Thematic relations and conceptual structure
So far, none of the theories we have discussed uses verb meaning to deter mine argument selection directly. In the traditional theta-role system, verbs are associated in the lexicon with theta-grids, and it is the hierarchy that determines how arguments get into a verb’s subject and object positions. As a result, even though the investigator uses the verb’s meaning to work out its theta-roles, and then uses the thematic hierarchy to work out its argument structure, the crucial dependence of argument structure on meaning isn’t explicitly formalized: argument structure cannot be read directly off the verb’s meaning, but proceeds via the intermediate stage of theta-roles.
This type of theoretical arrangement is somewhat redundant: looking at a verb’s meaning provided the justification for the theorist’s decision about what theta-roles a verb was given, but the precoded roles still constituted a separate aspect of the verb’s lexical representation, independent of its meaning. The precoded roles only exist for the syntactic purpose of getting the ‘right’ arguments as subject and object. In Dowty’s system, something similar is true: the theorist has to examine the verb’s meaning and decide in each case how many of the proto-role entailments are satisfied. There is no way in which argument structure can follow automatically from the verb’s meaning: since neither traditional theta-role theories nor Dowty’s proto-role system offers any way to represent verb meaning explicitly, a direct mapping from verb meaning to argument structure is impossible.
Explicit representation of verb meaning is therefore a prerequisite for a more unified theory of linking. Jackendoff (1987, 2002) offers just such a theory. As discussed in 8.1, Jackendoff’s theory of semantic representation involves a decomposition of meanings into primitive elements – BE, GO, CAUSE and so on. As we will see, according to Jackendoff it is the nature of the primitive elements within the semantic decompositions of verbs which directly determines argument structure. This allows Jackendoff to dispense completely with theta-roles or proto-roles, and to derive argument structure directly from semantics. Jackendoff explains the virtue of this as follows:
What is at stake is the issue of language acquisition. If a word’s syntactic behavior (including its syntactic argument structure) were always tightly linked to its meaning (including semantic argument structure), there would be far less lexical idiosyncrasy for the child to learn, always a desideratum . . . ( Jackendoff 2002: 138)
The primitives of Jackendoff’s system provide a new way of thinking about the nature of theta-roles. Consider Jackendoff’s representation of the conceptual structure underlying the following three sentences (also discussed in 8.1.2);
Sentence (18a) is the conceptually simplest, consisting simply of the state function BE with two semantic arguments: Thing (which refers to the door) and Property (which refers to openness), as diagrammed in (19):
(To say that Thing and Property are ‘semantic arguments’ is to say that they are necessary complements of the function BE – concepts unavoidably bound up with its meaning in this use: see 8.1.1 for explanation.)
Jackendoff defines the thematic role of ‘Theme’ as ‘the first argument of the functions GO, STAY, BE and ORIENT’ (1987: 378). The conceptual representation of (18a) contains the element ‘BE’, with two arguments, DOOR and OPEN. This means that the first argument in (18a), DOOR, which is realized by the lexeme door, is interpreted as Theme.
Now consider (18b). This has the same conceptual structure as (18a), except that it has added the ‘INCHoative’ function (Latin: ‘start’), which denotes the coming into being of an event. The door opened thus receives the following analysis:
The addition of INCH makes no difference to the thematic roles in the sentence: DOOR/door still satisfies the definition of ‘theme’.
Sentence (18c) adds the ‘CAUSative’ function to the previous structure:
CAUSE is a two-argument function: the first argument is a Thing (here, John), the second an Event (here, the door becoming open). Jackendoff defines Agent as the first argument of the Event-function CAUSE. This makes John Agent in (18c).
The other thematic roles are dealt with in exactly the same way. Source, described above as ‘the entity from which motion takes place’, is analysed as the argument of the Path-function FROM. Goal, the ‘entity towards which motion takes place’ is the argument of the Path-function TO. The experiencer argument, Jackendoff says, ‘presumably is an argument of an as yet unexplored State-function having to do with mental states’ (1987: 378).
This approach has an important consequence: ‘Agent’, ‘Theme’ and the other roles can be eliminated from the theory. Since Jackendoff defines the theta-roles in terms of underlying conceptual structure, it’s the under lying conceptual elements, not the theta-roles themselves, which do all the explanatory work. As Jackendoff puts it, the names of theta-roles ‘are just convenient mnemonics for particularly prominent configurations’ of underlying conceptual primitives. This means that we can restate the thematic hierarchy in terms of underlying conceptual configurations. Instead of Agent, for example, we substitute the first argument of the predicate CAUSE; instead of Theme, we substitute the single argument of GO, STAY, BE or ORIENT. Jackendoff is claiming, in other words, ‘that the terms Theme, Agent, and so on, are not primitives of semantic theory. Rather, they are relational notions defined structurally over conceptual structure’ (2002: 378–379). He compares this to the way ‘subject’ and ‘object’ are not primitive notions of generative syntax, but structural positions in the classical tree structure: subject is the NP immediately dominated by S, and object the NP immediately dominated by the VP.
The types of semantic distinction conveyed by the different theta-roles or, for Jackendoff, underlying conceptual configurations, are not the only semantic connections between a verb and its arguments. Verbs also place selectional restrictions on their arguments: the object of eat must be a solid, that of drink a liquid, that of read something visible, that of pay (in one of its uses) an amount of money. In Jackendoff’s theory of conceptual structure, these selectional restrictions are, like theta-roles, specified directly by the conceptual structure: they are not extra information which needs to be learnt in addition to the meaning of the verbs themselves. Jackendoff observes that a sentence like Bill paid inherently entails the information that what Bill paid was an amount of money; similarly, Harry drank contains the information that what Harry drank was something liquid. This shows that if an argument isn’t expressed lexically, information about its nature is still available. As a result, the selectional restrictions on the objects of drink and pay are ‘essentially explicit information that the verb supplies about its arguments’ (1987: 385). Selectional restrictions are ‘part of the verb’s meaning and should be fully integrated into the verb’s argument structure’ (1987: 385).
The lexical entry for drink, for example, is something like this (Jackendoff 1987: 386):
The Thing arguments which can be subcategorized by the verb are marked by the indices i and j. The first of these, indexed i, is the first argument of the CAUSE predicate: this makes it an Agent. The second, indexed j, is the first argument of the GO predicate: this makes it a theme. But the parentheses around ‘NPj’ in the third line of the representation indicate that drink doesn’t have to have a fully expressed direct object – the verb can be intransitive. But whether it does or not, the verb’s meaning itself contains the information that the thing being drunk is a liquid.
We can see another illustration of this approach to argument structure by considering the difference in meaning between the verbs butter and bottle, as in Harry buttered the bread and Joe bottled the wine. The meanings of these verbs are quite different: butter means ‘to put butter on something’; bottle means ‘put wine in a bottle’. Jackendoff represents the conceptual structures as follows (1987: 387):
In (23a) the subject argument (Harry in our example) is indexed with the letter i. As the first argument of CAUSE, this identifies the verb’s Agent. (Recall that terms like ‘Agent’ are just convenient shorthand here; what we mean is ‘first argument of a CAUSE predicate’.) The verb’s object – the bread – is also indexed, showing that it is subcategorized. As the argument of a TO function, bread would traditionally be called a Goal. Unlike the other two semantic arguments, the Theme argument, BUTTER, does not have an index linking it with an argument. This means that it is not connected to a subcategorized position. Jackendoff explains that, as a result, ‘this argument is totally filled in with information from the verb and is understood as “nonspecific butter” (1987: 387) – when we are told that Harry buttered the bread we don’t know anything about the identity of the butter involved. This contrasts with (23b). Here it is the Goal argument – the argument of the TO function – that doesn’t have an index. This means that it gets its interpretation entirely from the verb: all we know is that Joe bottled some specific wine, but we know nothing about the identity of the bottle into which the wine was put. Jackendoff notes that com paring these examples shows us that ‘the similarities and differences between butter and drink fall out directly from the notation adopted here. There is no need to interpose a level of argument structure to encode them’ (1987: 387). Argument structure and selectional restrictions are not separately coded pieces of information that have to be learnt as well as the meaning of the verb; they are part of the meaning of the verb itself. As a result, the learner’s task is simpler.
