Thematic role relations

  Much of Jackendoff's analysis is motivated by thematic role relations, i.e. the actor vs. patient relation and motion and location roles. He argues that actor/patient relations are orthogonal to other thematic relations, and formalises this distinction explicitly by introducing an Action tier and a Thematic tier into the representation of Events. The Action tier captures the actor/patient relations, while the Thematic tier captures the motion and location relations. The question for a computational linguist is whether it is necessary, or, more weakly, simply useful, to represent such relations explicitly in lexical semantic structure.

There has been much research, particularly within the Government and Binding grammatical framework, focused on how thematic relations can be used to explain syntactic phenomena such as binding and control (e.g. gruber:65, fillmore:68, Jackendoff jack:83,jack:90, grimshaw:90). There have been other attempts made at explaining these phenomena (e.g. within HPSG), but the results of research in this area are far from conclusive. Thematic relations may prove to be the best way of capturing the phenomena. In that case, it will clearly be in the interest of anyone interested in the syntax/semantics interface to include thematic roles in lexical semantic representation.

An explicit example in which the representation of, at least, Actor/Patient relations seems to be critical can be found in the sentences in ji8 (from Jackendoff 1990:130).

 

What Bill did to the books was load them on the truck. ? What Bill did to the truck was load the books onto it. * What Bill did to the books was load the truck with them. What Bill did to the truck was load it with books.

In each case, the expressed event is essentially the same: the books go onto the truck. What changes, as reflected through the syntactic form utilised, is which entity is viewed as the Patient - which entity is most directly ``affected'' by Bill's action. ji8a-b identify the books as the Patient, as that element (or an anaphoric reference to it) is in direct object position in the embedded sentence, while ji8c-d identify the truck as the Patient. The reason for the reduced acceptability of ji8b-c, then, is that there is a conflict between the entity (X) explicitly identified as being acted upon by Bill, Bill did something to X, and the entity suggested to be the Patient by the syntax of the embedded sentence. This example suggests that at least the semantic notion of Patient is relevant to the mapping between syntax and semantics, and that this thematic relation must be represented in the semantics to account for the distinction between the pairs of parallel sentences above. Parallel examples can be found to support the necessity of representing other thematic relations.

An additional active area of research with respect to the syntax/semantics interface is that of linking theory, a theory attempting to explain regularities in the mapping between semantic and syntactic arguments. So, for example, it is in the domain of a linking theory to explain why there is no verb quain in English such that ji29b is synomous with ji29a (Davis 1995).

 

The child trained/called/petted/fed/kicked the dog. The dog quained the child.

jack:90 and davis:95 inter alia argue that generalisations about linking depend on a lexical semantic representation of thematic roles. Elements of verb meaning, such as causation, can influence the syntactic realisation of a verb's arguments, and it is therefore necessary to embed the notion of thematic roles into a richly structured semantic representation. Linking rules can then use this structural information to determine how semantic arguments surface syntactically.

A principled explanation of the syntax/semantics interface, then, seems to depend on the notion of thematic relations within lexical semantic representations. Adopting linking theory within a computational framework would free the lexicographer from having to specify explicitly in the lexicon how arguments in a verb's subcategorisation frame correspond to its semantic arguments. In fact, it is possible that subcategorisation may be almost entirely semantically determined, thereby largely eliminating the need for ad hoc specification of subcategorisation frames for individual verbs. The predictive capacity of linking theory would therefore clearly reduce the amount of word-specific syntactic information which would have to appear in the computational lexicon. A linking theory is therefore desirable from both a theoretical and a practical perspective.

In sum, the representation of thematic role relations within the lexicon seems to be warranted by its necessity within explanatory frameworks for various syntactic phenomenon, and ultimately for a clearly defined, principled syntax/semantics interface.