CHAPTER 10

From Syntax to Semantics

10.1 A Review

The micro-graphics world illustrated several features and techniques which are widely applicable: (a) The translation of natural English into some other form of representation which was more useful. By 'more useful' we mean that the new form of representation can be processed readily to create some desirable end product (in our case a screen display). The form of representation was provided by an (assumed) existing system, and so we could concentrate upon the translation problem.

(b) The development of a specialised grammar (specifically for the subset of English involved) using both BNF and ATN notation.

(c) The realisation of an ATN in the form of a program (POPll and Prolog), with side-effects which are the way the internal text or representational form is generated.

(d) The problem of ambiguity and the need for 'back-tracking'. How back­ tracking can be achieved by means of recursion.

(e) Reference and the object of reference being a segment of the internal representation.

(f) The need to identify and isolate a segment of internal representation and label it as the object ofreference (or referent). We note that this ability to 'lasso' an arbitrary segment of code in the internal representation is of fundamental importance and requires complicated pattern-matching facil­ities. The entities identified in this way could be objects or actions.

(g) The techniques of pattern matching and the need for recursive (back­ tracking) mechanisms. The idea of instantiation.

(h) We noted that as there was an increase in the complexity of the subset of natural language being tackled, there was a requirement for a correspond­ing increase in the complexity of the representational form. This was evident as we introduced progressively the representation of colour, movement, objects which undergo changes of shape and form, and the notion of object ­histories extending over time.

(i) We introduced the idea of time-stamps, and noted that the properties of relative time were preserved (locally) for an entity when we copied and re­located it in absolute time.

(j) We introduced the idea of a 'concept' as a segment of internal representation for which the parameters had not been assigned specific values. We described this as a 'generic' representation - that is, capable of generating any specific example of the concept by the assignment of particular values to these parameters and by adding additional non­standard properties.

(k) We discussed the tense of verbs in terms of the time-stamps already introduced.

(l) We discussed the structure of a noun phrase, fIrst in terms of the micro­graphics world but also in general terms.

(m) We have developed a general grammar for a significant subset of English which is not confIned to the requirements of the micro-graphics world.



10.2 NL System Applications

The micro-graphics world allowed us to avoid asking the most important question of all. What do we require of an NL system?

NL systems have been developed for many applications including:

(a) Translation from one natural language to another.

(b) Processing natural language descriptions (newspaper reports) and reducing these to a standardised format which is appropriate for a conventional information retrieval system.

(c) The reduction of scientific reports to a standardised form for the same purpose.

(d) Processing natural language queries to an information retrieval system or database system and mapping these into a conventional query language.

(e) Processing student answers to questions posed by a computerised assessment system, and matching these answers to standard answers supplied by the author of the assessment system.

(t) Processing natural language statements in order to produce a paraphrase of the original statement (also in natural language).

(g) Conducting a dialogue with people in order to elicit information.

(h) The control of robotic devices in natural language.

Some of these applications are intensely practical and some are of purely academic interest. In each case the form of internal representation used is different and designed specifically for the application.

Although some of the systems actually developed have been moderately successful, the general goal of achieving a human-like performance remains beyond the current state of the art. One of the best known and successful systems yet developed was that due to Winograd. It was essentially a robot control application with a very restricted domain of objects which required manipula­tion. In it every reference is identified as a reference to some object within the domain.

Winograd himself has pointed out the limitations of his system. For example, it has no way of dealing with objects which do not exist within the domain, such as imaginary objects. Winograd's own example is the statement 'I want to own the fastest car in the world'. The simple-minded representation of this is to search the list of representations of all cars in the world and select the fastest as the referent. Yet the speaker is not referring to that car, he is actually referring to a car in his own imagination which is faster than the (currently) fastest car.

Charniak attempted to apply the techniques which Winograd had developed to another domain of discourse - the understanding of children's stories. He found that the techniques were totally inappropriate, and in a very influential thesis analysed the reasons for the failure of the techniques to transfer to another domain. The main reason was that an NL system requires the kind of understanding of its domain possessed by humans. In the case of the simple 'blocks world' used by Winograd there was not much to understand. But in the case of the children's stories the domain was the real world, and the wealth of information required for a proper understanding was overwhelming. In the real world people do not say exactly what they mean. They refer obliquely to things and expect others to make the appropriate deductions, and this comes so naturally to humans that we often overlook the fact that it is so.

The history ofNL processing is a story of optimism giving way to a realisation that the volume of information, and the capacity to process that information required by a human-like system, is very much greater than had been anticipated.

In the early stages of NL processing, workers in the field limited their goals. It was felt that topics such as metaphor, beliefs and intentions could be excluded as unnecessary complications, at least in the first instance. It has turned out that trying to produce a human-like system without these aspects is like trying to ride a stationary bicycle. It may seem that the movement of the bicycle is an unnecessary complication, but without it the problem is actually insoluble.

We can achieve limited success for limited goals. We can write an 'intelligent' front end system for a database, or process newspaper reports into standard form, but these systems do not match human performance. The essence of human performance is its flexibility. When we provide a human with some information there is no artificial limit on the way that information will be used. The human can choose to behave like a robot, translate the statement into another natural language, paraphrase the statement, match it to another statement which means the same thing, parse it, answer the statement with another, or ignore the statement altogether. A system which can mimic only one of these behaviours is not human-like. The only way for a system to behave in a human-like way is to process the natural language in the same way as a human would do. This does not mean that the actual mechanism of the computer system (in hardware terms) must be exactly the same as that of the human brain. It merely means that the mechanism must be an analogue of the human brain processes.

It is clear that the way humans represent the meaning of a statement is much more complex and much more general in its applicability than anything we have looked at yet.



When a human chooses to behave in a particular way in response to a statement, he or she does so for some reason. Motivation is involved. It may be the desire to please the speaker, or to insult the speaker or impress some other bystander. Without some analogue of such motivation a computerised NL system could not choose an appropriate response.

It is not clear that we actually want to create such a system. The limited and specialised goals of database access, robotic control, question answering etc. may be quite sufficient for our purposes. But the history of natural language processing has also been a record of dissatisfaction with current systems and continuing efforts to improve performance - to make the systems more human­like.

The successful achievement of that goal may not result in a system which we would actually want to use, but may provide us with a better understanding of ourselves.

10.3 An Outline of the next Chapters

In the next few chapters of the book we shall examine a number of techniques which have been used by various research workers. They are introduced not in chronological order but in order of increasing semantic complexity. Each technique has its uses, and if the reader is bent on practical implementations he or she should select and adapt the technique which is most appropriate to the application in mind. There is no sense in developing a system which is more complex than is actually necessary.

The progression of techniques not only exhibits a steady increase in semantic complexity but a steady decrease in the importance of syntactical analysis. It also represents a progression towards a more human-like capability.