A Semantic Ontology

  Jackendoff (1983, 1990) proposes an ontology which provides the building blocks of conceptual representation. Each representational constituent corresponds to one of the ontological categories: Thing, Event, State, Action, Place, Path, Property, Manner, or Amount. These categories are motivated linguistically in several ways.

1. Each of the ontological categories permits the formation of a wh-question ji3 (from Jackendoff 1983:53), except State. States are characterised by their contrast with Events, as shown in ji4.

2.  
3. What did you buy? Thing
4. What happened next? Event
5. What did you do? Action
6. Where is my coat? Place
7. Where did they go? Path
8. What was she like? Property
9. How did you cook the eggs? Manner
10. How long was the fish? Amount

11.  
12. What happened was that Events
13. ?What happened was that States

14. Each major syntactic constituent of a sentence corresponds to a conceptual constituent. Thus in John ran toward the house, John and the house are Things, the PP toward the house is a Path, and the entire sentence is an Event (example from Jackendoff 1990:22).
15. Several of the ontological categories support quantification, as in ji5 (from Jackendoff 1990:23).

16.  

17. Every dinosaur had a brain. Things
18. Everything you can do, I can do better. Actions
19. Anyplace you can go, I can go too. Places
20. Bill can shuffle cards in every way that bill can. Manner

These categories overlap to a certain degree with the types already assumed in standard HPSG, although there is not an exact equivalence as Jackendoff's categories concentrate on more fine-grained semantic distinctions than are incorporated in HPSG. They can, however, be integrated into HPSG quite straightforwardly, as we will see in Section 2.3.1.

Jackendoff's conceptual structures are built up using constituents in the ontology. Each constituent can be decomposed into function-argument structure. Functions impose conceptual constraints on the nature of and relations between function argument(s), which are themselves required to correspond to conceptual constituents. So, for example, a Place may be characterised as Place-function(Thing) and a Path might be characterised as Path-function(Place). Jackendoff seeks to identify functions which explain grammatical patterns of combination, using the ontology as a reference point. The existence of such an ontology provides a starting point for the identification of generalisations about how words can combine to form phrasal constituents and how phrasal constituents can combine to form larger constituents. Due to the creative capacity of language, it is clear that we do not simply learn lexically specific ways of putting words together -- there are regular, productive patterns of combination which apply to classes of words and phrases. By allowing for the mapping of words and phrases to general classes, the ontology supports the identification of such patterns of combination.

Why is such an ontology useful? It divides the linguistic domain up into types of entities to which each sentential constituent may correspond, thereby enabling the definition of broad constraints over the mapping between syntactic constituents and semantic entities. It further allows the definition of functions ranging over ontological entities, which indicate the ways in which entities can combine semantically. These functions are not fixed in advance by the ontology, but rather can be determined via linguistic analysis, in order to capture the range of relations expressible in language. The output of such functions is constrained to correspond to an entity in the ontology, thereby ensuring that there are at least very broad restrictions on what functions are possible. Thus the ontology reflects the types of entities which are expressible in language and guides the identification of coherent relations among these entities.

The definition of the ontology which Jackendoff uses stems in part from psychological claims about the projection of entities in the real world to a mental representation of those entities. Although the cognitive processes of categorisation and so forth are possibly not of direct concern to the computational lexical semanticist, the linguistic motivations for a semantic ontology as introduced here and the framework which such an ontology provides are reasons for adopting one. There are in addition semantic reasons for the adoption of an ontology, stemming from ``pragmatic anaphora'' (Hankamer and Sag 1976), in which anaphors like this and that can refer to entities of specific ontological categories compatible with the linguistic context, and psycholinguistic and developmental studies (see Pinker 1989 for an overview). Abstraction over cognitive structures is necessary to account for children's ability to learn language. Furthermore, Talmy talmy:83,talmy:88 has shown that most verb meanings cross-linguistically are built around recurring elements of meaning and their combinations. These studies strongly suggest that the adoption of an ontology is necessary to capture generalisations about language use.

The ontology as Jackendoff has proposed it, however, does not assume any explicit subdivision of the ontological categories, although he implicitly assumes a richer structure. These kinds of relations can easily be captured in a hierarchical ontology which has the conceptual categories introduced by Jackendoff at its top. This ontology would reflect categorisations of entities, events, etc. and relations between them. It therefore captures (a part of) world knowledge. Such an ontology is basic to constraint-based theories of grammar (e.g. HPSG) and have been widely used in computational approaches to natural language processing (see Copestake copestake:93b for an example). I will assume that an inheritance-based hierarchy is a critical representational component.

However, the lexical ontology which I will assume will be a linguistic ontology, following HPSG, as opposed to a world knowledge ontology, as emphasised in the introduction to this chapter. This means that information which is specifically relevant for the relationship between word forms and meanings will be represented here. For drink, for example, we must encode the information that the intransitive form of this verb has a specific default interpretation. That is, without more specific information ji31a means ji31b (Lascarides and Copestake lasc_copestake:97).

 

John drinks all the time. John drinks alcohol all the time.

This default interpretation cannot be explained solely on the basis of world knowledge since probabilistically drinking alcohol is not prima facie more likely than the drinking of other kinds of liquids. I will take advantage of the generalisations that can be made using an ontology, incorporating the notion of defaults which are associated with individual (groups of) lexical items.