The Hebrew dative: Usage patterns as discourse profile constructions

3.1 Data

I account for simultaneous parameters in a quantitative, bottom-up fashion, revealing usage patterns in a corpus. Quantitatively searching for patterns in a multivariate data, we can achieve a description of the corpus that is less dependent on the researcher’s presuppositions regarding the research question. For example, the mere theoretical decision about the particular function of the dative in each sentence is a highly subjective one, as shown above, largely depending on the researcher’s theoretical stance (Rudzka-Ostyn 1996). Using quantitative bottom-up methods we can avoid the need for such decisions, and compare the patterns emerging in the data with the traditional, subjectively defined participant roles.

The corpus serving as a database for the present research is the Israeli Parliament Corpus (IPC, Dattner 2015b): an approximately 1,760,000 words corpus of spoken Hebrew, retrieved from the transcriptions of 198 meetings of committees of the Knesset, the Israeli parliament. The corpus is composed of multiple registers of language, both formal and colloquial. From this corpus, all occurrences of a dative (l-) marked pronoun were extracted, yielding 9,694 tokens of dative uses.^[2]

Each token of dative use was coded for 15 variables. The choice of coding variables is based on the literature regarding dative constructions in Hebrew and in general. Thus, the coding includes the subject-predicate order, for example, as the predicate-first word order is said to be used with a particular dative function (Melnik 2006). As noted above, previous research found the notion of scalar, discourse-determined transitivity (Hopper and Thompson 1980) to play a crucial role in the use of dative constructions in Hebrew (Dattner 2008; Dattner 2015a). Transitivity, as defined by Hopper and Thompson (1980), is gradient, composed of a configuration of ten parameters. These configurations define different degrees of transitivity, such that an action is transferred from one participant to another with different degrees of effectiveness or intensity (Hopper and Thompson 1980): A high transitivity clause has two or more participants, and it conveys a telic, punctual action (rather than a state, or an atelic action) which is carried out by a volitional participant, high in Agency. If the patient (i.e. the endpoint of the action) is highly affected in the sense that it goes through a complete change of state, then the clause’s transitivity increases. A highly Individuated affected referent increases the transitivity of the clause even more: A human or an animate endpoint, who is concrete rather than abstract, singular, count, referential and definite, marked by a proper rather than a common noun, would render the clause higher in transitivity. The transitivity of a clause can be moreover increased if the clause if affirmative and contains a realis encoding of an event.

In Hebrew, as a semitic language, transitivity is also related to the verbal morphological paradigm (Berman 1987). The Hebrew verbal lexicon is morphologically organized in verb patterns, termed binyanim (lit. “buildings”). Seven such patterns exist, traditionally named kal, nifal, hifil, hufal, piel, pual, and hitpael. Each pattern is a paradigm related to a different degree of transitivity, and to a different type of argument structure (Berman 1993; Raz Zalzberg 2018): kal is underspecified for transitivity, piel and hifil are transitive, nifal and hitpael are traditionally analyzed as manifesting low transitivity, and pual and hufal are the passive patterns. Dealing with questions of transitivity and argument structure, the morpho-syntactic binyan system is highly relevant as a coding category for the present study (Ravid 2012).

The coding variables in the current analysis thus consider the following criteria, summarized in Table 1: (i) Features of the Subject argument and the predicate: the agency and individuation of the Subject referent (S/A.Agency. Individuation), the verbal morphological paradigm (the Hebrew binyan: Verb.Binyan), the lexical category of the predicate (Predicate.Lxical.Category), the type of predicate (Predicate.Type), the number of arguments in the clause (No.Of.Arguments), Subject-Predicate order (Predicate.First), the mode of the clause (Mode), the polarity of the clause (Polarity), ellipsis (Ellipsis), and voice: whether the predicate is in the passive or middle voice (Verb.Passive.Middle). (ii) Features of the Direct-Object argument: the type of non-dative argument in the clause (No.Dative.Argument), the level of individuation of the Direct-Object referent (O.Argument.Individuation), and the definiteness of the Direct-Object argument (O.Argument.Definiteness). (iii) Features of the dative referent: the person of the dative referent (Dative.Person), and its participant role (Dative.Function).

Regarding the last parameter, it is important to note that the subjective coding of dative functions was done only to provide a reference point for the bottom-up clustering process, in order to consider the objectively defined clusters in the light of traditionally defined participant roles. Thus, the dative function for each token was coded by the researcher according to the traditional analyses of dative participant roles (e.g. Cuervo 2003; Halevy 2016; Van Belle and Langendonck 1996; Van Langendonck and Van Belle 1998). Importantly, the dimensions of the Hierarchical Clustering on Principal Components reported bellow are constructed according to a set of active variables defined by the researcher. Only those variables that were hypothesized as candidates for explaining the distribution of dative interpretations are included as active variables. Therefore, the dative function parameter itself (Dative.Function), although coded in the corpus, was left out of the construction of the Multiple Correspondence Analysis map and the Hierarchical Clustering on Principal Components.

3.2 Method

The usage-based approach to language, which I adopt as a theoretical framework, is embedded in Cognitive Linguistics, a model of language that emphasizes the role of method and data (Glynn and Fischer 2010). A usage-based model of language calls for an empirical, quantitative approach to linguistic research, and specifically to corpus linguistics. Quantitative methods in linguistics aid in describing, explaining, and predicting linguistic phenomena (Baayen 2008; Gries 2013). The analysis presented in this paper is based on multivariate exploratory statistics, and specifically on cluster analysis using Hierarchical Clustering on Principal Components (Husson, Lê and Pagès 2011; Husson, Josse, Le and Mazet 2013; R Core Team 2014).

Cluster analysis provides some objectivity in determining groups of similar tokens (Divjak and Fieller 2014). Cluster analysis (in different manifestations) is recently used in linguistics in various areas. For example, it has been used in sociolinguistics (e.g. Moisl and Jones 2005), comparing language varieties and dialects (e.g. Gries and Mukherjee 2010; Hyvönen, Leino and Salmenkivi 2007; Szmrecsanyi and Kortmann 2009), computational linguistics (e.g. Kovatchev, Salamó and Mart\’ı 2016), lexical processing (e.g. Baayen, Feldman and Schreuder 2006), Construction Grammar (e.g. Dattner 2015a; Gries and Stefanowitsch 2010, accounting for semantic classes and prototypical cases of constructions), or analyzing behavioral profiles (Divjak and Gries 2006; Snoek 2011, and see Moisl (2015) for a comprehensive survey of the use of cluster analysis in linguistics).

Hierarchical Clustering on Principal Components (henceforth, HCPC) is a special case of cluster analysis, done on the principal components of a Multiple Correspondence Analysis (Divjak and Fieller 2014; Husson, Josse & Pagès 2010; Husson, Lê and Pagès 2011). Multiple Correspondence Analysis (MCA) is a Principal Component Analysis (PCA) for categorical variables (Husson, Lê and Pagès 2011). Performing hierarchical clustering on the principal components of the MCA, the MCA can be viewed as a de-noising method. Without the noise in the data, the clustering analysis is done on the signal, rendering the analysis more stable. The resulted clusters of tokens can be described according to the variables in the data, or according to the individual tokens each cluster is composed of. For example, the categorical variable number of arguments in Table 1 has three categories: one, two, and three arguments in the clause. Hypothetically, using HCPC we might learn that the category three arguments is linked to cluster one, while the categories one argument and two arguments are linked to cluster two. This observation can teach us that tokens with a three-argument syntactic structure share more features within themselves than they share with one- and two-argument tokens. That is, we can conclude that (i) beyond the fact that they share a particular category, they are similar on other levels as well, and (ii), even if two tokens do not share a category, they may still belong to the same cluster (i.e. if they are similar enough) since they share other features. Using HCPC we can consider the most prototypical tokens in a cluster, such that they constitute the center of gravity: these are the tokens that are closest to the center of the cluster. All other tokens in the cluster are related to these tokens through a chain of family resemblances. Moreover, HCPC provides us with the objects that belong to a cluster and are placed the furthest from other clusters’ centers; these are the cluster’s unique tokens.

Choosing the number of clusters is critical in exploring data using clustering methods (Husson, Josse & Pagès 2010). That is, in the process of clustering the data, each token can be treated as a cluster on its own, and on a different level, all the individuals belong in a single cluster. The number of clusters is thus chosen according to growth of inertia. For example, consider a data that can be divided into three, four, or five clusters. Each division to a particular number of clusters explains a certain amount of the variation in the data, such that a division to five clusters explains the most, and a division to three clusters explains the least. The preferable number of clusters is a function of the growth of explained variation. That is, if the difference between dividing the data into three clusters versus four clusters is greater than the difference between dividing the data into four clusters versus five clusters, than one should choose to divide the data into four clusters. This is so since the further clustering would not result in a significantly better representation of the data, but would add noise to the representation. In other words, the further clustering costs too much. The results of the clustering analysis of the data introduced above are presented and discussed in the following section.

4.1 Cluster one: Extended transitive discourse profile construction

The first cluster of tokens in the corpus gathers together 2,870 dative clauses that show a similar usage pattern, presented in Table 2.

Analyzing Table 2 we can see that the first cluster of tokens is linked to transitive predicates with concrete participants (these are termed Primary A; the classification of predicates in the corpus followed the semantic approach to grammar advocated by Dixon (2005)). Most of the predicates belong to the kal verb paradigm (underspecified for transitivity), and some to the the transitive piel paradigm (see Table 1 for the list of coding categories). Syntactically, Cluster One is linked to a three-argument structure with a nominal A, a nominal O, and a dative-marked participant in the second person, functioning as what a traditional participant role analysis would define as a Recipient, an Affectee, or a Possessive Dative. Cluster One is also linked to a realis mode of clauses, with an intermediate-high degree of Agency of the A referent, and an individuated O participant. These characteristics of Cluster One correspond to high transitivity, and a high degree of affectedness of the dative-marked participant (Dattner 2008; Hopper and Thompson 1980; Næss 2007; Rozas 2007; Thompson and Hopper 2001).

Besides its categorical variables, each cluster has a limited number of central and unique exemplars according to which it can be described. A central exemplar is the token closest to the center of cluster, in the sense that it shares more formal categories with other tokens in the cluster than with tokens in other clusters. A unique exemplar, on the other hand, is the token furthest from the centers of other clusters, in the sense that it shares few, if any, categories with the central tokens of the other clusters. The following are Cluster One’s central exemplars:

Looking at the three sentences in (8), we can observe a realization of the usage pattern presented in Table 2 as strongly linked to Cluster One: Verbs with concrete participants as their semantic roles, embedded in a three-argument structure with a highly agentive A participant (e.g. the minister or the representatives), and a nominal O participant. Cluster One’s unique exemplar, the token furthest from the centers of other clusters in the data, is the following:

(9) presents an accusative marked, relatively individuated O participant, a transitive, Primary A predicate in the kal paradigm (natan ‘give’), and an A participant with an intermediate-high degree of agentivity. These features, as we have seen above, mostly characterize tokens from Cluster One, and thus are unique to this particular cluster. Accordingly, the dative-marked participant in (9) can be termed a Recipient: The dative-marked participant is affected by a transitive motion event in which an Agent (metaphorically) moves a Theme from one point to another. However, it involves partial affectedness and indirect relations: the Recipient does not go through a complete change of state. The only entity that is construed as going through a change of state is the transferred Theme. The Recipient is affected only as a result of the Agent’s action on a different object, to the extent that he becomes a possessor of the transferred Theme. This type of three argument relation is elementary, and thus is lexically encoded by transfer verbs (Francez 2006; Haspelmath 2011; Malchukov, Haspelmath and Comrie 2010).

The first Discourse Profile Construction is thus related to high transitivity values, and is defined as the Extended Transitive Discourse Profile Construction: a usage pattern used to construe a relation between two participants (i.e. a transitive event) that produces a high degree of effect on a third entity (an extension to core, see Dixon (2005)). The third entity in the current case can be analyzed as a Recipient, an Affectee, or a Possessive Dative. This dative can be lexically governed by the predicate (as in (9) for example), or it can be a non-lexical dative, as in (10) bellow. It is a cover function unifying all extensions of an affective situation, in which the core event can be construed as complete without any reference to a third participant. It is in the Extended Transitive Discourse Profile Construction that a two-place event gains a relative, dependent status, anchored to a third, dative-marked participant. For example, it is only in the current Discourse Profile Construction (i.e. usage pattern) that the two-participant event of ‘drinking water’ is construed as related to a third entity, consequently affecting it:

This description of Cluster One is in line with Boneh and Bar-Asher Siegal’s (2014) account of non-lexical datives in Hebrew. Boneh and Bar-Asher Siegal (2014) suggest that the affected dative is a cover function, comprising the Possessor, the Benefactor, and the Malefactor. The common feature shared by these roles, according to Boneh and Bar-Asher Siegal (2014), is the marking of a participant who is affected by an event of which he is not a part of.

The following sentence contains such a non-lexical dative, which might pose a problem for interpretation. The solution proposed by the present Discourse Profile Constructions approach assumes that this sentence belongs to the Extended Transitive Discourse Profile Construction which guides the non-lexical dative’s interpretation:

(11) belongs to a usage pattern which emerges out of a set of utterances characterized as having a transitive predicate that belongs to a specific class of verbs, a realis mode, and a three participant event in which the affected participant is highly affected and the affecting participant has high agentivity and volition. That is, it is a clause with high transitivity, that construes a relation between a two participant event and a third participant. These (and some other) characteristics constitute a usage pattern, linked to a single conceptualization of the world. This link guides the interpretation of the dative-marked participant, be it lexically governed or not: In the prototypical case of the Extended Transitive Discourse Profile Construction, it construes a transfer relation. In the less prototypical cases, it can be any high transitivity event that can be construed as affecting a third participant.

4.2 Cluster two: Human endpoint discourse profile construction

The second cluster is presented in Table 3. It is the largest cluster in the data, grouping together 4,127 tokens of dative clauses (42.5% of the corpus).

Cluster Two is a coherent and homogeneous cluster, linked with transitive predicates from the types Primary B and Secondary C. Primary B verbs have concrete participants, but they can also be complemented by a clause (either instead of or in addition to the nominal complement), while Secondary C verbs introduce a subject that actually plays a role in bringing about the event or state referred to in the complement clause’s verb (Dixon 2005). These predicates are situated in a three-argument structure, but unlike Cluster One, the Direct Object slot is occupied by a clause or a Preposition Phrase rather than a Noun Phrase. That is, two arguments are referential, while the third is a clausal complement of the predicate conveying the content in a telling or showing event (Primary B verbs), or a Prepositional complement of the predicate (Secondary C verbs). The verb paradigms (binyan) linked with Cluster Two are the transitive hifil, which was not related to Cluster One, the underspecified kal, with the transitive piel showing a weaker link. Cluster Two is linked to the irrealis mode, and to a dative referent in the third person, which in participant role terms would be analyzed as one of the following: (i) an Addressee, the human endpoint of a message delivering action, its audience, or (ii), a Human Endpoint, which is the only (partially/mentally) affected endpoint of a two-participant event, marked as human. These characteristics can be detected in the following central exemplar:

Syntactically, the third argument in (12) is a Prepositional complement of the main predicate, usually indicating a secondary event the initiator of which is the dative-marked participant.

The unique exemplar of Cluster Two construes a telling event in which the dative-marked participant is mentally affected:

Note that the third argument in (13) is a finite clause indicating the content of a message delivered by the A referent to the dative-marked participant. That is, the dative-marked participant in both (12) and (13) is the only real affected-endpoint of the event, the clausal/Prepositional complement being non-referential (Thompson and Hopper 2001). The construal of the telling event conveyed in (13) profiles the audience of the telling action (the dative-marked participant) as the single endpoint of the action. This is in contrast with telling construals that belong to Cluster One, where the type of message is profiled and occupies a full nominal argument position, the O argument:

The dative-marked participant in the ‘type of message’ case (the audience in (14)) is not construed as the only affected entity in the event, but rather as indirectly affected by the metaphorical creation of a message. Thus, a telling act is construed differently when used within the usage patterns of Cluster One and Cluster Two: in the Discourse Profile Construction that emerges from the tokens of Cluster One a telling/showing event is construed as a creation of an abstract entity. This creation is done relative to another entity (the dative-marked participant), thus affecting it. In the Discourse Profile Construction emerging from the tokens of Cluster Two, however, a telling/showing event is construed as a two-pole action, with a teller/shower and an audience/perceiver, emphasizing not the creation of the message, but its content.

This is, then, the Discourse Profile Construction emerging from usage pattern uncovered as Cluster Two: a set of lexical, morpho-syntactic, semantic and pragmatic features that can be summarized as related to an intermediate level of transitivity, conventionally paired with an interpretation of the dative-marked participant as mildly/mentally affected by an action initiated by the A referent. This type of action is usually realized through either a telling/showing or helping/causing events as shown above, but not always, as can be seen in the following Cluster Two examples:

What groups these exemplars together with the central and unique exemplars presented above is the usage-pattern-based Discourse Profile Construction. That is, it is not a case of metaphorical extension from help to relinquish, or to drill. Rather, these cases construe an event with an intermediate degree of transitivity in which the dative-marked participant is the only endpoint, partially or mentally affected. These exemplars share more formal similarities with other exemplars belonging to Cluster Two than with exemplars from other clusters in the corpus: it is a combination of a transitive verb with a two-argument, Subject-first structure in an irrealis clause, or a three-argument structure with a non-nominal occupying the third argument position (i.e. no individuated O), and an intransitive verb in the active voice. This is the Human Endpoint Discourse Profile Construction.

4.3 Cluster three: Extended intransitive discourse profile construction

The third cluster in the corpus is composed of 1,110 tokens of dative clauses, and is summarized in Table 4. The tokens of Cluster Three are linked to the Secondary D, B, and A types of predicates (ordered according to strength of link), mostly in the nifal verb paradigm, and to the Primary A type in a weaker link. Secondary D verbs are intransitive verbs that take a complement clause in subject slot and another role (a stance taker), marked by the dative. Secondary B verbs have one independent role (the subject) in addition to the roles of the verb in the complement clause, describing the subject’s attitude towards some event or state. Secondary A verbs have no independent semantic roles and modify the meaning of another verb (the main verb of a complement clause), sharing its roles and syntactic relations. Primary A verbs have only concrete participants as their semantic roles. Cluster Three is linked to a two-argument, predicate-first structure, with the predicate either in the nifal paradigm, as mentioned above, or in the passive voice, thus appearing in the hufal and pual verb paradigms. The non-dative argument most strongly linked to Cluster Three is of an Adjectival nature, as in:

It can, however, be a clause or a nominal as well. In nominal cases, it is low in Agentivity. The dative referent linked to Cluster Three is in the first person, and it corresponds with the Experiencer or the Ethical Dative in traditional terms, as in the following central exemplar of Cluster Three:

While the Experiencer (17) is the participant role most strongly linked with Cluster Three, other types of participant roles can be found in the tokens close to the cluster’s center as well. For instance, the dative-marked participant in (18) can be analyzed as a Recipient or a Possessive Dative:

Nevertheless, the dative-marked participant in (18) is profiled as wishing to be in a state in which he has some money, therefore it is more of an Experiencer than a Possessor.

The unique token of the cluster, on the other hand, may be analyzed as consisting of an Addressee dative-marked participant, which is the human endpoint of a message delivering action. In Cluster Three, however, such an Addressee, a Recipient or a Possessive Dative are related either to intransitive predicates (18), or to the passive voice, as in:

Considering this variety of examples, trying to characterize the tokens in Cluster Three as related to a single participant role is an impractical task. The question thus remains what brings these (and other) tokens together; i.e. what are the important similarities they feature, such that they share a usage pattern, thus belonging to the same cluster. The tokens in Cluster Three are counted as similar not on the basis of a common participant role marked by the dative, but rather on the basis of the construal conveyed by the speaker, realized as a shared usage pattern. The following sentences belong to Cluster Three as well, and include an intransitive verb, a predicate-first structure, and other low transitivity features.

This usage pattern is related to a construal that anchors a situation to a human referent, the dative-marked participant. This is the Extended Intransitive Discourse Profile Construction. That is, while a stative construal consists of an S argument and a predicate, the Extended Intransitive construal is composed of an S argument, a predicate, and an extra participant marked by the dative (E, or extension to core, see Dixon (2005)), to which the relation between the S and the V are construed as relevant, much like the Extended Transitive Discourse Profile Construction emerging from Cluster One.

4.4 Cluster four: Evaluative reference point discourse profile construction

Cluster Four is presented in Table 5. It consists of 1,141 tokens, strongly linked to a two-argument syntactic structure with Adjectival predicates from the types Property and Value (Dixon 2005). The Property type includes adjectives such as barur, ‘clear,’ kashe, ‘hard,’ and nagish, ‘accessible.’ The Value type includes adjectives such as xashuv, ‘important,’ tov, ‘good,’ and nifla, ‘wonderful.’ Another distinctive feature of Cluster Four’s tokens is a predicate-first word order, with the non-dative argument realized as a non-finite clause. Cluster Four is linked to dative referents in the first person, traditionally analyzed as Experiencers, Judicantis, or Evaluative Reference Points in participant role terms. (21a)–(21b) are the closest exemplars to Cluster Four’s center of gravity, sharing the highest number of features with other exemplars in the cluster. (22a)–(22b) are Cluster Four’s unique exemplars, sharing the smallest number of features with tokens from other clusters.

Looking at Table 5 and considering (21)–(22), we can characterize Cluster Four as an Adjectival/Adverbial cluster. Its tokens construe an evaluation of a state of affairs relative to, or from the point of view of, the dative-marked participant. For example, the speaker in (22a) is not evaluating the arriving of the CEO as inappropriate; rather, this evaluation is anchored to a reference point – the dative-marked participant. And in (22b) there is a subjective reference to the situation, anchoring the evaluation to a reference point profiled by the dative-marked participant. That is, in (22b) the conveyed meaning can be broken into two parts, much like the type of construal we have seen in the case of the extended constructions emerging from Cluster One and Cluster Three: an evaluation, and its anchoring to a reference point.

In summary, the usage pattern of Cluster Four revolves around low affectedness of the dative-marked participant, and adjectival clause with low transitivity and predicate-first word order. This usage pattern is conventionally paired with a construal of an evaluation of a state of affairs, anchored to the dative-marked participant. This convention is the Evaluative Reference Point Discourse Profile Construction.

4.5 Cluster five

The fifth cluster is outlined in Table 6. It is the smallest cluster in the data, consisting of only 456 tokens. Clusters with a small number of tokens tend to be rather heterogeneous, and are less easily characterizable compared with other clusters presenting a uniform behavior.

The tokens of Cluster Five have a predicate-first syntactic structure, with non-verbal predicates, and an irrealis clause. They are mostly one-argument clauses, with the single argument being the dative-marked participant as in the following central (23a) and unique (23b) exemplars:

The dative construction in (23a) is a parenthetical comment concerning the speech act and its interlocutors in this case. The dative construction’s function in this type of parenthetical comments is to put the utterance in a relative, dependent state, anchoring it to an external entity (be it of the interlocutors, or another relevant referent).

The unique exemplar of Cluster Five (23b) is a gratitude utterance, aimed towards the dative-marked participant. It resembles dative clauses with other nominal predicates concerning manner and greetings (also belonging to Cluster Five) such as:

In some respects, these examples are not different from the Addressee participant role, or the Recipient. While in the Recipient cases it is a transfer event construal in which a transitive motion event is directed at the dative-marked participant, and in the Addressee cases it is a telling/showing event, here it is a greeting action, lexicalized in a nominal rather than a verbal construction. Thus, there is no substantial difference between the nominal dative clause in (23b) and its verb-headed equivalent in (25) that belongs to Cluster Two:

The only difference is in the construal of the state of affairs. While in (25) the greeting action is construed as a transitive event consisting of an Initiator and an Endpoint, the same action is construed in (23b) as a stative situation with no explicit reference to the Initiator of the action, profiling its Endpoint as the only relevant human entity.

We can see, then, that there is no general Discourse Profile Construction that emerges from the tokens of Cluster Five, due to the cluster’s heterogeneous nature. However, we can still detect a recurrent pattern agreeing with the types of construals we have seen so far as related to dative constructions.

4.6 Clustering analysis: Summary of the results

Summing up the results of the cluster analysis, we have shown that a corpus of Hebrew dative clauses can be categorized into clusters in a bottom-up manner to reveal hidden usage patterns. The clustering process yielded five clusters of tokens sharing usage patterns. Each of the five clusters in the data was shown to be both a cluster of tokens, and a cluster of lexical, morpho-syntactic, semantic, and discursive features. Four clusters were shown to be paired with unique conceptualizations. These correspondences between usage patterns and conceptualizations of the world have been defined as Discourse Profile Constructions. The first two clusters are transitive: The first cluster is a high transitivity pattern, defined as the Extended Transitive Discourse Profile Construction. The second cluster is an intermediate transitivity pattern, defined as the Human Endpoint Discourse Profile Construction. The third and fourth clusters are subdivisions of an intransitive pattern: The Extended Intransitive Discourse Profile Construction, and the Evaluative Reference point Discourse Profile Construction. Cluster five was shown to be incoherent due to its low frequency.

5.1 Discourse profile constructions: Implications for a usage-based grammar

In a seminal work, Goldberg (1995) shows that Argument Structure Constructions are the language’s basic clausal form-meaning pairings. These concern schematic clausal expressions and idealized event schemes. Goldberg points out that only such idealized events can lie at the bottom of basic clausal expressions; since specific events do not frequently recur, they are not incorporated into the grammar (Goldberg 1998; Kemmerer 2006). Thus, in any language we should be able to find a correspondence between basic clausal expressions and idealized event schemes with constant participant roles. However, regarding Hebrew dative constructions, the present paper argues that we cannot assign a single functional interpretation to correspond with a single Argument Structure, nor can we say for every Argument Structure that it construes a single idealized event with constant participant roles. Conducting a corpus research and using multivariate exploratory statistics, the present paper shows that from a usage point of view we should take into account multiple parameters that govern the interpretation of a clause.

Language use is the simultaneous interaction of different dimensions. Thus, one has to consider the various factors that take place in this interaction, both for an adequate description and as a model of language knowledge and use. The Discourse Profile Constructions approach adopts an exemplar-based model of language representation (Bybee 2006; Gahl and Yu 2006). In an exemplar model of language, exemplars are actual tokens of linguistic experience. These are categorized and matched with similar tokens of experience that have previously been stored as exemplars. Each exemplar of a morpho-syntactic construction includes information about the contexts of use, which might concern the local inferences made in such contexts. The concept of Discourse Profile Constructions accounts for similarity between exemplars in an objective, quantitative way using multivariate statistics (Dattner 2015b). Thus, Discourse Profile Constructions comprise not only a form-function correlation but also the network of cues and weightings that represent the language-user’s knowledge. Considering Discourse Profile Constructions, the differences and similarities between clausal patterns can be accounted for in a quantitative way. Adopting a concrete rather than abstract perspective, and treating multiple parameters rather than focusing on the syntactic-semantic pairing, the basic usage-relevant clause types of the language can be defined. The present paper suggests that from a usage perspective, the basic clausal form-function correlations in the language are not Argument Structure Constructions, but Discourse Profile Constructions: Emergent form-function correlations that consist of multiple sources of formal and functional information, conventionally pairing usage patterns of clauses with unified construals of the world.

The exemplar-like model proposed by the emerging Discourse Profile Constructions has another advantage, as it allows for a probabilistic approach to similarity that is not limited to form. Tokens that are similar enough on various dimensions (e.g. the agency of the A participant, or the mode of the clause), might be judged as functionally similar even though they do not share an Argument Structure Construction (as shown in (16)–(20) above, for example). The present paper explains this phenomena by arguing that while they might not share a formal Argument Structure Construction, they do belong to the same Discourse Profile Construction: Since a Discourse Profile Construction is multifactorial, similarity between tokens goes beyond syntax, and can be quantitatively assessed by several parameters at once. In an exemplar-based model of language representation, as a new token arrives, its similarity to existing exemplars is judged based on their family resemblance. Assuming Discourse Profile Constructions, the Argument Structure Construction is only one of the factors relevant for the comparison between new tokens and existing exemplars. That is, the Discourse Profile Constructions approach can accommodate a more naturalistic linguistic behavior, in which similarity is multifactorial. Constituting the usage-relevant basic clausal expressions in the language, Discourse Profile Constructions are in fact local optima; assemblies of different features which speakers weight against each other with respect to syntactic usage.

5.2 Summary

Previous research claims that the relevant information for interpreting a dative clause is its particular dative-marked participant role, its argument structure, or its specific predicate (e.g. Berman 1982; Blume 1998; Boneh and Bar-Asher Siegal 2014; Borer and Grodzinsky 1986; Cuervo 2003; Da̧browska 1997; Hole, Meinunger and Abraham 2006; Newman 1998; Šar\’ıc 2002). In contrast, the present paper suggests that it is usage patterns, realized as Discourse Profile Constructions, that are the key for interpretation. Discourse Profile Constructions are strong, emergent links between discourse-related construals of the world and their multifactorial formal labels. It is thus a broad definition of the Argument Structure Construction (Goldberg 1995; Perek 2015), which takes into consideration multiple sources of information: lexical, morphological, syntactic, semantic, and discursive, and moreover allows for probabilistic similarity. Discourse Profile Constructions resemble clusters of exemplars (Bybee 2006; Gahl and Yu 2006), organized and grouped together according to multiple parameters simultaneously. Using cluster analysis to reveal hidden usage patterns in corpus, the present paper shows that traditional analyses of dative roles are not constantly reflected in usage, to the extent that there is no one-to-one correspondence between the usage patterns revealed in the corpus and traditionally defined dative functions. The cluster analysis uncovers four meaningful dative usage patterns (two transitive, two intransitive), each corresponding to a conceptualization of a state of affairs, defined as Discourse Profile Constructions: (i) a high transitivity pattern, defined as the Extended Transitive Discourse Profile Construction; (ii) a medium transitivity pattern, defined as the Human Endpoint Discourse Profile Construction; (ii) an intransitive pattern, the Extended Intransitive Discourse Profile Construction; and (iv), an intransitive pattern, the Evaluative Reference point Discourse Profile Construction.

In sum, the present paper suggests that, from a usage-based perspective, it may not be the fine-grained differences between very similar semantic roles which are important for interpretation, but rather the Discourse Profile Construction the token resembles. That is, the cluster analysis provided us with an explanation for the principles that categorize tokens together into a single cluster, constituting a usage pattern which corresponds to a single conceptualization. The Discourse Profile Construction approach thus enables us to better understand the notion of usage as a factor in the conventional pairing between form and function. However, this conclusion is currently limited to the description of the present corpus. A psycholinguistic research looking for the psychological reality of Discourse Profile Constructions is needed in order to strengthen the claim, and so is a learning-oriented study accounting for the role of Discourse Profile Constructions in language acquisition.