12.1 Classification and interpretation

In the domain of the semantics of compounding, the central question is that of semantic interpretation. This is the question of how the meaning of the compound relates to the meaning of its components. Obviously, the answer to this question depends on the individual compound considered. In order to make generalizations, it is necessary to classify compounds. However, there are different views as to the intrinsic importance of classifications.

A good overview of the discussion of classification is Scalise and Bisetto (Reference Scalise, Bisetto, Lieber and Štekauer2009). On the basis of a thorough study of earlier classifications, they make a “new proposal” (Reference Scalise, Bisetto, Lieber and Štekauer2009: 49–52). A problem with this proposal is that the classification is presented as a purely hierarchical tree structure, although the labels they attach to the nodes of the tree suggest otherwise. At the top level, they have three classes: subordinate, ATAP (attributive or appositive) and coordinative. The first two of them are divided into two further classes. At that point each of the classes is divided into endocentric and exocentric. Technically, the fact that each of the five classes is divided into the same two categories means that we should not have a tree structure in the first place. The opposition between endocentric and exocentric is obviously independent of the distinction among the five classes they apply to. Therefore, we should have a matrix structure instead of a tree structure, so that, for instance, endocentric compounds constitute a class at the same time as attributive compounds.

Another type of problem is that some of the expressions they treat as compounds are rather dubious. In their examples, Scalise and Bisetto (Reference Scalise, Bisetto, Lieber and Štekauer2009: 50) include blue-eyed, sans papiers and mother-child. As argued in ten Hacken (Reference ten Hacken1994, Reference ten Hacken2010), expressions such as blue-eyed have very little in common with compounds and much more with derivations. The only reason why such expressions have sometimes been treated as compounds is that a derivational analysis requires the suffix -ed to be attached to a phrase. This violates the level-ordering hypothesis. However, as shown by Bauer et al. (Reference Bauer, Lieber and Plag2013), for a language like English, this hypothesis has in any case no foundation in the data. Semantically, it is obvious that blue-eyed (‘with blue eyes’) and similar derivations should be analysed as a regular application of the rule attaching -ed in a way also found in bearded (‘with a beard’). Once we accept the possibility of phrasal input to derivation, there is no reason to exclude the same type of analysis for French sans abri (lit. ‘without shelter’, i.e. homeless [person/people]), which can be considered a conversion of a PP, as I argue in ten Hacken (Reference ten Hacken, Hajičová, Kotěšovcová and Mírovský2003). In the case of mother-child, the status as a compound is somewhat dubious because of the limited distribution. Such expressions can only occur as the non-head of compounds. As proposed in ten Hacken (Reference ten Hacken1994, Reference ten Hacken, Hajičová, Kotěšovcová and Mírovský2003), this can be explained from the fact that this position does not require a syntactic category.

As long as we use the labels of compound classes as pretheoretical descriptions, without any theoretical claims attached to them, there is no reason to bother too much about such issues. In this volume, much of the discussion concerns the interpretation of the relationship between the components of a compound. In this context, it is not necessary to know for each expression which category of compound it belongs to. It is worth considering, however, how the different frameworks highlighted in this volume relate to classification.

Jackendoff’s Parallel Architecture (PA), as presented in this volume, is perhaps the clearest example of a framework that is not interested in classification at all. Jackendoff (this volume) presents 13 “basic functions”, but they cannot be taken as the basis for a classification. It is interesting to compare them in this respect to Levi’s (Reference Levi1978) Recoverably deletable predicates (RDPs). For Levi, a (non-verbal) compound is 12-ways ambiguous because of the nine RDPs, three of which are reversible. None of these readings is marked as the right one. The list Levi (Reference Levi1978: 280–284) gives in her appendix is an illustration of how the RDPs should be interpreted, not a classification of compounds. When a compound is lexicalized, the reading may or may not correspond to one of the RDPs.

Jackendoff’s system for the characterization of the meaning of compounds can be seen as an elaboration of the idea of RDPs, but without the precisely specified number of readings. His basic functions are meant to interact with elements of the meanings of the components, in particular proper function and other qualia along the lines of Pustejovsky (Reference Pustejovsky1995), so that we end up with a generative system. Not only is there no reason to distinguish rigorously delimited categories of compounds, even for compounding and indeed for word formation, but Jackendoff also refuses to specify any meaningful boundaries. As a starting point, this is certainly defensible. Any proposal to draw boundaries in such a way that special mechanisms are at work has to be argued for. In ten Hacken (Reference ten Hacken2010, Reference ten Hacken, Bloch-Trojnar and Bloch-Rozmej2012a, Reference ten Hacken, ten Hacken and Thomas2013a), I make the argument that in the framework of PA, word formation should be in a component that is distinct from the lexicon, which in PA contains lexical items, idioms, and formation rules.

Given the concern to account for the meaning rather than to classify expressions, it is not surprising that Jackendoff (Reference Jackendoff, Lieber and Stekauer2009: 115–117) accepts what he calls promiscuity. As I understand this concept, it should be distinguished from ambiguity. An expression is ambiguous if it has more than one reading. In the case of promiscuity, we have several analyses corresponding to basically the same reading. This can be illustrated on the basis of (1).

1. (1) vuurtoren     (lit. ‘fire tower’, i.e. lighthouse)

   1. a. toren with vuur characteristically on it

   2. b. toren that makes vuur

   3. c. toren from which shooting weapons can be gevuurd

The Dutch equivalent for lighthouse can be described with the basic functions be (X, AT/IN/ON Y) as in (1a) or make (X, Y) as in (1b). For lighthouse, Jackendoff (this volume) gives the latter. They are different ways of describing the same meaning and therefore we have promiscuity here. In (1c) there is another possible meaning. It is not lexicalized in Dutch, but it is interpretable with enough context. Here we have a separate reading, because it is no longer a tower of the same type as in (1a–b). Therefore, vuurtoren is ambiguous between (1a–b) on one hand and (1c) on the other.

In Lieber’s framework, the focus is somewhat different. The distinction between skeleton and body makes it easier to determine a part of the meaning of expressions that can give rise to precise classifications. The skeleton is characterized by seven binary features (cf. Lieber, this volume). As the presence or absence of a feature is also used contrastively, this yields quite a large number of theoretically possible combinations. In addition, skeletons have a structure that goes beyond a simple combination of features. We can therefore not easily determine a priori how many skeleton classes there might be. Nevertheless, it is much easier to determine that two expressions have the same skeleton than how close their conceptual structures in PA are.

When we compare the distribution of information between skeleton and body with the treatment of this information in PA, we can observe the following. Jackendoff (Reference Jackendoff2002: 345–350) distinguishes conceptual structure from spatial structure. Both represent meaning, but whereas conceptual structure interacts with language, spatial structure interacts with the visual system, the haptic system and the action system. Detailed shapes of objects are represented in spatial structure, but not in conceptual structure. Compared to Lieber’s division between skeleton and body, Jackendoff’s conceptual structure is much richer in information and less formally structured than Lieber’s skeleton. Part of what Lieber encodes in the body finds its place in Jackendoff’s conceptual structure.

In Štekauer’s onomasiological approach, the eight onomasiological types (OTs) provide an obvious and explicit basis for classification. However, when we use them to classify existing compounds, we do not use them in their originally intended function. Given the onomasiological nature of the model, they are meant as steps in the speaker’s decision process, that is, on the way from conceptual structure to phonological structure. When we classify existing compounds, we focus on an intermediate step in the putative naming process that resulted in the form-meaning combination we can observe. It should be noted that such a classification is actually what Štekauer (Reference Štekauer1998: 93–117) uses as the basis for his discussion of productivity. Therefore, although it is not compelling to do so within his system, Štekauer’s own use of OTs in this sense justifies the classification of compounds into OTs. Of course, compounds are not in any way distinguished from other types of word formation in this model.

To summarize, in Jackendoff’s PA, it would be wrong to use the basic functions as a basis for classification. The skeleton structure in Lieber’s framework offers a better mechanism for classification, although she does not present the skeleton in this role. Štekauer’s OTs can be legitimately used as classes of word formation, but not specifically of compounds. In fact, in all three models the interpretation of compounds is more central than their classification.

12.2 Some examples of primary nominal compounds

The best way to compare the three frameworks in their role to account for the semantic interpretation of compounding is to consider what they have to say about the same compounds. Let us first consider the Dutch compounds in (2).

1. (2)

   a.	fietsenrek	(lit. ‘bicycle-en-rack’, i.e. bicycle stand)
   b.	treinkaartje	(lit. ‘train-card-dim’, i.e. train ticket)
   c.	zonnepaneel	(lit. ‘sun-e-panel’, i.e. solar panel)

In (2a) and (2c), the Dutch compounds have linking elements. They will not concern us here. In (2b), the diminutive suffix is semantically necessary, as treinkaart would rather be interpreted as a map of the railway system (cf. also ten Hacken Reference ten Hacken2013b). The compounds in (2) are primary in the sense that the semantics of their components does not provide a predicate to make the link between the two components explicit.

As a benchmark, we can consider what the older theories of Levi (Reference Levi1978), Allen (Reference Allen1978) and Selkirk (Reference Selkirk1982) would have to say on (2). Selkirk (Reference Selkirk1982) basically excludes the semantics of primary compounds from her system. She does not say anything more than that they have a right-headed structure [N N]_N. Allen (Reference Allen1978) adds the Variable R condition, which only states that the semantic relation depends on the component nouns. Levi (Reference Levi1978) gives RDPs, but they are quite coarse-grained. Both (2a) and (2b) have for, whereas (2c) has use. Of course, this should be interpreted as a shorthand for the statement that all three compounds in (2) are 12-ways ambiguous, but these particular RDPs are the ones a speaker using the compounds intends the hearer to conclude to. Crucially, anything going beyond the RDPs can only be added in (the unanalysed process of) lexicalization.

As I stated in the introduction to this volume, Jackendoff’s system can be considered as an improvement of the expressive power and precision of Levi’s RDPs. In (3), I give tentative analyses of (2a) and (2b).

(3)

1. a. fiets₁enrek₂ = [_Thing REK₂^α; [PF (CAUSE (PERSON, GO (FIETS₁ (TO (IN α)))))]]

2. b. trein₁kaartje₂ = [_Thing KAARTJE₂^α; [PF (LET (α, BE (PERSON^β, ON (TREIN₁^γ))); [PF (CAUSE (γ, GO (β, PATH)))])]]

In (3a), a fietsenrek is described as a thing (more precisely a rek) whose proper function is for people to put a fiets into. This is clearly much more specific than just rek for fiets. Moreover, it is very economical on functions introduced. All the functions except for REK and FIETS are used generally in Conceptual Structure, as illustrated by many examples in Jackendoff (Reference Jackendoff1990).

In (3b), the structure is much more complex. Treinkaartje is described as a thing (more precisely a kaartje) whose proper function is to let someone be on a train. The proper function of the train is to cause the people on it to go along some path. Following Jackendoff (Reference Jackendoff, Lieber and Stekauer2009: 126), I assume that the PF automatically attaches to the last-named item. Therefore the first PF is the PF of kaartje and the second one the PF of TREIN. Whereas Jackendoff (this volume) uses MOVE as the PF of boat in his analysis of steamboat, I think that for scheduled train services the Path is crucial. Therefore I used CAUSE and GO in (3b), so that there is a natural place for the Path constituent. As in (3a), no idiosyncratic functions have to be introduced.

For (2c), it is more difficult to come up with a similarly detailed, yet theoretically economical analysis. Van Dale (Reference van Dale, Geerts and Heestermans1992) gives the definition in (4) under zonnepaneel.

1. (4) paneel dat zonlicht moet opvangen, m.n. zo’n paneel met foto-elektrische cellen om zonlicht om te zetten in elektrische energie

   ‘panel that should collect sunlight, in particular such a panel with photo-electric cells to transform sunlight into electrical energy’

The basic function that (4) suggests is MAKE. Jackendoff (Reference Jackendoff, Lieber and Stekauer2009: 124) proposes a basic function MAKE (X, Y, From Z). This would be useful, because the panel transforms sunlight into electrical energy. Jackendoff (this volume) has two basic functions with MAKE: one indicating source material (as in olive oil) and one indicating product (as in honey bee). In (5), the two analyses corresponding to these versions are given.

(5)

1. a. zon₁nepaneel₂ = [_Thing PANEEL₂^α; [PF (MAKE (α, ENERGY, FROM (ZON₁)))]]

2. b. zon₁nepaneel₂ = [_Thing PANEEL₂^α; [PF (MAKE (α, ENERGY^β)); [MADE (β, FROM (ZON₁))]]]

In (5a) we see the three-place predicate MAKE as proposed by Jackendoff (Reference Jackendoff, Lieber and Stekauer2009), whereas (5b) splits this into two different predicates: one for the product and one for the source. In both versions in (5), the use of FROM is slightly awkward, because it diverges quite markedly from Jackendoff’s examples, all of which imply that, as in olive oil, the object of FROM is used up in the process.

Let us now consider the compounds in (2) from the perspective of Lieber’s theory. As I mentioned in the introductory chapter, this system can be seen as an elaboration of Allen’s (Reference Allen1978) Variable R condition, because it adds a theory of slots. Lieber divides the semantic properties into a skeleton, the more formalized features of the body and the less formalized features of the body. In (6), I give a tentative analysis of (2a).

1. (6)

   fiets	rek
   [+material ([i])]	[+material ([i])]
   <–animate>	<–animate>
   <+artefact>	<+artefact>
   {transport, pedals, … }	{for storage, … }

Both components in (6) refer to material things and they are coindexed because they constitute a compound together. The nature of the relation between the components can be constructed from the elements of the body. The step from knowing that a relation has to be constructed to a hypothesis about such a relation is much less formalized than in (3a). This can be seen when we compare (6) to the tentative analysis of (2b) and (2c) in (7).

1. (7)

   a.	trein	kaartje
   	[+material ([i])]	[+material ([i])]
   	<–animate>	<–animate>
   	<+artefact>	<+artefact>
   	{transport, public, payment per journey, … }	{information, proof of payment, … }
   b.	zon	paneel
   	[+material ([i])]	[+material ([i])]
   	<–animate>	<–animate> <+artefact>
   	<–artefact>	<2 dimensions>
   	{radiation, light, heat, … }	{for covering sth, … }

In (7a), we find that the components of the compound have the same skeleton and formalized aspects of the body as in (6). The only difference is in the encyclopedic knowledge encoded in the non-formalized part of the body. The difference in the relation between the components of (2a) and (2b) is expressed much more explicitly in the contrast between (3a) and (3b) than in the difference between (6) and (7a).

A problem in devising (7b) is that paneel in isolation gives only very few indications of what it means in zonnepaneel. It generally refers to what is inside a rectangular frame. This can be a painting or a set of buttons and switches used to operate a complex electrical device, etc. Therefore, only very little information can be included about paneel. This contrasts with the situation in (5), where the PF is the proper function of a zonnepaneel. The contrast between (5) and (7b) highlights a difference between Jackendoff’s and Lieber’s approaches. Whereas Jackendoff describes the meaning of zonnepaneel and refers to the components, Lieber describes the meaning of zon and paneel and derives the compound meaning.

Finally, let us consider how the compounds in (2) are treated in Štekauer’s onomasiological model. A full account would track the entire process from the identification of a type of object in the real world to the phonological form of the name associated with it. Some of the steps are clearly more interesting than others, however. Štekauer (this volume) concentrates on determining the OT and assigning semantic features to each of the components. From this perspective, the compounds in (2) are very similar and can collectively be represented as in (8).

1. (8)

   OT3	DingM	DedM	Base
   	R	0	R
   	Object	Action	Object
   a.	fiets(en)	0	rek
   b.	trein	0	kaartje
   c.	zon(ne)	0	paneel

In all three compounds, we have an unexpressed Action linking two Objects. In OT3, the Determining component of the Mark and the onomasiological base are expressed, but not the Determined component of the Mark. It is only at the onomatological level that the three compounds are distinguished, because individual morphemes are attached to the expressed positions. The linking elements in brackets are added at phonological level.

At this point, it is interesting to consider how similar the analyses in the different approaches are and where the differences stem from. As noted above, Jackendoff starts with the full compounds, whereas Lieber starts with the components. The contrast between Jackendoff and Lieber can be represented as a polar opposition. Given the onomasiological nature of his theory, Štekauer starts from the unnamed concepts. It would be tempting to analyse this approach as being in direct opposition to the semasiological approaches pursued by Jackendoff and Lieber. I think for Lieber we can indeed see such an opposition. However, if both Jackendoff’s and Štekauer’s model are in a kind of polar opposition to Lieber’s, the question is how Jackendoff’s and Štekauer’s models relate to each other. We will come back to this question in section 12.6.

12.3 Some examples of verbal compounds

In the generative tradition, primary compounds are opposed to verbal compounds. Since Roeper and Siegel (Reference Roeper and Siegel1978), verbal compounds have attracted most of the theoretical attention. Some Dutch examples are given in (9).

1. (9)

   a.	fietsenstalling	(lit. ‘bicycle-en-storage’, i.e. bicycle shelter)
   b.	boekenlegger	(lit. ‘book-en-lay-er’, i.e. bookmark)
   c.	wespensteek	(lit. ‘wasp-en-sting’, i.e. wasp’s sting)

In each of the compounds in (9), the second noun contains a verbal element. In (9a), stallen is related to stal (‘stable’), but its meaning has been extended from animals to vehicles (especially two-wheeled ones) in Dutch. The presence of the verb gives an explicit indication of the relation between the two components. In older approaches, this has been modelled differently. Levi (Reference Levi1978) adopts the position that such compounds have four possible analyses based on nominalization, in addition to the 12 RDP-based analyses. Selkirk (Reference Selkirk1982) takes the argument structure of the verb as the basis for her treatment of such compounds, while also admitting that the first noun can have a non-argument interpretation, as in tree eater (‘someone who eats only when in a tree’).

In each of the three models under scrutiny here, we also find a clearly distinct treatment of verbal compounds. In PA, the compounds in (9) can be analysed as in (10).

(10)

1. a. fiets₁enstal₃ling₂ =

   [_Thing OBJECT₂^α; [STAL₃ (INDEF, FIETS₁, IN α)]]

2. b. boek₁enleg₃ger₂ =

   [_Thing OBJECT₂^α; [LEG₃ (INDEF, α, IN BOEK₁)]]

3. c. wesp₁ensteek₂ =

   [_Event STEEK₂ (WESP₁, INDEF)]

There is a clear distinction between (10a–b) on one hand and (10c) on the other. (10c) is an example of a pure application of the Argument Schema, (23) in Jackendoff (this volume: 25). In all three cases, the second noun provides the function that connects the two components, but in (10c) the second noun is equal to this function. Therefore, we have a transposition. In transpositions, we have identical conceptual structures corresponding to different syntactic structures. In the conceptual structure in (10c), it is not visible whether the syntactic realization will be a sentence or a compound noun (cf. ten Hacken 2015).

Jackendoff (this volume: 25) also treats verbal compounds such as hair dryer as a result of the application of the Argument Schema, as in his (25). This is the analysis adopted in (10a–b). Jackendoff (Reference Jackendoff, Lieber and Stekauer2009: 125), however, gives a different analysis of hair dryer. Applying this older approach to (9a–b) results in analyses such as (11).

(11)

1. a. fiets₁enstal₃ling₂ =

   [_Thing X₂^α; [PF (STAL₃ (PERSON, FIETS₁, IN α))]]

2. b. boek₁enleg₃ger₂ =

   [_Thing X₂^α; [PF (LEG₃ (PERSON, α, IN BOEK₁))]]

The differences between (10a) and (11a) are four. First, whereas (10a) applies the Argument Schema, (11a) is an application of the Modifier Schema. The second difference, following from the choice of Schema, is that in (10a) stal provides the function, whereas in (11a) the PF of stalling provides the function. The other two differences are less crucial. Whereas (11a) has an unspecified profiled item, it is specified as OBJECT in (10a). Conversely, whereas (10a) has INDEF as the first argument of the function, (11a) specifies that this is a PERSON. I will assume that the change of analysis is intended and that Jackendoff prefers (10a–b) over (11).Footnote ¹

Lieber (this volume) makes a distinction between synthetic compounds such as (9a–b) and non-affixal deverbal compounds such as (9c). In both types, the skeleton provides much more information than for primary compounds such as (2), so that the features of the body need not be referred to immediately for their interpretation. Tentative skeletons for (9a–b) are given in (12).

(12)

1. a. [+material ([_i])] [+material, dynamic ([_j], [+dynamic ([ ], [+dynamic, +IEPS ([_i], [+LOC ([_j])])])])]

        fiets(en)                    -ing                        stal_V

2. b. [+material ([_j])] [+material, dynamic ([_i], [+dynamic ([ ], [+dynamic, +IEPS ([_i ], [+LOC ([_j ])])])])]

        boek(en)                   -er                          leg_V

In (12) three basic components are distinguished, in the same way as in (10) and (11). For both verbs, the skeleton is based on Lieber’s (Reference Lieber2004: 59) representation of put. For stallen it would perhaps be more natural to consider what in (12a) is the third argument as a modifier instead, but this would make it impossible to coindex the argument linked to -ing with anything. This argument with -ing is necessary, because a fietsenstalling is an object rather than an action. In (10a), the location is also represented as an argument. It might be questioned why -ing and -er in (12) have the feature [dynamic]. Here I follow Lieber’s (this volume: 40) structure of -er, her (4).

When we consider the coindexation in (12), we observe first of all that in both cases the two available elements (the noun and the suffix) are coindexed with arguments of the verb. Explaining the choice of the argument to coindex each element with is not so straightforward. De Haas and Trommelen (Reference de Haas and Trommelen1993) give as the first meaning of the affixes the process for -ing and the agentive for -er. Examples with synthetic compounds are given in (13).

1. (13)

   a.	vliegtuigkaping	(‘aeroplane hijack-ing’)
   b.	wijnkenner	(lit. ‘wine know-er’, i.e. wine expert)

In (13a), -ing is not coindexed with an argument of kapen, because (13a) designates an event, see Lieber’s (this volume: 45) structure for celebration in her (12). In (13b), -er would be coindexed with the first argument of the verb. Readings of (9a–b) parallel to (13) are possible, but not lexicalized. They can only be activated by providing enough context.

In (12), the nouns fiets and boek are coindexed with different arguments of the verbs. It may be possible to explain this choice in part on the basis of properties encoded in the body. Both can be interpreted as a space, but in its role as containing a text a boek has a linear structure which a fiets does not have. In any case, much of the coindexation can only be explained in terms of the need to name certain objects. It is only because of this need that the lexicalized interpretations of fietsenstalling and boekenlegger in (9) are not parallel to (13).

For the non-affixal deverbal compound (9c), the skeleton can be represented as in (14).

1. (14)

   [+material ([sentient-i])] [–material, dynamic ([+dynamic ([sentient-i],[ ])])]

   wesp(en)                       steek

The structure in (14) is parallel to the one Lieber (this volume: 46) gives for dog attack, her (16a). In (14), the feature sentient is taken from the body and it is used as a condition on coindexing of the verbal argument. However, it would not exclude a reading parallel to autogebruik (‘car use’), in which the first noun is coindexed with the second slot of the converted verb. Again, knowledge of the type of event a name is needed for determines the choice.

When we now turn to Štekauer’s model, we find the same distinction between (9a–b) on one hand and (9c) on the other. Here (9a–b) belongs to OT1, whereas (9c) belongs to OT5. In (15) and (16), analyses of the three compounds are proposed.

1. (15)

   OT1	DingM	DedM	Base
   	R	R	R
   a.	Object	Action	Location
   	fiets(en)	stal	-ing
   b.	Location	Action	Object
   	boek(en)	leg	-er

1. (16)

   OT5	DingM	DedM	Base
   	R	R
   	Agent	Action	Event
   	wesp(en)	steek

Verbal compounds of the type in (9a–b) are Štekauer’s main class of examples for OT1, the onomasiological type in which all three elements of the onomasiological structure are expressed. In (15a) and (15b), we see the by now familiar reversal of the roles of the first and last elements. In this framework, this is indicated by the semantic roles they are assigned. There is no parallel to the problem of selecting the correct reading of the suffix in contrast to the reading in (13), because the central question in Štekauer’s approach is how a name is chosen for a particular concept. Therefore, alternative interpretations of the forms are not considered.

In (16), we see how conversion is analysed. The component steek realizes the determined part of the mark as well as the onomasiological base. As such, steek has the same role as kaping in (13a), where the two are realized by a verb and a suffix. In general, the DedM expresses an action, process or state.

In the analysis of verbal compounds, Jackendoff’s theory does not make a strong distinction to primary compounds, whereas for Lieber, they have a very different analysis, as she accounts for the meaning of primary compounds on the basis of the body of the components, but for verbal compounds a large part of the meaning is encoded in the skeleton. For Štekauer, verbal compounds belong to another OT, but within the class of verbal compounds, there is also a distinction between the OT of compounds with a derived head and those with a converted head.

12.4 Implicit predicates

The difference between the primary compounds in (2) and the verbal compounds in (9) is that the former have in general a much wider range of possible relations because their components do not express any preferred relation. The verb in the compounds in (9) directs the interpretation by suggesting a preferred relation. The compounds in (17) can be considered intermediate between these two types.

1. (17)

   a.	fietsendief	(lit. ‘bicycle-en-thief’, i.e. bicycle thief)
   b.	taxichauffeur	(‘taxi driver’)
   c.	vleesmes	(lit. ‘meat knife’, i.e. carving knife)

In each of the head components in (17), there is a clear implication of the verb attached to it. This is most obvious in the Dutch-English pair in (17b). Whereas chauffeur is a borrowing from French, which is not morphologically complex in Dutch,Footnote ² the English driver is a clear agent noun with an explicit verb. Levi (Reference Levi1978) discusses car thief. She adopts a generative semantic analysis in which the underlying Deep Structure of thief is the same as the one for stealer. In (17c), mes can be paraphrased as snijder (‘cutter’) in a parallel way to thief and stealer. The question is to what extent such an analysis is desirable.

In Jackendoff’s system, a crucial issue is what he calls lexicalization of a conceptual structure.Footnote ³ This can be illustrated on the basis of the representations in (18).

1. (18)

   a.	fiets1endief2 =
   	[Thing PERSON2α; [STEAL (α, FIETS1)]]
   b.	taxi1chauffeur2 =
   	[Thing PERSON2α; [OCC (DRIVE (α, TAXI1)]]
   c.	vlees1mes2 =
   	[Event MES2α; [PF (CUT (INDEF, VLEES1); WITH α)]]

The lexical conceptual structure for dief has an open slot where (18a) has FIETS. As STEAL is part of the lexical entry of dief, there is no need to use any other mechanism to provide a predicate for the compound if the most common sense is intended.Footnote ⁴ The structure in (18b) is similar, except that the action modality of OCC is used. In Dutch, a taxichauffeur is a professional driver of a taxi. An occasional driver would have to be referred to as a taxibestuurder (lit. ‘taxi steer-(d)er’, i.e. taxi driver). In these structures, Jackendoff’s theory resembles Levi’s (Reference Levi1978) generative semantic approach. The main difference is that in Levi’s approach the semantic structure underlies the syntactic and phonological structures, whereas in PA the three representations are parallel, without any predetermined directionality.

The structure in (18c) can be compared to the one for fietsenrek in (3a). Both involve a proper function, but it can be argued that the proper function of a mes (‘knife’) is more prominent and more determinate than the proper function of a rek (‘rack’). The PF in (18c) is associated with mes, whereas in (3a) the PF is a property of the entire compound fietsenrek rather than of rek. This difference is a matter of degree and it is entirely conceivable to think of any number of intermediate cases bridging the gap between vleesmes and fietsenrek. The lack of a clear boundary between synthetic and primary compounds is characteristic of Jackendoff’s approach.

In Lieber’s model, by contrast, there is a clear distinction. She characterizes subordinate compounds of the type illustrated in (9a–b) as “those that involve indexation between one element of the compound and an argument of the base verb from which the other element is derived” (this volume: 42, emphasis added). In (17) there is no base verb, so that the meaning of the compound is necessarily derived in the same way as for the compounds in (2). For each of the compounds in (17), the two components have a skeleton characterized as [+material [ ]] and these two skeletons are coindexed in their only available position. In their skeletons, they are identical to (6) and (7). The information that a dief steals and a chauffeur drives (occupationally) is encoded in the body of the meaning representation and can be used to interpret the coindexation of the two components.

Štekauer’s model requires us to choose an OT. For the compounds in (17), there are two options, either OT3, as for fietsenrek in (8), or OT5, as for wespensteek in (16). It seems plausible to divide the examples in (17) so as to arrive at the analyses in (19) and (20).

1. (19)

   OT5	DingM	DedM	Base
   	R	R
   	Object	Action	Agent
   a.	fiets(en)	dief
   b.	taxi	chauffeur

1. (20)

   OT3	DingM	DedM	Base
   	R	0	R
   	Object	Action	Instrument
   	vlees	0	mes

In determining whether OT5 or OT3 is the best analysis, the question is whether the head of the compound incorporates an action or not. In (19), dief and chauffeur are analysed in the same way as if they were converted verbs. In (20), the Action is treated as entirely unexpressed.

At this point, it is worth considering what such a distinction actually means. Ten Hacken and Panocová (Reference ten Hacken and Panocová2011) analyse the role of the speech community in Štekauer’s theory and conclude that it is fundamentally compatible with a competence-based model of language, as long as we take for granted that the speech community is not an entity performing actions or taking decisions, but only a community of speakers who as individuals can perform actions and take decisions. The question we have to ask about the significance of the distinction encoded in (19) and (20) is, then, whether we would notice if another speaker had a differently encoded lexical entry in their mental lexicon; (21) illustrates this possibility.

1. (21)

   OT3	DingM	DedM	Base
   	R	0	R
   	Object	Action	Agent
   	taxi	0	chauffeur

Suppose that one speaker of Dutch, Annet, has taxichauffeur in her mental lexicon as (19b) and another speaker, Berend, has it in his mental lexicon as (21). It is doubtful whether Annet and Berend will ever find out that they have a different lexical entry for taxichauffeur in their mental lexicons. The form is the same and the meaning is close enough so that no difference is noted.Footnote ⁵ The only situation where Annet and Berend are likely to notice the difference is if they are discussing the analysis of the word in Štekauer’s framework. This does not mean that the distinction between OT3 and OT5 is a spurious one. There are obvious cases of OT3 (e.g. fietsenrek) and obvious cases of OT5 (e.g. wespensteek), so that we need both OTs. As long as we focus on the account of word formation as a naming mechanism, it does not matter that taxichauffeur could be either. Such borderline cases only become a problem when we try to use the OTs as a classification.

The treatment of compounds with implicit predicates gives us further evidence for the nature of the distinction between primary compounds and verbal compounds. Lieber takes the absence of a verb in the form as a sufficient reason to treat compounds such as (17) as primary compounds, deriving the relation between their components from properties of the body only. Jackendoff is not forced to make a clear distinction and can use the predicates whether or not they are lexically expressed. Therefore, the transition between the primary compounds and verbal compounds need not be marked by any sharp boundary. Although Štekauer’s model encodes primary compounds as OT3 and verbal compounds as OT5, the discussion of the compounds in (17) puts this distinction into context. Even though the naming process takes one of these routes in each individual instance of naming, it is not necessary to have a unique attribution of an OT for each compound. The gradual transition can be modelled as different speakers of the same speech community having a different OT for the same compound.

12.5 Compounds with a verbal non-head

As a final set of data, let us consider compounds in which the non-head provides a predicate that can be interpreted in the meaning of the compound. Some Dutch examples are given in (22).

1. (22)

   a.	parkeergarage	(lit. ‘parking garage’, i.e. covered car park)
   b.	bromvlieg	(lit. ‘humming fly’, i.e. bluebottle)
   c.	strijkkwartet	(lit. ‘stroking quartet’, i.e. string quartet)

In Dutch, the form of the non-heads in these cases is the verb stem. In the glosses in (22), this has been translated by an -ing form. As I argued in ten Hacken (Reference ten Hacken1994), the non-head position of a compound is non-categorial. Syntactic categories are based on syntactic distribution, but the non-head of a compound is not syntactically accessible. This makes it possible to have components here that are barred from syntax because they lack a syntactic category. In ten Hacken (Reference ten Hacken, Hajičová, Kotěšovcová and Mírovský2003) I give the example of attribute-value pair, where attribute-value is a coordinate construct of a type that cannot appear outside compounding. Therefore, whether parkeer in (22a) stands for the verb parkeren or the nominalized infinitive parkeren cannot be determined. However, the meaning does not depend on this classification.

A remarkable property of all compounds in (22) is their degree of lexicalization. Parkeergarage in (22a) can refer to a multi-storey car park or an underground car park, not to a garage for two or three cars. A bromvlieg in (22b) is an animal belonging to the family of calliphoridae, not just any fly that makes a humming noise. Most specifically, a strijkkwartet in (22c) is a chamber ensemble with two violins, a viola and a cello. Whereas in the English translation these instruments are called by the strings that vibrate to produce the sound, in Dutch it is the action that is highlighted. At the same time, strijkkwartet can also refer to a piece of music written for such an ensemble. This ambiguity is the same as in English.

In Jackendoff’s framework, we can construct conceptual structures as in (23) for (22a–b).

(23)

1. a. parkeer₁garage₂ =

   [_Thing GARAGE₂^α; [PF (PARKEER₁ (PERSON, CAR, IN α))]]

2. b. brom₁vlieg₂ =

   [_Thing VLIEG₂^α; [HAB (BROM₁ (α))]]

In both cases in (23), the non-head provides the function, but it is not a relation between the two components. This is most striking in (23b), which has a one-place predicate as its function. It is interesting to compare parkeergarage in (23a) with fietsenrek in (3a). The two functions have the same argument structure, but in (3a) the object is expressed and in (23a) the function. CAR in (23a) is a defeasible default. In several Dutch cities there are fietsparkeergarages (lit. ‘bicycle parking garages’, i.e. underground or multi-storey parking facilities for bicycles), but without the explicit mention of fiets, it is understood that a parkeergarage is for parking cars. Jackendoff (Reference Jackendoff1983) calls such defaults preference rules.

In the case of strijkkwartet in (22c), we encounter the situation that the meaning of the compound is only in a very small part retrievable from the meaning of its components. Jackendoff (Reference Jackendoff1975) introduced the mechanism of redundancy rules, which is very well suited for such situations. The idea is that we describe the meaning of strijkkwartet in all relevant detail, including, for instance, the information about which instruments the quartet is composed of. The meanings of strijken and kwartet are described in separate entries. The entry for strijk-kwartet refers to the entries for strijken and for kwartet and those parts of the entry for strijkkwartet that are covered by the entries referred to do not count in calculating the information load for the storage of the entry for strijkkwartet. The information provided by the entries of the components includes much of the form, the syntactic category and aspects of the meaning; for example, that it is a group of four. The connection to strijkinstrumenten (lit. ‘stroking instruments’, i.e. strings) is partly expressed, but not the convention which ones are included.Footnote ⁶

Lieber (this volume: 46) treats compounds with a verb in non-head position as non-affixal deverbal compounds, thus emphasizing the similarity between examples such as (22) and cases like wespensteek in (9c). On the basis of the analysis of pushcart in her (16b), we can assume structures such as (24) for (22a–b).

(24)

1. a. [–material, dynamic ([+dynamic, +IEPS ([ ], [ ], [+LOC ([_i])])])] [+material ([_i])]

   parkeer garage

2. b. [+dynamic ([_i])] [+material ([_i])]

   brom vlieg

The structure in (24b) is interesting because it yields a minimal pair with the structures in (6) and (7). Whereas in (6) and (7) the features of the body were crucial in determining the sense of the coindexation, in (24b) the coindexation fills an obvious slot in the frame for brommen. In (24a), the verbal frame in the first line offers three possible slots for coindexation and the features of the body of garage indicate which one is the most likely. One can imagine a parkeerbediende (‘parking servant’, i.e. a person employed by a hotel to park the cars of guests) asking about a parkeerauto (‘parking car’, i.e. car to be put in the car park), where bediende fills the first slot and auto the second slot of parkeren as represented in the first line of (24a).

For strijkkwartet in (22c), Lieber’s system does not provide the amount of detail that makes it a special case in Jackendoff’s framework. The word formation process involved is separated from the information about which instruments play in it. What is important is the distinction between the two readings, one as a chamber ensemble and the other as a piece of music. They can be encoded by different coindexation, as in (25).

(25)

1. a. [–material, dynamic ([+dynamic ([_i], [ ])])] [+material ([_i])]

           strijk                                                        kwartet

2. b. [–material, dynamic ([+dynamic ([ ], [_i])])] [–material ([_i])]

           strijk                                                        kwartet

The difference in the readings in (25) is in part expressed by the value of [material] for kwartet, because players are [+material] but the music they play is [–material]. In (25b) I assume that strijken has as its object the piece of music rather than the instrument. In Dutch, spelen is used for both instead of strijken, as illustrated in (26).

(26)

1. a. Carla speelt/*strijkt Beethoven’s Harfenquartett.

   ‘Carla plays/*strokes Beethoven’s Harfenquartett.’

2. b. Carla speelt/*strijkt altviool.

   ‘Carla plays/*strokes viola.’

In (25) we have to assume, then, that strijken stands for spelen and that there are two readings with different requirements of the features of the second argument. The ambiguity between (25a) and (25b) is systematic, but limited to expressions denoting chamber ensembles.

In Štekauer’s model, the verbal elements in (22) have to be interpreted as the determined part of the onomasiological mark. For (22a) and (22c), this means that OT2 is the relevant type. This is illustrated in (27).

1. (27)

   OT2	DingM	DedM	Base
   	0	R	R
   a.	Object	Action	Location
   	0	parkeer	garage
   b.	Object	Action	Agent
   	0	strijk	kwartet
   c.	Agent	Action	Object
   	0	strijk	kwartet

In (27a), parkeren is represented as having an unexpressed object. It is not a counterexample that we can have a stationsparkeergarage (‘[railway] station car park’), because the object of parkeren is still implied. The two readings of strijkkwartet have a reversed attribution of Agent and Object roles, where (27b) is the ensemble reading and (27c) the piece-of-music reading.

In OT2 we can only encode cases with transitive verbs. For intransitive verbs as in (22b), we need OT7, as illustrated in (28).

1. (28)

   OT7	Mark	Base
   	R	R
   	Action	Agent
   	brom	vlieg

As with the other types of compound discussed here, the lexicalization aspect is treated differently to the other frameworks, because the concept to be named is the starting point. In the case of strijkkwartet, the composition of the ensemble is determined before any component of the name. Therefore, there is no point at which the meaning of strijkkwartet is specialized by giving the instruments involved. Similarly, the family of insects is the starting point for linking bromvlieg to this family of animals.

Compounds of the type in (22) give further evidence of the differences between the three frameworks. In Jackendoff’s formalism, the meaning of a verbal element can be used in conceptual structures independently of the position where it is realized in phonological and syntactic structures. In Lieber’s framework, these compounds are modelled on the basis of the skeleton in a way that is very similar to verbal compounds discussed in section 12.3. Štekauer’s system uses two further OTs, depending on whether the verb is transitive or not.

12.6 Similarities and differences

Having considered analyses of four types of compound, we have collected some contrastive data about the three frameworks that we can use for a characterization. Contrasts are most clearly seen when two frameworks are compared. Therefore I will start by looking at each of the three pairs separately.

Let us start by comparing Jackendoff’s and Lieber’s approaches. A striking difference between the analyses we have seen in this chapter is the degree of specificity. Lieber describes compounds on the basis of the skeleton. This skeleton has a limited structure with slots that can be coindexed. For the precise interpretation of coindexation, Lieber refers to the properties of the body, the less formalized part. Jackendoff proposes a much more elaborate conceptual structure. Although not all details of the conceptual structure for an expression are formalized, the formalization is scalable. It is as if we can use a telescope to home in on particular features whenever desired. There is no restricted domain for formalization of the type of Lieber’s skeletons.

This difference in depth of coverage and formalization of the semantic representation is connected to a different approach to syntax. As Jackendoff (Reference Jackendoff2002) notes, Chomsky’s models are all strongly syntactocentric. Against this background, we have to understand Selkirk’s (Reference Selkirk1982) reduction of (word formation) morphology to syntactic rules, an approach that inspired Distributed Morphology (DM). In her overview of compounding in DM, Harley (Reference Harley, Lieber and Štekauer2009) concentrates on the generation of the correct forms of compounds. As ten Hacken and Thomas (Reference Thomas, ten Hacken and Thomas2013: 15–18) argue, DM is not interested in the semantics of word formation for its own sake, but assigns it a place in its model to get it out of the way. In contrast to DM, Lieber’s approach embraces semantics. However, semantics for Lieber includes an ontological category substance/thing/essence, which she considers “the notional equivalent of nouns” (Reference Lieber, Lieber and Štekauer2009: 81).

The contrast between Lieber’s and Jackendoff’s frameworks is perhaps most clearly visible in the way the distinction between verbal compounds and primary compounds is modelled. For Lieber, the difference between them is that verbal compounds have a much richer skeleton structure, which encodes the argument structure of the verb and can be exploited to derive the meaning of the compound from its component parts. Primary compounds have a rather impoverished skeleton structure. For Jackendoff, the difference between verbal and primary compounds is gradual. It is not visible in the conceptual structure itself whether we are dealing with a verbal or a primary compound. This only becomes clear when we consider the coindexation of individual elements with the corresponding syntactic and phonological structures. This also means that Jackendoff can accommodate borderline cases such as compounds with implicit predicates as intermediate cases, whereas Lieber has to assign them to one of the two categories. In Jackendoff’s framework, taxichauffeur in (17) can be assigned the same conceptual structure as taxi driver, but in English the predicate DRIVE is coindexed with syntactic and phonological equivalents, whereas in Dutch it is not. In Lieber’s framework, taxichauffeur is a primary compound like fietsenrek in (2).

Another striking difference between Lieber’s and Jackendoff’s approaches is that Lieber aims to build up the meaning of a compound from the meaning of its parts, whereas Jackendoff describes the meaning of the compound, referring to the meaning of the parts. For Jackendoff, word formation rules are first of all redundancy rules, which is very useful when trying to describe compounds with specialized meanings; for example, strijkkwartet in (22). As she emphasizes the compositionality of word formation rules, Lieber makes a much more clear-cut distinction between the regular meaning and the added specialization.

The comparison of Lieber’s and Štekauer’s frameworks is in a sense much easier, because there is a single, obvious difference that is at the basis of their perspectives. Štekauer’s perspective is onomasiological and Lieber’s is semasiological. Whereas Štekauer’s onomasiological types (OTs) are a stage in the decision process determining the name to be used for a particular concept, Lieber, as mentioned above, aims to derive the meaning of a compound from the meaning of its parts. Both approaches have a role in modelling the speaker-hearer’s use of word formation rules, but Štekauer concentrates on the speaker’s perspective and Lieber on the hearer’s. The onomasiological perspective has the advantage that it can straightforwardly explain the emergence of specialized meanings. This is most obvious in strongly lexicalized compounds such as strijkkwartet. The composition of the ensemble does not have to be derived from the form, because it was the starting point of the decision process. A semasiological orientation can only aim to encode the part of the meaning that can be derived from the form, leaving the rest to world knowledge.

The third pairing, Jackendoff’s and Štekauer’s theories, is arguably the most difficult to compare. As noted in section 12.2, both are opposed to Lieber’s theory, but not in the same dimension. It would be wrong, for instance, to characterize Jackendoff’s theory as semasiologically oriented. The parallel nature of PA makes it inherently non-directional. In this sense, one could see PA as declarative, as opposed to the procedural, directional approaches of Štekauer’s and Lieber’s theories.

The semantic representations in Jackendoff’s framework are much more detailed than Štekauer’s. This can be explained by the different aims. Jackendoff represents semantic structure and turns to word formation because it has semantic implications. Štekauer accounts for the naming process in word formation and uses those aspects of semantics that guide the decision process leading to a name for the concept. This is why it is sufficient to use labels such as Location and Action.

An important difference is also the approach to naming. Jackendoff interprets word formation rules as formation rules or redundancy rules. A formation rule is a lexical entry. As such, as shown by Jackendoff (Reference Jackendoff2002: 152–182), it is in principle not distinguished from words, idioms or syntactic formation rules. Lexical entries provide the building blocks for the generation of triples of phonological, syntactic and conceptual structures. A redundancy rule, as elaborated by Jackendoff (Reference Jackendoff1975), is an emergent generalization about lexical entries. It is not a rule that can be applied to produce a new form or a new meaning, but it reduces the cost of storing entries. This means that the formation of a new lexical entry can only be modelled as the storage of a complex expression in the lexicon. When the frequency of use makes it more efficient to retrieve a complex form from the lexicon rather than building it up by means of rules in the course of processing, we get a new entry. Naming does not play any autonomous role in this process.

For Štekauer, by contrast, naming is the essential trigger of word formation. It is the need to name a new concept that determines the use of word formation rules. Word formation rules supply possible names and impose certain conditions, but they do not determine the meaning of the resulting name, because this meaning has been determined before the naming process started. This gives us a very elegant and compelling account of the phenomenon I call onomasiological coercion (cf. ten Hacken Reference ten Hacken2013b). It is onomasiological coercion which explains that fietsenstalling and boekenlegger in (9) do not have the same meaning structure as vliegtuigkaping and wijnkenner in (13). The reason why fietsenstalling is not an action noun is not to be found in the word formation rule but only in the concept to be named. In Jackendoff’s framework, only the resulting meaning is represented.

On the basis of the three comparisons, we can arrive at the following brief characterizations of the three frameworks in their role of accounting for the semantics of compounding.

Lieber’s framework starts with the form of compounds and tries to derive as much as possible of the meaning from the skeletons of the basic parts. The more articulated the skeletons, the more successful this approach is. Therefore, compounds with verbal elements are the object of special attention. As far as the meaning cannot be derived from the coindexation of elements in the skeleton, it has to be supplemented on the basis of the body of the individual components.

Štekauer’s framework takes naming as its starting point. It is a model of the decision process leading from a concept to be named to a name for the concept. In the various steps, semantic features are selected to fit into predetermined onomasiological structures. Once an OT has been chosen and semantic features attached to its part, we can choose the morphemes to constitute the name. Compounding does not play any specific role here, but compounds can be accounted for. In fact, most if not all OTs can lead to compounds, so there is no natural class of compounding here. However, compounds in a particular OT can normally be characterized in terms of, for instance, primary or verbal compounds.

Jackendoff’s framework is a general theory of language, developed from a theory of semantics. As such, it assigns no special role to compounding, or indeed to word formation. It covers the meaning of expressions in detail, without making a principled distinction between verbal and primary compounds. Word formation rules are used as redundancy rules to reduce the impact of lexical entries on storage capacity. In the case of compounding, they can also be used in the same way as syntactic rules are. New entries for compounds occur when they are used frequently, not because they fulfil a naming need.

As a final thought, let us consider how these frameworks are in competition or could supplement each other. Given the opposite perspectives, it would seem natural to see Lieber’s and Štekauer’s frameworks as supplementing each other, giving the hearer’s and the speaker’s sides of processing, respectively. It should be noted, however, that an at least equally natural view is that the speaker will consider the hearer’s perspective in choosing a name for a concept and the hearer will try to reconstruct the speaker’s reasoning. Therefore, neither of the two needs the other to give a full account of word formation.

Jackendoff’s theory has a much larger domain. The domain Lieber’s theory covers is included in Jackendoff’s, but they make incompatible background assumptions. Therefore, it would be very hard to use the two theories together. This is not the case with Štekauer’s theory, because Jackendoff does not cover naming. However, it would require PA to include a separate word formation component, along the lines proposed in ten Hacken (Reference ten Hacken, ten Hacken and Thomas2013a), which is not in line with Jackendoff’s views of what he calls “semiproductivity”.

Although in some respects the three frameworks are more or less compatible, researchers in the semantics of compounding will therefore have to choose one of the frameworks as their basis. Considerations from other frameworks can only be integrated on an individual basis.