CHAPTER 11

Case Grammar

11.1 Case: Syntax or Semantics?

It is considered ungrammatical to say or write: 'I done it'

'You goes'

'She did it to he'

The notion of 'case' was introduced into traditional grammars to describe this phenomenon and give precision to the restrictions which render these statements ungrammatical. Three cases were recognised - the subjective, objective and genitive (or possessive). For example, the pronoun 'I' is always associated with the subject of a sentence. 'Me' is always associated with the object of a sentence and 'mine' is the genitive form. In this type of analysis 'case' is a syntactical phenomenon.

More recently, however, an argument has been developed for a notion of 'case' which is associated with some role which an entity is playing in the events or situation described. As such, 'case' is a semi-semantic notion, and the role refers to the entities which are identified as the meanings of various noun phrases in a sentence rather than to individual nouns. In this form of analysis the main verb determines the various case roles which can be associated with the nouns or noun phrases in the sentence.

Consider for example the two sentences: SI 'I did it'

S2 'It was done by me'.

Traditionally, in S1 'I' has subjective case and in S2 'me' has the objective case. But, in S1 'I' is the agent and in S2 'me' is also the agent (i.e. the person doing the action).

11.2 Traditional (or Syntactical) Case Classifications

A sentence with a transitive verb has, in its simplest form, the pattern: (Subject) (verb) (object)

The main noun in the noun phrase which corresponds to the (Subject) slot is said have the subjective case. The importance of case, for syntactic analysis, lies in the fact that when pronouns are used they must be used in the form which agrees with the case slot they fill.

subject = I, we, he, she, they, who

object = me, us, him, her, them, who(m)

genitive = my, our, his, her, their, whose

The possessive or genitive case can also be recognised in nouns, so that we have such phrases as 'the boy's book'. Here the word 'boy's' has the genitive case. If the verb has passive voice the pattern is:

(Subject) (passive-verb) by (object)

The subjective form 'I' is used in both the sentences 'I kicked the boy' and 'I was kicked by the boy' even although in one sentence 'I' is carrying out the action and in the second sentence 'I' is the recipient of the action. This is why we describe this type of case analysis as syntactic.

Where the verb is intransitive the surface structure pattern does not contain an object. The simplest pattern is:

(Subject) (verb)

e.g. 'John slept'

The name 'subject', which is the case of the noun which precedes the verb, indicates that it is the main subject of discourse. It is also the focus of attention.

Although the early notion of case was intended to be purely syntactical, we can see in the notion of focus, and in the distinction between transitive and intransitive verbs, some notion of semantics creeping in. The difference between a transitive and an intransitive verb is that the fmt is associated with some action being done to something, while the second is associated with some action which does not require the participation of another entity. The genitive case (or possessive) also has some hint of semantics about it, although the emphasis is on the form of word to be used, not the role which it is playing. These might be thought of as very 'broad brush' semantic properties.

We will not develop the discussion of traditional surface structure case classifications any further because it has not played a significant role in the development of NL-systems. It should be noted, however, that the identification of the correct surface structure case of a noun has often been used as a stepping stone to the identification of the more semantically based case classification.

11.3 Semantic Case Classifications

In this type of case analysis we are concerned with the roles which various entities are playing with respect to the meaning of a sentence. We are concerned with whether a person is the active 'agent' of an action or the passive recipient (or 'experiencer') of that action. We also note that sometimes an entity is an 'instrument' participating in some action without being the causative agent.

These descriptions 'agent', 'experiencer' and 'instrument' are examples of case classifications. A list of possible cases is given below. For each example the noun shown in capital letters is the one with the case being discussed. Various authorities have suggested several different case categories and an agreed set of case classifications appears to be elusive. The list below includes some from different sources.

Agent. The entity is the animate causative agent which causes the events described. Although it is not explicitly indicated, there is an assumption that the action is deliberate. This assumption can be removed by inserting an adverb such as 'accidentally' or made explicit by means of the adverb 'deliberately'.

e.g. JOHN killed Bill.

Experiencer. The entity is the target of some action, the object to which the action is done. That is, its position changes, or its condition or state of existence is modified in some way as a result of the action taken.

e.g. John killed BILL. BILL died.

Instrument. The entity is the means by which the action is carried out or is the immediate stimulus for the action.

e.g. John killed Bill with a KNIFE. A HAMMER broke the window.

Result. The entity came into existence as a result of the action taken.

e.g. John built a HOUSE.

Dative. The entity is being given something.

e.g. John gave a book to BILL. Comitative. The entity accompanies something.

e.g. John went to the pictures with MARY.

Goal. The entity is the target location of some movement.

e.g. John went to LONDON.

Source. The entity is the location from which some motion took place.

e.g. John came from LONDON.

Other case classifications have been suggested. These include: counter-agent, patient, beneficiary, force, vehicle, material, referent, range, causal-actant, theme, donor, and many others. The multiplicity of these testifies to the fact that no classification scheme for case has been shown to be entirely satisfactory.

11.4 Case-frames

The basic method of determining the case of an entity depends upon the idea that each verb has a predisposition to be associated with particular cases. From this we get the notion of a 'case-frame' - a structure which has slots available into which various entities can be slotted if they have the right syntactical and semantic properties. The idea is analogous to chemical bonds. A molecule has a predisposition to bond with other molecules (or atoms or ions etc) if they are in the correct position and have the correct form (cf. syntactical properties) and if they have the correct chemical properties (cf. semantic properties). While each molecule has unique chemical properties of its own we can identify classes of molecules which have similar bonding characteristics, and the same is true of verbs. We can identify classes of verbs which appear to share a common case­frame structure.

In the description which follows we shall distinguish between the syntactical case and the semantic case by prefIxing these with syn- and sem- respectively.




The case frame above is somewhat cryptic. The explanation is as follows. In any sentence containing the verbs 'kill', 'hit', 'smash' etc. the noun phrase associated with the 'agent' case will be found to be the grammatical subject (in an active voice sentence). If, on the other hand, the sentence has the passive voice (e.g. 'He was killed by John') then the agent case will be found to be associated with the noun phrase preceded by the preposition 'by'.

The 'experiencer' case will be associated with the grammatical object in an active voice sentence, and with the grammatical subject in the passive voice sentence.

The 'instrumental' case is associated with the grammatical subject in an active voice sentence, if that subject is inanimate, for example 'A hammer broke the window'. It is also often introduced by the preposition 'with', (e.g. 'John broke the window with a hammer'). The instrumental case presents particular difficulties and the reader will immediately begin to see possible objections to this analysis. In the sentence 'John was killed by a bus' our analysis would suggest that the bus was the agent, whereas in the sentence 'John was killed by Bill with a bus', the analysis casts the bus in the role of instrument. It is clearly possible for an entity to be both agent and instrument simultaneously. The problem has been discussed extensively in the research literature, and to date no completely satisfactory solution has emerged.

The basic idea is that the case frame will be implemented as a data structure which is associated with each verb classification. Associated with each slot in the frame will be the conditions which should be true of any entity which fIlls the slot. Since these properties often refer to syntactic properties, it is anticipated that the case-frame analysis will follow upon a syntactical analysis, during which each noun phrase will have its syntactical category determined and tacked on to it.

For comparison we provide a second case-frame, this time for verbs which are intransitive and therefore do not normally take a grammatical object.




Here again the reader might raise objections. While one does not normally 'die something' , or 'sleep something' , one does from time to time encounter sentences such as 'He died a quick death' or 'He slept a dreamless sleep'. Both of these constructions suggest a more active role than that of 'experiencer' for 'he'. We can also have such constructions as 'He died of poison', which is equivalent to 'He was killed by poison'. This appears to justify the role of'experiencer' for 'he'. These problems are the result of trying to force a case-frame analysis on a set of verbs as though all members of the set were in this respect identical. In fact all are different in meaning, and therefore possess subtle differences of case dependency which will show up if we seek them out. We shall return to an analysis of the defIciencies of case-grammar in section 11.5.

The prepositions (by, in, on, to, beside, above,...) indicate locative case, but since some of these prepositions also denote other relationships in other circumstances this is not always helpful. One can invent sentences in which the locative case is indicated by the preposition 'in' but in which the same preposition is used for other purposes, e.g.:

'John kicked Bill in a rage, in the kitchen, in our house, in Aberdeen, in December, in a flash, in the stomach, in defIance of my instructions, in full sight of everyone, in my book.'

'He hit the girl with his hand'.

'He hit the girl with red hair.' It is clear from these examples that the prepositions 'in' arid 'with' cannot be accepted as an unambiguous indicator of the locative case or the instrumental case. It also indicates the extent to which we must rely upon examination of the semantic properties of each entity in order to determine its case. Similar examples could be generated for every other preposition.

In most case-frame based systems the case-frame includes slots for:

Tense of the verb (present/past/future/infinitive),

Aspect (perfect/imperfect or continuing! completed),

Mood of the sentence (imperative/declarative/question), Modality (actuaVpossible),

Voice (active/passive).

These factors, which are often carried by auxiliary verbs, are interrelated to the case-slots. For example a transitive verb with passive voice does not need to have an agent but must have an experiencer, (e.g. 'John was killed').

Although, as we have hinted, a case grammar does not provide a good foundation for an NL-system, it has been included here because it represents an important stage in the development of more useful techniques, and because an examination of the reasons why it is unsatisfactory provides a clue to the nature of more promising approaches. It is also the case that the notion of case can be used to augment some other formalism to produce an improved system overall. One of the difficulties often found with an ATN grammar is the problem of dealing with prepositional phrases. The extreme example we gave above for the preposition 'in' shows how much ambiguity is often associated with preposi­tional phrases. A case-frame for the verb can indicate a predisposition for certain prepositional phrases. The verb 'to blame' for example has a potential case-slot for a prepositional phrase beginning with 'on', (e.g. 'He blamed the failure on . Joe').

The diagram below illustrates the way a case-frame can be used along with a simple ATN.

The dotted-line arrows are intended to indicate the creation and update of data structures as a side-effect of traversing certain arcs. The traversal of the verb arc creates the appropriate case-frame. Since this does not exist at the time the first np-arc is traversed, traversal of the first np-arc merely creates a data structure representing the 'subject' case. After the case-frame has been created the traversal of other arcs can update its case-slots directly.

An alternative approach involves the use of 'demons'. These are functions which are placed within each case-slot. When the case-frame is created these are activated one at a time, and each begins to seek out the entity which fits its conditions best. This is a powerful technique which has a number of applications.




11.5 Why a Case Grammar is unsatisfactory

One argument against case grammar can be summarised as follows. A grammar consists of a set of categories within which all words are classified, and a set of grammatical patterns or sentence-patterns which are allowable with respect to these categories. The idea is crucially dependent upon all words within a single grammatical category behaving in exactly the same way with respect to the patterns. It should be possible to replace any word in a pattern by another word in the same grammatical category, without rendering the sentence ungrammat­ical (although the meaning may be and probably will be changed).

The cases of a case grammar are semantic categories. They relate, not to the words involved, but to what the words mean, and the roles played by the entities associated with those meanings within the overall interpretation of a sentence. The idea is based upon the notion that actions can be categorised with respect to these roles, so that we can identify role-patterns just as we identify grammatical patterns, and that we can categorise verbs according to those patterns.

It is arguable that the meanings of words, and particularly of verbs, are much more complex than can be captured by a set of role-patterns, and that therefore we can never define a universally satisfactory set of case-frames which are tied to verbs alone.

Role-playing is intimately concerned with cause and effect relationships which form chains linking events. Verbs are the main way that information about cause and effect chains is communicated, but in some cases some of the information' is carried by, oris implicit in, a noun. Consider, for example, the cause-and-effect information (and therefore the role-slot information) which is communicated by the use of nouns such as 'fertiliser', 'poison' and 'computer program'. When these are used within a sentence they convey information which would otherwise need to be conveyed by the verb. An understanding of the word 'fertiliser' involves knowing that plants grow by taking nutrient from the soil, and that placing certain materials in the soil aids this process. The noun 'fertiliser' may identify the material concerned, but understanding would be incomplete if it did not contain the additional knowledge about the actions and their consequences. An understanding of the verb 'to fertilise' involves exactly the same items of knowledge, even though the verb is the label of the action and not the material. To a large extent, therefore, the two words, noun and verb, can substitute for each other in a sentence in the sense that they indicate the roles which can be played, provided the other words are chosen to make good the elements each lacks. The case-frame idea, however, in its simplest form, is dependent upon the idea that the verb and the verb alone carries or conveys the information about roles.

In the set of verbs provided as examples of case-frame I transitive verbs above, we might have been tempted to include the verb 'to poison' (on purely semantic grounds). It has, after all, a similarity to 'hit', 'hurt' and 'kill'. We can use it in similar sentence patterns:

'John killed Bill'
'John poisoned Bill'
'John killed Bill with arsenic' 'John poisoned Bill with arsenic' 'Arsenic killed Bill'
'Arsenic poisoned Bill'


Consider, however, the sentence (Sl) 'John poisoned Bill's drink'. There is no equivalent for 'kill'. The sentence (S2) 'John killed Bill's drink' may be grammatical but it is semantically anomalous and 'drink' has a different case ('experiencer?'). We can have (S3) 'John put poison in Bill's drink'.

In these sentences we have empirical evidence that 'to poison' is to be classified in a different way from 'to kill'. But our argument goes further.

In the case of the twQ'sentences Sl and S3 most people would agree that they have very similar if not identical meanings. However in S I 'Bill's drink' fIlls the experiencer case-slot, while in the latter Bill's drink fills the location case-slot. The difference is due to the fact that in one sentence much of the cause-and-effect information is carried by the verb 'poisoned' whereas in the other the same information is carried by the noun 'poison'. That is, the verb 'poisoned' conveys the same information as 'put poison in'. We might therefore (taking liberties with the grammar) describe the group of words 'put poison in' as the 'verb' of the second sentence, and this new composite verb would have the same case-slots as the verb 'to poison'.

The meaning of 'to poison' covers both the act of placing a poisonous substance in a suitable medium for ingestion, and the possible consequences which follow. If the syn-object is animate then it is assumed that the consequences have occurred; if the syn-object is inanimate then the verb is taken as describing the action of placing the poisonous substance in some medium identified by the syn-object.

The semantic representation of 'to poison' must therefore have a structure which in simplified form has the following ingredients: The actors in this play:

The action:
(A causes B to enter C)
(D causes C and B to enter D)
(entry of B into D causes D to be sick)
(D dies)


From this 'scenario' we can identify the uninstantiated elements (actors) A,B,C and D. There must, therefore, be possible case-bonds for each of these, and so the case-frame structure emerges as a consequence of the semantic structure. Note, however, that the information contained in this scenario can be communicated both by the verb 'to poison' and the noun 'poison'. For case grammar to be effective in capturing the semantic content of a sentence, it would need to take into account the information content of the nouns involved, not just that of the verbs.

Basically case-frame I above is telling a little story. It says that someone (the agent) did something (the action - no case-slot) to something (the experiencer), possibly by causing something (the instrument) to cause the action to happen to the experiencer. We could add other elements to the story, such as that the action took place somewhere (the location) and so on. The frame with its blank slots tells that basic stereotyped story, and the f1lling in of the various case-slots fleshes it out and particularises it. What that story lacks, however, is any suggestion of detail, and any sense of consequence. It could be argued that we erred in placing the verb 'to kill' and the verb 'to poison' in the same case-category. There are other verbs for which a common case-category is harder to invalidate. Consider, for example, the verbs 'to hurt' and 'to kill'.

'John hurt Bill'
'John killed Bill'


and for (almost) every other sentence structure involving 'hurt' the word 'kill' can be substituted without loss of grammatical sense and without destroying the case-slot pattern. Consider however 'John was hurt until screaming'. The sentence is somewhat contrived, but its meaning is clear enough. The question to be asked concerns the correct case classification for 'screaming' in the case-frame' of 'hurt', and whether there can be any equivalent case-slot associated with 'killed'. Hurting can continue over a period of time 'until' some condition or event; death is an instantaneous event even although it may be preceded by a period of approach to death.

Case grammar is unsatisfactory because it simply does not go far enough in the analysis of the semantic structure of the units it tries to characterise. A case­frame is really a syntactical device trying to accommodate semantic information. Case-frames are too rigid, and cannot accommodate the changes which are necessary to deal with exceptional circumstances. A case grammar identifies the case-slots associated with a verb, but does not make use of the information about roles which can be associated with other words.

A case-frame representation of a 'story' is rather like someone telling us the plot of a film by saying: 'It had a poor fast-shooting good guy, a rich bad guy, a beautiful widow who owned a ranch, two wild Indians, a corrupt sheriff and an amazingly intelligent horse'. No doubt we can fill in the blanks for ourselves without too much difficulty, but as a narrative it leaves something to be desired.

Earlier we made an analogy between case grammars and the idea of chemical bonding between molecules. The classification of chemicals based on ideas of bonding has been extremely successful. That classification scheme, however, is based upon a well-established theory of molecular structure. The bonding characteristics of a chemical are a consequence of its structure. A case grammar is similar to an empirical law of bonding which is not illuminated by having an underlying theory. As a consequence we may be tempted to classify in the same way two examples of 'case' which are actually different although they may have superficial similarities. Likewise we may be tempted to separate two verbs because they have some different characteristics with respect to their case-frame although they have an underlying similarity. Later we shall develop an argument in favour of the scenario approach to semantic representation. leis of interest to note here, however, that the scenario approach can explain the existence of case-slots.

In spite of the limitations of case grammars, they have been a spawning ground for a number of very interesting NL-systems.