A detailed history of ISSCO's research themes

From the beginning, the Institute's work had been somewhat different from that of a conventional University department. Where traditional linguistics since the later 1950's has mainly been concerned with the formalisation of the syntactic structure of a text, paying relatively little attention to semantics or pragmatics, the Institute, starting from the hypothesis that the process of comprehension was fundamental, had developed an approach based on unifying relevant work in linguistics, psychology and Artificial Intelligence with the aim of describing verbal communication as a whole.

This led to much work on knowledge representation and on discourse analysis, and to an emphasis on semantics in linguistics. The basic aim of describing and facilitating man-man and man-machine communication automatically implied that a number of different disciplines were relevant: research workers over the years have included psychologists, mathematicians, linguists and computer scientists. The range of interests was reflected in the list of working papers published by the Institute.

The working papers have a wide international distribution. Many of the early ideas on semantics developed within the Institute can also be found gathered together in a book Computational Semantics (eds. Charniak and Wilks, North-Holland, 1976) which was the result of a tutorial given by the Institute in the spring of 1975.

A further basic postulate of the group's work was that they were not only concerned with communication between people, but also with man-machine communication. Any theory they developed must, therefore, have been testable via programming on a computer. The vast majority of linguists and psycho-linguists were unwilling to work within this constraint, which they regarded as too restrictive.

This instistence that a theory must be testable via a computer programme embodying it, was reflected in the theme of a second tutorial, organised in July 1981, which concentrated on parsing - ways of deriving a specified representation from an original text via a series of programmed operations. The results of this tutorial were published in 'Parsing Natural Language', ed. King, Academic Press (1983).

Much of ISSCO's work can be seen as an attempt at finding formalisms which would allow computer treatment of language. When the Institute was founded (in 1972), semantic networks had just been proposed as a way of representing the sense of words and of whole texts. Thus, early research aimed at elaborating this representation in order to render it adequate for such a purpose (see WP's 10, 19, 20, 23). An inspection of semantic networks revealed some of their weaknesses. The representation certainly does give important information about sentences. It tells us about the agent, the object being acted on, etc. But the representation is static, it can give no dynamic information of how an action unfolds. Yet much of language comprehension is based precisely on knowing the normal pattern followed in an action. The development of Conceptual Dependency (WP's 1-6, 12) theory allowed us to make significant progress towards a more complete representation. The theory proposes that all actions can be reduced to a few (a dozen or so) primitive acts, used within quite complex case frames. Semantic primitives also form an integral part of the theory of preference semantics, also heavily used in the work of the Institute (WP's 17, 18, 30). The use of a structured set of primitives allows the representation of information which cannot be captured in a semantic network, for example that 'travelling' is a particular kind of movement where the agent and object of the action are identical. It is also considerably easier to include inference information in a primitive representation.

But the problem of inference is very complex, since it is often based on world knowledge rather than directly on the semantics of a particular word. Thus, there is a difference in kind between the following three types of inference problem:

1. The soldiers shot the women and I saw several of them fall.

2. The farmer's wife sold the cow because she needed money.
   The farmer's wife sold the cow because she gave no milk.

3. The town councillors refused a permit to the women because they feared violence.    The town councillors refused a permit to the women because they advocated revolution.

In the first example, the semantics of the word 'shot' might be used to decide that the 'them' is 'women'. In the second, although part of the meaning of cow might be 'animal giving milk', it is not so easy to say that part of the meaning of 'wife' is 'person needing money' and it is therefore correspondingly more difficult to imagine resolving the ambiguity via a precise semantic coding of individual words. And in the third example, it is impossible : it is only our general knowledge, based on our own culture, about the likely attitudes of town councillors which allows us to decide who 'they' are in the two separate readings.

A first attempt at solving this problem involved including the result of an action directly in the representation (WP 10), but this still covered only individual words and individual actions. A more satisfactory approach came from the theory of frames (WP's 13, 14, 15, 28, 29, 31). Such a representation allows the inclusion in the frame for 'travel', for example, of our knowledge about train guards, ticket buying or eating on a train.

More theoretical work on precisely what kind of world knowledge must be included within a knowledge representation can be found in some of the papers already mentioned and in WP's 34, 35 54, 55, and 56. Similarly a discussion of an analogy between frame theory and human cognitive structures can be found in WP's 32, 44.

Go to page:     1    2    3