent than one based on a period of language which spans over a millennium.

Second, as explored in further detail below in Section 5, we find reason to doubt the reliability of using Affectedness as a feature in this hierarchy, depending as it does on the same kinds of semantic changes better marked by COS alone. Finally, where Luraghi (2010) only considers direct object semantic properties in framing the Transitivity Hierarchy (COS and Affectedness), the present study expands the range of semantic properties considered in the analysis by examining all the Subject and Object Proto-role Properties proposed in Dowty (1991).

Much like Luraghi, Lavidas (2009) also notes the importance of COS as a property of objects that correlates with ACC case-marking and indicates enhanced transitivity. Lavidas also provides an overview of language change across the historical development of Greek, with a particular focus on how verbs’ meanings condition object case. Lavidas 2009 is specifically focused on the mechanisms by which verbal transitivity and meaning change over time, providing an overview of verbal transitivity across the development of Greek. Therein it is claimed that verbs taking COS object arguments represent the “central or prototypal class of transitive verbs,” (p. 2), in agreement with Luraghi (2003, 2010). Pivotally, Lavidas also recognizes the important role that verb meaning plays in determining the morphological and syntactic structures that accompany the use of any verb. Lavidas is concerned in particular with understanding changes in DO case-marking across periods of AG—especially with respect to the semantic roles that correlate with GEN and DAT DO case-marking. He notes the tendency toward an extension of ACC case, which is gradually extended to objects with semantic roles previously marked by GEN, DAT, or various prepositions. It should be noted that, unlike Luraghi, Lavidas does not stipulate any hierarchy for Greek verbal transitivity, taking transitivity as he does not to be “a unified phenomenon” (p. 9) which can be described in such terms (even though he does take COS verbs to be the most prototypically transitive).

Of case-marking in Classical Greek. Lavidas does not address the plethora of available case-markings for DO s, saying only (p. 64) that

… the accusative case was the usual case of the subjects of infinitives and the typical case for direct objects; therefore, the accusative case was a particularly frequent case. Some verbs idiosyncratically governed objects in other cases, of course. The dative case marked indirect objects, arguments that are interested in the action described by the verb, possession with the verb ‘to be,’ agents with some passive verbs, or partial effect on the object.

So, while Lavidas (2009) provides a foundational description of diachronic changes to transitive verbal constructions, he does not contribute much to an understanding of why CG marks particular DO s with GEN and DAT case. In fact, we find from our own data that his description of DAT case-marked objects fails to adequately describe the range of cases wherein DAT marking of objects does arise (we also note the absence of any description by Lavidas of GEN marked objects). Furthermore, in his description of CG’s DO cases, Lavidas falls victim to the mistake of taking an aerial approach that spans several periods of Ancient Greek, rather than pursuing an examination of specific periods. As before, this approach often results in broad generalizations that obfuscate important facts about any given stage.

Nonetheless, Lavidas (2009) provides a fundamental understanding of the construction of transitivity across the temporal cline of Greek, and Luraghi (2003, 2010) provides a more specific look at the way this phenomenon informs case-selection properties across Ancient Greek. The present research then builds on these works by applying these principles to an even more specific data set, in analyzing how such processes can be understood in terms of Classical Greek only.

3 Data domain and data collection methodology

While previous considerations of CG case have attempted to generalize across all of Ancient Greek, we concern ourselves only with Classical Greek (c. 500–300 BCE). By doing this, we can examine a data domain that is internally consistent and avoid potentially confounding semantic changes that might arise over a longer time span. Our study is also restricted to works of prose, excluding poetic and theatrical works. This is done in order to avoid the anomalies that might arise in works that are stylistically fashioned (and distorted) by meter and which often involve violations of canonical grammatical form.² Prose works, e.g., historical and rhetorical texts, tend to adhere more faithfully to grammatical norms and provide more reliable indicators of DO case-marking parameters of CG. The following authors then make up the source material for our investigation: Herodotus (c. 484–425 BCE), Thucydides (c. 460–400 BCE), Lysias (c. 445–380 BCE), Isocrates (c. 436–338 BCE), Xenophon (c. 431–354 BCE), Plato (c. 428–348 BCE), Aeschines (c. 389–314 BCE), Demosthenes (c. 384–322 BCE), and Aristotle (c. 384–322 BCE). While the writings of these authors do involve a small range of dialect variation, these dialectal differences are not sufficiently large as to impact the canonical case marking patterns of verbs’ objects.

To preclude confounds that might arise from an examination of multiple classes of verbs (e.g., ditransitives and experiencer verbs), the study was limited to an examination of sentences containing monotransitive predicates in active constructions with a single direct object complement. Accordingly, psychological predicates were excluded from consideration, since the subject-object relationship of verbal complements in these constructions tends to be rather fluid, as in example (3). Filip (1996: 3) asserts, apropos of this concern, that “linking in the domain of psychological predicates is problematic if we assume that there is a direct and uniform association between thematic (or lexical semantic) arguments and syntactic arguments”.

(3) a. Snakes frightened Sam.

b. Sam feared snakes.

Propositional attitude verbs were also excluded from the study, since these predicates typically take clausal (propositional) complements rather than direct object NP s, and do not usually assign object case, as in example (4).

(4) Tom believes [Sam fears snakes].

Data that conformed to the above-mentioned constraints was collected using the Perseus Online Digital Library (Crane 2005–2019). We note that this corpus is not searchable by grammatical marking, only by lexemes. Thus, once verbs were identified in preliminary research, they were then searched for in the corpus and items that were both (i) attributable to Classical Greek prose authors and (ii) active monotransitive constructions were added into the data set for analysis. Each verb identified as selecting an ACC-, GEN-, or DAT-marked direct object was searched for in all of its possible forms (i.e., in all six principal parts and inflections). Thirty-seven verbs were initially identified, six of which were excluded for a lack of tokens, resulting in the analysis of 300 instances of direct object sentences across 31 verbs.³

4 Theoretical background: argument structure

In our study, we reassess the properties used in Luraghi’s Transitivity Hierarchy, and find both that Affectedness is not a sufficiently precise feature on which to base such a hierarchy and that the properties of a verb’s subject can be just as important as its object in determining the transitivity of the clause as a whole. In framing our methodology, we have thus been led to consider a wider range of semantic features (and only semantic features that are precise enough to measure). We have also determined to examine, simultaneously, the semantic properties of both the subject and object arguments of the transitive clause, to ensure that subject properties are fairly considered alongside the object ones.

In this regard, our coding of subject and object semantic features cannot rely on the traditional inventory of theta-roles as they are usually conceived of in Generative Grammar. These thematic/semantic roles (i.e., theta-roles or θ-roles), though taken by some to be grammatical primitives encoding the semantic properties of a verb’s arguments and part of the interface between lexical and syntactic representations, generally fail to be general or precisely defined enough for a project of this sort. The typical thematic role categories of agent, patient, goal, source, experiencer, and theme leave far too many open questions about what they actually represent, whether they are primitive properties or composites, and why their application to arguments is so profoundly affected by context.

To date, there is no reliable, completely consistent inventory of thematic roles from which one might build a coherent and complete lexical semantic framework for a language, and the inventories most commonly invoked do not adequately account for the full range of observable data (Jackendoff, 1987; Dowty, 1991). There is disagreement on the commonly proposed theta role inventories in terms of the number of roles, the role labels, and the linkage of these roles to grammatical functions. Jackendoff (1987: 371), for instance, argues against the unnecessary conflation of syntactic arguments and semantic theta-roles, referring to the reliance on theta-roles as a “thinly disguised wild card to meet the exigencies of syntax”. Eschewing any isomorphism between syntax and semantics, Jackendoff (p. 372) suggests that the three “autonomous levels of structure: phonological, syntactic, and semantic/conceptual” each have their own “characteristic primitives and principles of combination, and [their] own organization into subcomponents”. Dowty (1991) extrapolates on such primitives in his explanation of “thematic proto-roles,” and departs from the strict view of theta-roles as primitives and as one-to-one, absolute correlates of syntactic roles.

For this reason, we appeal to accounts of lexical semantic properties that are both more nuanced and more precisely measured. One of the more carefully constructed accounts of these properties is found in Dowty (1991), wherein an inventory of “Proto-role Properties” (i.e., a set of entailments) is proposed. Rather than relying on discrete theta-roles, Dowty (1991) proposes a group of semantic properties that can serve as ‘tendency-makers’ that push semantic participants into grammatical slots based on those properties. This relationship is one that is more fluid than traditionally rigid theta-role assignments. Rather than giving a list of autonomous theta-role primitives, Dowty (1991) posits semantic role prototypes, in which there are different degrees of membership in two primary categories—(i) proto-agent and (ii) proto-patient. Proto-roles serve as entailments proceeding from the relation of a verb with a particular argument of that verb (see Table 2).

To instrumentalize these properties, Dowty proposes an argument selection principle, which is again more fluid than previous conceptions of syntactic and semantic argument structure. This principle states that “the argument for which the predicate entails the greatest number of Proto-Agent properties will be lexicalized as the subject of the predicate; the argument having the greatest number of Proto-Patient entailments will be lexicalized as the direct object” (pp. 576). With this selection principle in mind, Ackerman & Moore (2001) provide an extension to Dowty’s theory, supporting its use as a theoretical apparatus, and providing the Paradigmatic Argument Selection Principle, described by Hovav (2003: 648) as dealing “with cases in which there is an alternation in the encoding of a specific argument, accompanied by a change in interpretation”. Ackerman and Moore’s (2001) approach, in which sets of arguments are compared, further emphasizes the importance of both subjects and objects simultaneously determining the relevant roles in the determination of object case-marking. Thus, Dowty’s Proto-role theory and research emanating from it serves as the foundation on which the present study rests, taking relevant properties of both the subjects and objects in CG constructions to yield an analysis of DO case-marking.

Utilizing Dowty’s Proto-roles, a number of researchers have delved more deeply into their significance and use, and have endeavored to make them more precise. Among those who have worked to make these proto-roles more explanatory are Ackerman & Moore (2001), Hovav (2003), Næss (2007), Beavers (2010, 2011), and Grimm (2011).

Næss (2007) takes a closer look at Dowty’s (i-a) volitionality and (i-b) sentience, showing both that sentience covers a wider range of cases than volitionality, and that volitionality is more nuanced and context dependent. While both are deemed to be inherent properties of agents, it is volitionality which figures as a more useful distinction and more critical to the concept of agentivity. Essential to this, as Næss notes, is that volitionality is preserved under negation, while sentience need not be; cf. John didn’t listen to the song (as he didn’t want to) vs. John didn’t hear the song (as he had died that morning). It also holds up if an agent refuses to act, as in John wouldn’t break the glass. Næss (pp. 187–188) also reports on the importance of volition for case-marking, especially regarding psychological predicates such as hear, above. In Kannada, the verb ‘see’ takes a NOM subject when the action described a volitional seeing (similar to English look at or watch). However, when ‘see’ is understood as a psychological experiencer predicate, and the subject is not volitional, it takes a DAT subject. All this leaves volition as a more consequential semantic feature than sentience.

Dowty’s (ii-a) change of state (COS) and (ii-c) affectedness are two other Proto-role Properties which overlap to some degree and have generated closer examination in the literature, particularly in Beavers (2010, 2011) and Grimm (2011). Observations noted in both Beavers’ and Grimm’s papers complicate Luraghi’s Hierarchy, since her hierarchy relies on both COS and Affectedness being discrete, binary features. Beavers and Grimm both contend that Affectedness cannot be adequately captured as a binary and is at the very least tertiary. Grimm (2011: 521–522) challenges the supposed “binary distinction” between affected and unaffected, on the grounds that it obfuscates instances of something being “partially affected,” a distinction that he argues is lost in most typological studies invoking the feature. For his part, Beavers (2010, 2011) suggests that Affectedness is best captured as a degree or scale of COS.

To properly address the matter of affectedness, Beavers (2010) proposes an Affectedness Hierarchy, which involves three types of change and four degrees of affectedness. The types of change are: change of state, change of location, and creation/consumption. The degrees of affectedness in the hierarchy are quantized change > non-quantized change > potential change > unspecified change. With both quantized and non-quantized change, affectedness accompanied by change is entailed, but only in the case of quantized change is it measured, with the degree of change reported. COS is not entailed by potential change, and this degree of affectedness implies an action having taken place which affects an object in a manner which may or may not lead to a change in the object. To display the difference between potential change and unspecified change, Beavers (2010) points to the difference between John hit the rope and John hit at the rope (838), wherein the first entails that the rope receives the brunt of an action which may or may not change it, while the latter does not entail that the object is affected in any way, let alone changed.

Given that Dowty’s proto-roles provide a more highly articulated set of metrics by which to assess the semantics of subject and object arguments, along with the research pursuant to his proposal since that time, we determined to use them as a basis from which to approach the question of DO case-selection, testing the range of meaning-relationships indicated by object case-marking in CG. It is important to understand, as Dowty suggests, that the properties are not completely determinative in mapping arguments to subject and object positions. Rather, subjects tend to exhibit more (rather than always exhibiting all) agent Proto-role Properties, and objects tend to exhibit more patient Proto-role Properties. It is therefore possible (and actually common) for subjects to exhibit some patient-like properties, and for objects to exhibit some agent-like properties. Different verbs may then illustrate unique trends across the assignment and implication of such roles. For this reason, in our examination of the data, we sought to test the full range of both agent and patient Proto-role Properties.

5 Statistical analysis

In this section, we provide a statistical examination of the CG, to determine (i) whether the explanatory device of Luraghi’s transitivity hierarchy is borne out in a more carefully curated set of CG data and (ii) whether a differently constructed hierarchy (feature-wise) might be more predictive. Collected tokens were tested first in accordance with Luraghi’s Hierarchy (Section 5.2.1) and then through the lens of Dowty’s proto-roles (Section 5.2.2).

From this examination we found that Luraghi’s predictions for a COS-based hierarchical distinction between the ACC and GEN marked objects are borne out, but that her predictions for an affect-based distinction between the GEN and the DAT do not. In fact, we find, quite surprisingly, that DAT objects do not fit into this hierarchy at all. Accordingly, while the first two levels of Luraghi’s Hierarchy (High Transitivity being indicated by +COS and selecting the ACC, and Medium Transitivity by -COS and selecting the GEN) are confirmed, DAT objects are unaccounted for and have no place in the hierarchy, regardless of what features are applied. In other words, DAT objects cannot be made to fit into the transitivity hierarchy through the invocation of the COS-feature and something else.

5.2 Statistical analysis

This section presents the results of two statistical analyses of object case-marking drawn from the CG data that we collected and coded for semantic properties. First, we tested whether Luraghi’s Hierarchy can account for the distribution of ACC, GEN, and DAT case-marking. Finding this not to be the case, we next tested the distribution of case against Dowty’s Proto-role Properties, to see if any subset of these semantic features could be used to devise a workable hierarchy similar to Luraghi’s, which could predict levels of transitivity and through case-marking. In our examination of the 2010 Transitivity Hierarchy, we did find support for the correlation of (ACC object marking) high transitivity and (GEN object marking) medium transitivity levels with the presence or absence of the COS property. However, we also did not find support for the inclusion of DAT-marked objects in the Hierarchy on the basis of a correlation with an Affectedness feature as the 2010 Heirarchy would have it. Thus, the present study devises a revision of the Transitivity Hierarchy that proposes only two levels of transitivity and which places DAT object-marking verbs outside of it for reasons that become clear below.

To test Luraghi’s Hierarchy against the CG data, all sentence tokens were coded for the object properties +/−COS and +/−affected, and the transitivity score of each sentence was based on the following arrangement of these features: +COS/+affected = high transitivity; −COS/+affected = medium transitivity; and −COS/−affected = low transitivity. We then determined to what extent the transitivity scores reliably predicted object case-marking in accordance with Luraghi’s claims: high transitivity sentences have ACC objects, medium transitivity sentences have GEN objects, and low transitivity sentences have DAT objects. This analysis was performed with a Chi-square test, using the chi.square function in R Studios (R Core Team, 2018). Since Chi-square tests are suitable for determining whether there is a significant interdependence between two categorical variables, the test was used to find out whether the determinative features used in the Transitivity Hierarchy correctly predict the case-marking of objects.

Finding that COS is predictive but that Affectedness is not, the CG data was then re-examined to determine whether the COS feature might be coupled with other Proto-role Properties to successfully predict the ACC>GEN>DAT Hierarchy. For this purpose, the CG data was coded for the full set of Dowty’s Proto-role Properties (excepting, of course, for affect), to determine whether any of these might interact in a significant way with the data and predict the case marking of objects. By using both subject and object semantic properties, we aimed to test the predictive capacity of each of them, on the understanding that both subject and object properties can collectively determine case marking of the object. Each verb-token was coded +/− for the five Agent Proto-role features of their subjects (i.e., triggering COS, volitionality, movement, independent existence, and sentience) and +/− four Patient Proto-role features of their direct objects (i.e., undergoing COS, incremental theme, lack of movement, and dependent existence). Note that only nine of Dowty’s ten Proto-role Properties are used, because (as discussed in Section 4) Affectedness is not a reliably binary property and because it was already shown to not be predictive (to the extent that it could be coded properly). We also note, in regard to the coding of COS properties, that we only counted the first two stages of Beaver’s (2010, 2011) COS hierarchy (Quantized Change and Non-quantized Change). Accordingly, constructions which entailed quantized change or non-quantized change were coded as +COS, and constructions which entailed potential change or unspecified were coded as −COS.

With this coding in place, Multinomial Logistical Regressions (MLR s) were performed on the full range of data via the multinom function (Venables & Ripley 2002). As MLR s use the Maximum Likelihood Parameter to test for the best fitting predicative model for a categorical Dependent Variables with 3+ levels (here, DO case) across multiple Independent Variables (here, Dowty’s nine Proto-role Properties), we find this to be the best approach to the current data—the MLR s thus helped to determine which features were relevant to the prediction of case-marking on the DO. The results of the MLR are discussed below in Section 5.2.1. The data was coded by the two authors separately, and then discussed. Any discrepancies between coding were discussed and clarified; any discrepancies that could not be resolved were discarded from the data set.

From this examination, we find that COS is predictive of ACC and GEN case-marking, but none of the other semantic features were found to DAT-marking on objects in interaction with COS. We found instead, in addition to being −COS, that DAT objects can also indicate subject movement towards a goal. However, as discussed in section 5.3, this turns out to be reason to believe that DAT-selecting verbs are not best represented within the same hierarchy as ACC and GEN, and are better understood in terms of Blume’s (1998) interaction verbs.

5.2.1 Results: Chi-Square test

This section presents the results of the first step of our analysis, i.e., testing whether predictions made by Luraghi’s Hierarchy are confirmed by a Chi-Square test. It turns out that they are not. While High Transitivity (+COS/+affected) strongly correlates with ACC case-marking, Medium Transitivity (−COS/+affected) does not correlate with the GEN-case, but in fact correlates far more strongly with the DAT case-marking. Furthermore, we find that Low Transitivity (−COS/−affected) better correlates with GEN case-marking and has no correlation at all with the DAT. This indicated that while +COS does in fact set ACC apart from the GEN and DAT case-marking verbs, +/−COS and +/−affected coding does not predictably distinguish between GEN and DAT case-marking verbs. These results call into question the use of +/−affected as a binary feature by which to predict CG case-marking of objects.

If Luraghi’s Hierarchy and the features that are said to underlie it were correct, then the number of instances of ACC-case verb constructions should be highest in the first column under (+COS, +Affected), the number of instances of GEN-case verb constructions should be highest in the second column under (−COS, +Affected), and the number of instances of DAT-case verb constructions should be highest in the last column under (−COS, −Affected). However, our findings disconfirm Luraghi’s predictions. Table 3 shows the largest number of ACC-case verb constructions in the first column (n=30), the largest number of GEN-case verb constructions split across columns two (n=31) and three (n=32), and the largest number of DAT-case verb constructions in column two (n=46) with just a few in columns one (n=2) and three (n=4).

Luraghi’s Hierarchy thus appears, descriptively, to be disconfirmed by the data. Our descriptive assessment is confirmed through statistical analysis. A Chi-Square analysis reveals a highly significant interaction between transitivity and the assignment of case ([X2(4, 154)=130.7, p < .0001]). However, these findings do not fully support Luraghi’s Hierarchy. The Chi-Square residuals in Table 4 contradict Luraghi’s claim that GEN objects are more transitive than DAT objects, where a positive value communicates a positive factor by which the case-marking and level of transitivity can be correlated, and a negative value indicates lack thereof. Thus, while the interaction is highly significant, the results show DAT case-assigning verbs to be more transitive than GEN case-assigning verbs, contrary to the ranking proposed in Luraghi 2010. The correlation plot of the residuals in Figure 1 further illustrates this problem with Luraghi’s predictions.

Table 3

Luraghi’s Transitivity Hierarchy

	High	Medium	Low
	+COS/+Affected	−COS/+Affected	−COS/−Affected
ACC	30	8	0
GEN	1	31	32
DAT	2	46	4

Table 4

Chi-Square Pearson’s Residuals (Luraghi’s Hierarchy Transitivity)

	ACC	GEN	DAT
High	7.660	–3.433	–2.739
Medium	–2.833	–0.728	3.229
Low	–2.980	4.405	–2.339

Accordingly, it must be either that the transitivity ranking (ACC > GEN > DAT) is incorrect or that the properties +/−COS and +/−Affected do not together determine the ranking. Not only does the analysis result in the unexpected hierarchy ACC > DAT > GEN, but it also splits GEN case-marking verbs across two levels of transitivity, Medium and Low. Given these results, along with doubts about using +/−Affected as a binary feature, we are inclined to find more fault with the features than with the hierarchy.

There is also an additional reason to trust the hierarchy above the features, namely the fact that there are CG verbs which optionally mark their objects with ACC or GEN case. This, combined with the fact that there are no CG verbs that display alternating ACC/DAT case-marking of their objects, suggests that ACC and GEN marking are more closely aligned than are ACC and DAT marking. In fact, verbs which case-mark their objects with DAT case only mark them with DAT, suggesting perhaps that DAT case-marking itself lies outside of the transitivity hierarchy altogether.

View Full Size

Figure 1

Correlation Plot of Chi-Square Residuals (Luraghi’s Transitivity Hierarchy)

Citation: Journal of Greek Linguistics 23, 1 (2023) ; 10.1163/15699846-02301004

• Download Figure
• Download figure as PowerPoint slide

5.2.2 Results: Multinomial Logistic Regression

This section reports a test of the nine of Dowty’s Proto-role features that we used in coding the CG sentence data, in order to determine if any subset of them can be used to construct a hierarchy similar to that of Luraghi’s, one which accurately predicts case-marking patterns found in CG prose. To best test this, we employed Multinomial Logistic Regressions (MLR s). In performing various MLR s across the possible combinations of identified features, we found the two most interpretable models and the two with the lowest Akaike Information Criterion (AIC) scores⁵ to be: first, a model which tests the interaction between COS, volition, and object lack of movement (called here MLR-A, AIC = 172.3508), and second, a model which tests COS, subject movement, and object lack of movement (MLR-B, AIC = 188.0661). Reducing our regression model to only two of the above-mentioned features, we find the best-fitting model to be that of COS and subject movement (MLR-C, AIC = 191.0066).⁶

Though the reduction of features in MLR-C leads to a slightly higher AIC than MLR-A, we find it more compelling because volition in MLR-A is not independently predictive at all, and is only predictive when interacting with COS. Despite this, COS is equally predictive with or without volition, indicating that with or without volitionality, the determinate feature remains COS (P-test values show that COS alone has exactly the same p-value as both COS and volitionality, and volitionality alone is not predictive at all). The same can be said for MLR-B concerning object lack of movement, which is no more predictive than COS alone. As such, MLR-A and -B work only to reaffirm what we already know: COS correlates with ACC case-marking, and a lack of COS correlates to non-ACC. However, these two tests do not provide any information as to what sets GEN and DAT case-marking predicates apart?

Testing for MLR-C, we finally find interactions that set the ACC, GEN, and DAT cases apart from each other. COS, as predicted by Luraghi and MLR-A and MLR-B, continues to demarcate the distinction between highly transitive ACC and the other cases. GEN case-marking verbs are 99.99 % less likely to have objects that undergo COS than ACC case-marking verbs (p < .0001). However, alongside that, we also find that subject movement can distinguish DAT from both GEN and ACC. While subject movement is not a significant feature of GEN case-marking verbs, DAT case-marking verbs are 99.93 % less likely to have subjects which do not move than ACC case-marking subjects are at a rate that is significant (p = .013).

Thus, the only feature which consistently and reliably differentiates the case-marking of CG objects is that of COS, setting apart the ACC case-marked DO s from those marked with either GEN or DAT case. COS is always present in the most interpretable models having the relatively lowest AIC scores, and even when models include other variables, those variables are only as significant as COS is alone. The only exception to this is the interaction seen with the variable of DAT case-marking. However, it cannot even be said that DAT case-marking verbs most often have subjects which exhibit subject movement, only that they have them more often than do ACC or GEN case-marking verbs. While this does lead to a successful featural analysis of the distribution of these three object cases, it does not permit us to establish the three-tiered hierarchy of transitivity based on these two features.

Table 5

MLR-C Exponentiated Coefficient Values

Object Case	COS	Subject Move	COS: Move
GEN	.000521	9.479 e-08	1
DAT	3.569 e-08	.0635	1

6 Conclusions and directions for further research

The observations, analyses, and conclusions we have presented here are by no means the last word on this topic. Further analysis is warranted into other Classical Greek texts and genres and other syntactic environments, expanding on the data-domain presented herein. What we have hopefully accomplished is to present a coherent, corpus-driven analysis of the semantic correlates of object case marking for one period of Ancient Greek (i.e., Classical Greek). In delimiting our corpus to a 200–300 year range of texts, we hope to have avoided the confounding contributions of language change, which might have affected and blurred our analysis were we to have included texts over the entire ~1000-year period of Ancient Greek.

The observations made here confirm the basic thesis of Luraghi (2003, 2010) which posits a “transitivity hierarchy” determining the case of transitive verbs’ objects, though we find reason to believe it does not operate in the exact ways that she details. Our observations for, and analysis of, Classical Greek do show, however, that (i) this hierarchy is sensitive to COS, (ii) it is most appropriate to apply this hierarchical relationship to the ACC and GEN only, given the behavior of both non-alternating and alternating ACC- and GEN-marking verb classes, and (iii) DAT-marked objects seem to interact differently from ACC- and GEN-marked objects relative to this hierarchy.

Thus, we propose that this research can and should be taken a step further in order to investigate the case of case-marking alternations in CG verbs, in order to answer in certain terms, whether this alternation can be said to reflect the same sorts of relationships first posed by Luraghi and then further developed herein. If such answers can be provided, it may be possible to then apply these results to close readings of CG texts, texts which display verbs having alternating case-selection patterns. This might lead to a better understanding of the contextual or pragmatic implications of such case alternations that have been previously unrecognized (as Danesi et al., 2018 find, regarding the use of oblique cases to mark subjects in certain CG modal constructions). Turning these results back onto a reexamination of the corpus data may also give way to insights into what might otherwise be anomalous instances of case-marking alternations seen with a small number of verbs, and which indeed have, until now, been considered as such.

This study breaks new ground in its careful application of Dowty’s 1991 thematic proto-role features to the variable case-marking on verbal objects, and in applying this analysis to a large-enough body to yield statistically significant results. In terms of CG, we can attest to the relevance of COS, but can also point to the lack of determinacy of Affectedness as might be predicted by Beaver (2010, 2011) and Grimm (2011).

What remains to be accomplished, and what is well outside the scope of this study, is to apply this method to other periods and other varieties of Ancient Greek. Doing so could provide evidence that the transitivity hierarchy (as a correlation between semantic features and grammatical cases) is stable over time or possibly that it is subject to change. While our present research has taken a synchronic approach, it is our recommendation that further research on this topic might strengthen the findings presented herein by applying these methods to other periods of Greek, thus taking a diachronic perspective. The methods and conclusions of this study also beg application to other languages that might have regular correlations between semantic features and case marking, to discover whether the transitivity hierarchy might be cross-linguistically generalizable and whether it is semantically stable, feature-wise. In such an approach, one hopes to gain insight not only into the development and interaction of semantic and syntactic case within Greek itself, but, perhaps, to shed light on this interaction at large.

Appendix A

Classical Greek verbs analyzed under the present analysis

ἁμαρτάνω ‘mistake, do wrong’

ἀμελέω ‘neglect, be careless’

ἀπαντάω ‘meet’

ἀπειλέω ‘threaten’

ἀποκτείνω ‘kill’

ἅπτω ‘fasten, touch, understand, kindle’

ἀρέσκω ‘please, make amends’

ἄρχω ‘rule’

βασιλεύω ‘rule over’

βοηθέω ‘aid, help’

γράφω ‘write’

δουλεύω ‘be a slave to’

ἔοικα ‘be like’

ἐπανίστημι ‘rise against’

ἐπιμελέομαι ‘take care of’

ἕπομαι ‘follow, pursue’

ἐσθίω ‘eat’

εὐνοέω ‘be friendly to’

ἐφικνέομαι ‘arrive at, reach, attain’

ἡγέομαι ‘lead, begin’

καταλύω ‘destroy’

καταγελάω ‘mock’

κοινωνέω ‘have, share’

κορέννυμι ‘satiate oneself’

λανθάνω ‘escape the notice of’

λείπω ‘leave’

λίσσομαι ‘beg, beseech, pray’

μάχομαι ‘fight’

μέτειμι ‘be among, have part in’

μετέχω ‘partake’

νικάω ‘conquer’

πειράω ‘try, make an attempt’

περιτυγχάνω ‘meet, happen on’

πίνω ‘drink’

πρέπω ‘befit’

τυγχάνω ‘succeed, obtain’

τυρβάζω ‘disturb’

Appendix B