CHAPTER 7

Nouns and Noun Phrases

7.1 Why Noun Phrases?

In Chapter 1 the grammars developed for dealing with assignment statements and line definitions were ad hoc and peculiar to the application under consideration. By extending the facilities of the micro-graphics world to deal with events in the past and the future, we have been required to introduce the idea of 'tense', and in consequence we were required to describe a grammar for conventional verb phrases. These can be found in any natural language sentence and are not peculiar to our application. Having discussed verb phrases it is appropriate, therefore, to go on and discuss the noun phrase, as it occurs in all natural language sentences, although we shall still refer back to our micro­graphics application for illustration. The verb phrase and the noun phrase are the main structural units of a natural language sentence.

If we draw an analogy between a sentence and a play (on the stage) then noun phrases are concerned with the parts to be played, and the verb phrase is concerned with the action of the play. The overall structure of the sentence indicates which role is associated with which part. Other bits of a sentence are often concerned with scene setting and with stage directions.

7.2 Entities and Concepts

In section 2.4 we introduced the term 'concept' to describe the structural unit which is in effect the definition of a word such as 'triangle'.




We described this structure as 'generic' because it could be used to generate specific examples of the concept triangle by assigning specific coordinate values to the points X, Y and Z, and then constructing what is in effect a macro­expansion of the definition.

We need to distinguish between the generic structure and the examples which can be derived from it. We call the generic structure a concept, and each example we shall call an entity or object. In the micro-graphics world an entity is always a concrete thing like a triangle or a square, but in natural language in general an entity can be abstract, like 'an idea' or 'an embarrassment'. An entity is usually identified by a noun phrase. For example:


A concept is usually identified by a single noun. For example:

Horse, idea, pictures, bicycle, man

A noun is the most important constituent of a noun phrase. It tells us what generic concept or class of entity our specific example of an entity belongs to. That tells us what properties we can expect it to have.

Other words in the noun phrase may provide us with additional information, and perhaps modify the properties identified by the noun (if they are unusual or non-default). For example, the noun 'canary' tells us that the entity under consideration is a bird, has wings, is yellow, and so on. The words' ..green canary' tell us that although the entity concerned has most of the expected properties of a canary, it is unusual with respect to its colour.

Our discussions in sections 3.3 and 3.4 have shown that an entity can live through several 'incarnations' and that we need some method of 'throwing a lasso' round the lines of code which represent these incarnations so that we can refer to the entity as a whole.

7.3 Proper Nouns

A proper noun (e.g. John, Fred) is a name label or identifier for one particular entity, that is, one for which the parameters have been assigned particular values. In the micro-graphics world a proper noun is represented by a label such as 'P' or 'Q' or 'triangle ABC'. We already have the facilities for dealing with this in our symbol table.

Because a proper noun is the name of an entity, it can form a noun phrase on its own. In general a proper noun must be defined before we can associate it with a particular entity. There are some proper nouns, however, which have certain prior expectations associated with them. The proper noun 'fred' for example, suggests (without definition) that the entity referred to is human and male. The proper noun 'fido' suggests a dog, and so on. There is no reason why these default associations cannot be overridden, but it is interesting to note that in some cases we should be prepared to associate some proper nouns with concepts as though they were normal nouns. For example we might describe someone as 'a Hitler'.



7.4 The Indefinite Article 'a'

The phrase 'a triangle' identifies a single example of the concept triangle. Where there.,is.no .reason to suppose that the triangle has.any particular parameter values we may assign arbitrary values to the parameters. It may be wise, however, not to give such a triangle values which make it a special case. We can think of the triangle as being a random selection from the set of all triangles, but the selection is not entirely random since what is selected will be a 'typical' triangle. What 'typical' means will vary from indiVidual to individual (and from computer system to computer system), but normally it will not be isosceles or right-angled orzero angled, and normally it will be locate roughly in the centre of the screen or the Qlank area of the screen in such a way that if the available area was subdivided as shown in the diagram below the triangle would more or less rul the centre square.

The avoidance of special cases means that if the user makes the statement 'typical triangle', the system will not draw a triangle which is too small to be seen, or which occludes other figures on the screen, or for which one of the sides coinCides with the side of the screen, and so on.

This notion of 'typical' case is inherent in natural language. If we.speak of 'a man', we imagine a 'typical' man. We do not normally imagine a one-legged man, or a blind man, ora giant, or a man with any abnormal characteristic. Indeed, if we are asked afterwards. what was the colour, of hair of the man we imagined on hearing the statement 'a man'; it is doubtful if we could give a simple answer. The probability is that our imagination did not supply that level"of detail. It was simply 'a normal or standard man'.

If we wish to draw attention to some characteristic we speak of 'a tall man' or 'a fat child', 'a large mouse' and 'a small elephant'. The terms 'tall', 'fat', 'large' and 'small' do not correspond to absolute dimensions but are related to the typical dimensions of objects of these various classes. A small elephant is still much larger than a large mouse. Being 'near to' a mountain does not imply a distance of separation of the same scale as being 'near to' another person. Our sense of 'normal' is strongly related to the properties of the object being discussed, and that means influenced by the concept labelled by the noun in the relevant noun phrase.




To represent the entity referred to by the phrase 'a triangle', therefore, we need two things:

(a) A representation using formal parameters which can be given absolute values to reproduce any required triangle.

(b) A set of 'typical' parameter values which can be used as default values if no clue is given to the actual parameter values. They also serve as a standard of comparison for the interpretation of adjectives such as 'small' and 'large'.

We use the generic representation of a concept to generate an entity and assign to it the parameter values we find there. These are the 'typical' or 'default' values of the concept. In the case of our micro-graphics world, these default values may be relative to the free space which is currently available on the screen.

One way to look at the processing of the phrase 'a triangle' is to think of the NL process picking out of its memory a generic concept structure identified by its label 'triangle'. This defines an infinite set of all possible triangles, because the parameters have no values. In the micro-graphics world the set may not actually be infinite because it is bounded by a fmite screen, but it is as good as infinite for practical purposes.

The meaning of the determiner 'a' is represented by a function call which performs a selection process upon that infinite set to pick one typical example. This is the first time we have introduced the idea of the meaning of a word being represented by a function which operates upon the representations associated with other words, but it is a useful idea which we shall utilise frequently.

In terms of the ATN processing of a noun phrase, we can think in terms of side-effects. As the system traces the arc corresponding to the word 'a', it must place a function identifier in a global register. When the next arc corresponding to 'triangle' is traversed, the side-effect must be the placing of the appropriate concept definition in another 'argument slot' of the global register. When the 'exit' arc of the ATN is being traced, the side-effect must be the application of the function to its arguments yielding the internal text required.

This is a much more complicated procedure than the one we described earlier, in which we recognised but otherwise ignored the presence of the determiner 'a'.



7.5 The Defmite Article

The phrase 'the triangle' will be processed in very much the same way as the phrase 'a triangle', but the use of 'the' instead of 'a' introduces a reference to some previously mentioned triangle, and will therefore initiate a search backwards to discover the referent.

The processing is identical for the two phrases up to the point where the internal text has been generated. The search associated with 'the', then begins and uses this internal text as a template for the matching process. The interpretation of 'the' therefore involves the use of two functions, one to select the entity example from the infinite set of possible examples, and the other to carry out the search. The adjective 'red' in 'the red triangle' adds information to help make the search more precise and resolve any ambiguity.

The use of the definite article 'the' also refers to and determines the focus of attention. This explains the special effects which can be obtained by the use of 'the'. For example:

'The tall man was approaching steadily, and as he did so he glanced towards Julia.'

If this was the first sentence in a story where there is no previously mentioned 'tall man' to which this can refer, then the writer is using 'the' to establish the 'man' in question as the focus of attention.



7.6 Standard Parameter Values

The colour adjectives in our graphics language example are absolute (e.g.: 'red' is always a standard red). But in other contexts this cannot be assumed. The 'red' of 'red hair' is not the same red as the 'red' of 'red sun'. Each is interpreted as:

RED X = 'redder than the standard colour for objects of type X'.

There will not only be a standard colour for each object type but a standard range of variation. This allows us to interpret the phrase 'a very red sun'.

One approach which might be adopted for the representation of these standard values uses the same type of mechanism which we used in the representation of time (Chapter 4). We might provide each standard parameter value with an arbitrary unique label, and then provide information about how these are related to each other. Thus we can have STD-RED-I and STD-RED­2, and a relationship which tells us that STD-RED-I > STD-RED-2 (with respect to redness). The same method will serve for measures of 'nearness'. Each object will be associated with a normal or standard distance of separation. Eachcan be assigned a unique label STD-DIST-I, STD-DIST-2 and so on. To describe an object as 'near' we assign to the distance of separation a new unique label (DIST-3) and then note that DIST-3 < STD-DIST-l.

Such a technique might suffice for a very limited micro-world in which there are only a few standard values to be considered, but for the world in general it is inadequate. The prior assignment of a specific standard size and a standard distance of separation to every object would be a mammoth task. Every parameter value must have its standard value, and so this idea of prior assignment would overload our system with unique values which may never be required. It would be preferable to have some way of generating only those values we actually need, as the need arises.

A possible solution lies in the idea of having a function which takes an object label and the property concerned as arguments, and yields a unique standard value as a result. We can then refer to the unique standard value by means of the function name and its arguments, instead of referring to it by the unique label which it would generate. In effect the activation of the function (or its application to its arguments) is held in suspense, and the function call itself is used as the label for its result. If we wish to represent a value 'red' of some object X of concept Y which is redder than standard red for some such objects, we fIrst give the object X a colour parameter value = R, and then write R > STD-RED(Y).

This does not solve the problem completely. It removes the need to assign unique labels to every possible standard value, but it leaves us with the problem of indicating what the symbol '>' means. For simple numeric values its meaning is obvious, but in other cases it could mean 'redder' or 'greener' or 'darker' or 'lighter' or 'bigger', or 'farther' and so on. Its interpretation is in fact related to the parameter values we are considering.

One possibility is to replace '>' by a specific parameter-related predicate:

redder(R,STD-RED(Y))

Another possibility is to associate each parameter with an 'axis of discrimina­tion' (in this example - degree of redness). We will discuss a way of implementing this idea later (Chapter 21 onwards).

Functions which replace a possible infmity of parameter values are known as 'Skolem Functions'.



7.7 Adjectives

Loosely speaking an adjective is a 'describing' word. More strictly, and in terms of the representational approach we have been describing, the meaning of an adjective is a modification carried out on the parameter values of a concept representation.

Consider the processing of the words'. ..red triangle'. The traversal of the ATN arc corresponding to 'red' will create as a side-effect a representation of a concept (say X) which has only one known property - it has the colour red. The tracing of the next arc corresponding to the word 'triangle' generates as a side-effect the generic structure for the concept 'triangle'. The next process is to try to unify these two structures, and the simplest way to do that is to unify the unspecified concept (X) with the concept 'triangle'. The result is a new concept (red triangle). Note that it is the representation of a concept, not an entity. The word 'a' or 'the' is required to process the concept representation and turn it into an entity.

The virtue of this two-step mechanism is that it deals with situations in which there is no concept to which (X) can be unified. Consider the sentence 'The red is very bright'. Here there is no concept specified, which is being described as 'red'. But there is a concept of something. It is after all impossible to imagine the colour red without also imagining something which is that colour. (Even an unbroken field of colour is still a 'something'.)

There is not really a great deal of difference between an adjective and a noun. The difference lies in the roles which they normally play within a noun phrase, and depends on which is the dominant role. Consider, for example, the two phrases 'the red pillar-box' and 'the pillar-box red'. In the latter case the result of interpretation will be an unspecified entity with the parameter value equal to a particular kind of red (the standard colour of pillar-boxes). The focus of attention is on the colour, not the entity. .

Since nouns identify concepts, and concepts have default parameter values, a noun can be used as an adjective to apply that default value to some other concept which does not have that particular value as its default value.

7.8 Gerunds

A gerund is a verb which is being used in the role of a noun. We might refer, for example, to 'the moving of the triangle'. Here the noun phrase 'the moving' has no noun, but has instead the present participle of the verb 'to move', In these circumstances the noun phrase is referring not to an object, but to a segment of internal text which describes an action carried out on an object. The prepositional phrase 'of the triangle' is appended to, and is part of, the noun phrase, it provides additional information (to allow identification of the reference entity or segment of code).

As in the case of the adjective, we see that the important aspect, from the point of view of the interpretation, is the role that the gerund plays within the noun phrase. The word 'moving' describes the process to be carried out on some entity. If this entity is unspecified (e.g.: 'The moving was completed') we must, as for the adjective, create a representation of some dummy construct representing a movement of some kind. Its properties are provided by the prepositional phrase 'of the triangle'.

A classic case of ambiguity occurs in the case where it is not clear which aspect of a concept is being described by the gerund. This ambiguity is illustrated by the example 'The shooting of the hunters'. Here it is not clear whether the shooting is being carried out by the hunters or being carried out on the hunters, This problem is generated by the fact that the hunters are entities which can both shoot and be shot, and so we do not know how to unify the structures for the two concepts. There is no similar ambiguity with the phrases 'The shooting of the rifles' or 'The shooting of the targets',

These examples illustrate the importance of semantic analysis and the impossibility in some cases of determining the correct syntactical analysis in the absence of semantic considerations.



7.9 The Syntax of a Noun Phrase

Once we understand what we are trying to do to a noun phrase (np) let us see if we can establish its word pattern (or grammatical structure). In its simplest form we have phrases such as:




To save ourselves the bother of writing out many alternative structures we shall use the notation:




where the '?' symbol indicates that the adjective mayor may not be present. Of course the adjective can also be repeated, e.g. 'the large red triangle', and so we shall use the ,*, symbol to indicate that a term may be repeated several times.




Next we notice that the adjectives often occur in pairs: 'the fast moving triangle'. Here the word 'moving', which is not normally thought of as an adjective, is being used as part of a word group 'fast moving', and that word pair forms an adjectival group. There are many more such adjectival groups, e.g. 'right angled', 'ham fisted', 'double talking', 'barrel chested', 'clear thinking', 'man eating', 'fly bitten', and so on. Note that in each group the second word is a verb and the fIrSt word is either an adjective or a noun.

Consider first the noun-verb pairs. These fall into two categories - those with the present participle (e.g.: 'talking') and those with the past participle (e.g.: 'chested'). In the case of the present participle the noun of the noun phrase pair is the entity carrying out the action, and the noun of the noun-verb adjectival group is the recipient of the action. In the past participle form the noun of the noun phrase is the recipient of the action described by the verb. For example, 'a man eating tiger' and 'a man eaten tiger' are two very different things. In some cases the verb of the adjectival group does not normally have an agent (e.g.: 'fisted' (with exceptions». In these cases the noun of the noun-verb adjectival group is being used as an adjective. The first word of the adjectival group word­pair can be an adjective or an adverb (e.g. 'clear' in 'clear thinking').

These are complex problems which we shall discuss in more detail later. For the moment we are interested in establishing a usable syntactical structure. In view of the possibility of the adjectival group, our grammatical structure ought




to replace the simple adjective with an adjectival-group component (adj-group):




Next we note that the determiner is not present if the noun phrase refers to more than one entity - 'triangles have three sides', or 'men were breaking down the door'. Therefore we ,have:


Lastly we replace the simple 'noun' with the 'noun-group' which includes the possibility of it being a gerund.




To these we can add prepositional phrases, e.g. 'the triangle above the square'.




where 'pp' is a prepositional phrase with structure given by:




This grammar is also represented by the ATN shown in Figure 7.2

One of the difficulties associated with a grammar such as this is that it is not easy to show the agreements which must exist between the various options. For example, the fact that a determiner may be present or not present may be shown clearly, but what is not shown is that the determiner is not present when the noun phrase is plural. Likewise the adjectival group used must agree for number with the noun, and so on.

One way to cope with this is to subdivide the classifications noun, adjective, determiner etc. into sub-classifications, and refer to these in the grammar. Thus we can have:




and so on. With these sub-classifications we can be more precise about what is grammatical and what is ungrammatical. It is possible to take this strategy to ridiculous lengths, and it is our belief that many grammarians have done exactly that. The end-point of this process will be reached when every word is in a classification of its own, at which point syntax will (virtually) have become semantics.

In Chapter 9 we have provided a reasonably detailed grammar of English which uses many more sub-classifications than we have described here, but it is certainly not complete. Much of the latter part of this book represents an attempt to abandon the classical approach to NL processing, which is based on syntactical analysis, and to develop a purely semantic approach.

In Chapter 8 we discuss the problem of pattern matching which emerged clearly in this section as a requirement. We also consider better ways to implement a grammar.