CHAPTER 14

Conceptual Dependencies

14.1 The Theory of Conceptual Dependencies

The theory of conceptual dependencies has been influential, and has been used as the foundation for a number of interesting systems which have achieved a limited measure of success. Although it has an ad hoc flavour which is philosophically unsatisfactory, it addresses important questions of meaning and interpretation which have largely been avoided by many other NL systems. The theory is based upon the idea that the meaning of statements can be analysed in terms of three basic types of mental construct. These are acts, picture producers (PP), and states. These three constructs are the building bricks of the theory. They are combined to produce structures known as 'conceptualisations' and the relationships between the various components are called 'conceptual dependencies'. We shall describe each component in turn and then consider the theory as a whole.

14.2 The Eleven Primitive Acts

The theory of conceptual dependencies suggests that there are eleven 'primitive acts'. The description 'primitive' indicates that they do not require analysis into more fundamental actions, and they do not require explanation. These acts consist mainly of the moving about of ideas and physical objects. The actions described by, or associated with, the use of particular verbs are then analysed in terms of these primitive acts, and in most cases they can be shown to be equivalent to one or more of the primitive acts, perhaps with the value of one or more of the case-slots predefined. If the primitive acts are thought of as the 'atoms' of meaning, then the actions associated with the verbs can be thought of as the 'molecules' of the system. The case-slots or 'dependencies' which are associated with each verb emerge as a property conferred upon them by their 'molecular structure'.

The primitive acts each have three basic case-slots - the agent, the object and the direction. The direction is subdivided into 'from' and 'to' case-slots. In some cases both do not apply.

Every primitive act can also be extended by adding an instrumental case which is often another primitive act. Thus a person might INGEST a glass of milk, where the instrument is the primitive act of physically transferring (or PTRANS­-ing) the glass to his lips. In some instances we have used expressions such as 'mindof(X)'. This is intended to indicate the conscious mind of a person or animal. The nature and number of primitive acts described by this theory have varied over a period of years as the theory was refined, tested and modified. A recent count places the number at eleven.




The eleven primitive acts are:

(1) MOVE

This is the movement by a person or animal of a part of its own anatomy. That is, it moves its hand, or some other part of itself, to a new position. The case-slots associated with this act are:

MOVE:
    the  agent:  A
    the  object  moved  :  partof(A)
    the  direction
        where  from:  B 
        where  to:      C



(2)  PROPEL



This is the act of causing something (other than oneself) to move to a new position. The case-slots are:

PROPEL:
    the  agent  :  A
    the  object  propelled  :  B
    the  direction
        where  from  :  C
        where  to  :  D
    the  instrument  :  a  MOVE  at  /  another  propel  act  /  a  grasp  act



(3)  INGEST


This is the act associated with the assimilation of food or drink. The case-slots are:

INGEST:
    the  agent  :  A
    the  object  ingested    :  B
    the  direction 
        where  from  :  C
        where  to  :  inside(A)/  mouthof(A)
    the  instrument:  a  PTRANS  act



(4)  EXPEL



This is the act associated with the elimination of material from the body (e.g. blood, faeces, urine, air, etc.). The case-slots are: EXPEL:
    the  agent  :  A
    the  material  expelled  the  direction  :  B
        where  from  :  inside(A)
        where  to  :  C
    the  instrument  :  a  PROPEL  act  /  a  MOVE  act



(5)  GRASP
This is the act associated with the grasping of an object in the hand, or in some other way. The case-slots are: GRASP:
    the  agent  :  A
    the  object  grasped  :  B
    the  direction 
        where  from  :  C
        where  to  :  handof(A)
    the  instrument  :  a  MOVE  act



(6)  ATRANS
This is the act associated with an 'abstract' transfer from one state to another. A typical example is provided by the sale of goods from one person to another. In some cases the object itself does not move at all (e.g. a house) but the object has been 'transferred' from the ownership of one person to the ownership of another. The case-slots are:

ATRANS
    the  agent :  A
    the  object  transferred:  B 
    the  direction
        where  from  :  C
        where  to  :  D
    the  instrument  :  a  PTRANS  act/  an  MTRANS  act  /  a  MOVE  act



(7)  PTRANS



This is the act associated with the physical transfer of an object (including the movement of itself) from one place to another. The case-slots are:

PTRANS:
    the  agent  :  A
    the  object  moved  :  B
    the  direction
        where  from  :  C
        where  to  :  D
    the  instrument  :  a  MOVE  act  /  a  PROPEL  act



(8)  SPEAK



The  nature  of  this  act  is  obvious  from  its  name.  The  case-slots  are:



SPEAK:
    the  agent  :  A
    the  object  (thing  spoken)  :  B
    the  direction 
        where  from  :  A 
        where  to  :  C
        the  instrument  :  a  MOVE  act  (one's  mouth)



(9)  ATTEND



This act is the converse of SPEAK. It is associated with the reception of some communication. The case slots are:

ATTEND:
    the  agent  :  A
    the  object  (sense  organ)  :  B
    the  direction
        where  from  :  C
        where  to  :  A



(10)  MTRANS



This is the act associated with the transfer of ideas or information from one person to another. The case-slots are:

MTRANS:
    the  agent  :  A
    the  object  (information):  B
    the  direction
        where  from  (medium)  :  C 
        where  to  :  mindof(D)
    the  instrument  :  a  SPEAK  act  /  an  ATTEND  act



(11)  MBUILD



This act is associated with the mental construction or working out of ideas from simpler ideas leading to decisions. The case-slots are:

MBUILD 
    the  agent  :  A
    the  object 
        input  :  B
        output  :  C
    the  direction 
        where  from  :  memory(A) 
        where  to  :  mindof(A)
    the  instrument  :  an  MTRANS  act



That completes the list of primitive acts.

14.3 Picture producers (PP)

The strange name 'picture producers' arises from the notion that many mental constructs trigger a pictorial visualisation in our minds. A picture producer (or PP) corresponds both to the generic concept structure which we have described in previous sections and to specific examples of these concepts. Many nouns are associated with PPs, but not all nouns. A concept such as 'honesty' is not a PP. (It is not clear how a concept such as 'honesty' is represented in the theory of conceptual dependencies.) Being a concrete entity, a PP can be the agent of an act. A PP can be implemented as a record structure or frame with slots for its properties. These properties may be states. The idea is very similar if not identical to the object-frames described in section 12.2.

14.4 States

A state is a condition or property of the world. Something being in a particular location is a state. Something having a particular colour is a state. Someone being in a particular emotional condition is a state.

A state can be represented by a record structure with two elements: the name of the state (or 'state-id') and the value (or 'state-val'). The terminology is that proposed by those who developed the theory of conceptual dependencies (CD), but a 'state' in CD terminology appears to be identical to a 'property' which has a 'property name' and a 'property value', rather in the same way as we suggested for the dictionary structure in section 8.5.

The record which represents a state can itself be the value of one of the properties of a PP or one of the case-slot values of an act. For example, to indicate that a PP has a particular location we might give it a property:

state-id  =  LOC  (for  location), 
and  state-val  =  London.



This data structure could be assigned as the slot fIller for the 'where to' case-slot of a primitive act.

The theory identifies states corresponding to (or having the property names) length, colour, light intensity, mass and speed. It also identifies states of control, part(of), ownership, containment and proximity.

To deal with emotional states of mind, CD theory has introduced the concept of 'mental location' (MLOC) and defines an arbitrary numeric scale of values ranging from -10 to 10 (catatonic to ecstatic). The act of pleasing someone is therefore represented as an act which causes (DO) an increase of the numeric value of that person's mental state (A TRANS from MLOC -1 to MLOC +1 (say).

Other mental states which also have their numeric scales of values are 'anger', 'fear', 'disgust' and 'surprise'. The theory also includes states such as 'health', 'hunger', consciousness' and several others.

The arbitrary nature of these states and their numeric scales of values is one of the most unsatisfactory aspects of the theory, but at least it represents an attempt to tackle these problems in a fairly direct way.

14.5 Implementation: from Input Text to Conceptualisation

Various mechanisms for getting from the original input text to the internal representation (or conceptualisation) have been tried. Some have used special parsers and others have abandoned the process of parsing altogether. We shall describe one possible method in outline only.

The basic data structures are:

(1)  INLIST:  the  input  text.
(2)  CST  ACK:  a  stack  of  conceptualisations.  It  is  initially  empty.
(3)  RSTACK:  a  stack  of  'requests'.  The  RSTACK  is  also  initially  empty.
(4)  LEXICON:  a  list  of  all  the  words  which  the  system  can  handle. 
                          Each  word  is  associated  with  a  PACKET.
(5)  PACKETS:  a  set  of  requests.
(6)  REQUEST:  a  demon,  or  function  (see  section  12.3)  which  can  be  stored 
                          and  fired  when  required.  A  request  has  the  form:­



        TEST:  a  condition  which  must  be  true  for  the  request  to  be  granted; 
        ASSIGN:  a  set  of  variables  and  values.  If  the  request  is  granted 
                        the  values  are  assigned  to  these  variables;
        NEXT-PACKET:  a  list  offurther  requests,  which  are  loaded 
                                  if  the  first  ­request  is  granted  (and  after  the 
                                  assignment  of  the  variable  values).



The  text-processing  algorithm  is  as  follows:



while  inlist  is  not  empty
        Remove  the  first  word  of  inlist  and  assign  it  to  x.
        Look  up  x  in  the  lexicon,  extract  its  packet  and 
                  place  the  packet  on  RSTACK,  while  (RSTACK  is  not  empty) 
                  and
                (there  is  a  request  there  which  can  be  fired)
        Look  for  a  request  with  test-condition  =  true;
    if  found  then
                      assign  the  variable  values
                      and  carry  out  any  other  actions  requested 
                      remove  the  packet  from  RSTACK
                      add  its  next-packets  to  RSTACK  endwhile;
endwhile;



Note that in carrying out the actions indicated by the request, the system may place a conceptual dependency structure on the CSTACK. It can also access and modify the values assigned to the variables within conceptual dependencies already on the CSTACK. The CSTACK therefore provides a mechanism for communication between processing handling different parts of a statement ­the same facilities provided by the side-effects of our ATN processing technique. When processing is completed, the contents of the CST ACK will represent the meaning of the statement.

The real work of the process (analogous to the side-effects of our ATN processing) is therefore carried out by the requests, which vary from word to word in the lexicon (just as the side-effects varied from arc to arc of the ATN).

An example of the packet associated with the word 'went' (adapted from Inside Computer Understanding by Schank and Riesbeck) is shown below. Go­varl, go-var2 and go-var3 are local variables (within PTRANS). Part-of-speech, subject and conceptual-dep are global variables which are updated as processing is carried out. The variable conceptual-dep is always set equal to the most recent conceptual dependency analysed.

PACKET  for  "WENT"
------------------(start  of  "went")
TEST:  --­
ASSIGN:
part-of-speech  :=  verb
conceptual-dep  :=  PTRANS
                  agent  :=  go-varl
                  object:=  go-varl
                  to  :=  go-var2
                  from  :=  go-var3
            go-varl  :=  subject
            go-var2  :=  nil
            go-var3  :=  nil
            NEXT-PACKET
-----------------(start  of  "to")
                    TEST  (next-word  =  "to")
                    ASSIGN:  ---­
                    NEXT-PACKET
                    ------------------(start  of  "np")
                          TEST:  (part-of-speech  =  np)  ASSIGN:  ­
                          go-var2  :=  conceptual-dep
                          TEST:  (next-word  =  "home")
                          ASSIGN:  go-var2  :=  "house".
                    ------------------(  end  of  "np")
            ------------------(end  of  "to")
--------------------(  end  of  "went")



Note  that  this  could  be  translated  into  the  program:



define  went;
    part-of-speech  :=  "verb";
    v1  :=  subject;
    v2  :=  nil;
    v3  :=  nil;
    conceptual-dep  :=  PTRANS(vl,vl,v2,v3);  *
    if  next-word(inlist)  =  "to"  then
            *
if  part-of-speech  =  np  then
                  v2  :=  conceptual-dep;  **
endif;
    else
if  next-word(inlist)  =  "home"
                  v2  :=  "house";
endif;
    endif;
enddefine;



At the points marked ,*, the program is entering a new 'request'. At the point marked ,**, the value of the conceptual-dependency which is assigned to v2 is the conceptual dependency belonging to the noun phrase (np).

The way this system behaves is complex and difficult to understand because of the number of variables, and the way these are updated at different times by procedures which are themselves dependent upon the other words in the sentence being processed. It is important to note, however, that the idea of a parser has almost vanished from the system. Syntactical information is embedded in the packet associated with each word. Each 'knows' what words and syntactical structures to expect in its immediate vicinity, and provides the system with this local information as it is required. Note that the series of packets given in the illustration are testing for two possible sequences of words:

went  to  (somewhere) 
went  home


The tests involved in the 'requests' correspond logically to the tests carried out in an ATN before deciding which arc should be followed. The updates of the variables correspond to the side-effects generated when an arc is traced.

14.6 The 'DO' Act

To the eleven primitive acts is added a twelfth general purpose act called DO. It simply means 'to cause something to happen' and is often used in conjunction with the primitive acts by making one of these fill its object case-slot. That is -

DO 
    the  agent  :  A
    the  object  :  another  act  or  a  state
    the  instrument  :  a  MOVE  act,  a  PTRANS,  etc.


14.7 Tense in Conceptual Dependencies

To simplify things we have omitted any mention of tense. The conceptualisation structure can be extended to include additional case-slots, one of which would be a tense parameter. The requests associated with the package for any given verb can include a request to insert an appropriate tense value.

14.8 Examples of Conceptual Dependencies

In section 14.5 we provided an example of the 'package' which is used for the verb 'went'. This package creates and places in the CSTACK a conceptualisation with the structure PTRANS(agent,object,from,to), so that the representation of the sentence 'John went to London' would be:

PTRANS 
    agent=John 
    object=John 
    from=  ? 
    to=London



We also stated in section 14.1 that the basic idea in CD theory is to interpret the meaning of verbs (and other words) in terms ofthe primitive objects (PPs), acts and states. To illustrate this we provide below some examples of sentences and their interpretation in these terms. Most are adapted from examples given by Schank and Riesbeck.

(a) John took a plane to London

PTRANS agent = John object = John from = ? to = London instrument = PROPEL agent = plane object = plane from = ? to = London and PTRANS agent = John object = John from = ? to = plane

(i.e. John went to London by placing himself on a plane which propelled itself to London)



(b) John drank a glass of milk

INGEST agent = John object = milk from = glass to = mouth of John instrument = PTRANS agent = John object = glass from = table to = mouth of John instrument = MOVE agent = John object = hand of John from = ? to = glass and GRASP agent = John object = glass from = ? to = hand of John

(ie. John drank a glass of milk by moving the glass to his mouth by moving his hand to the glass and grasping the glass)

The reader will recognise the alarming complexity of this structure, and the fact that some additional instrumental or intermediate acts have been omitted. For example, after grasping the glass and moving his hand to his mouth (thus transporting the glass to his mouth) John tipped the glass so that the milk flowed into his mouth, and then he swallowed...etc.

(c)  John  decided  to  go  to  London



MBUILD
    agent  =  John
    object
    input  =  ?
    output  =  PTRANS
                          agent  =  John 
                          object  =  John 
                          from  =  ?
                          to  =  London
    from  =  memory
    to  =  conscious  mind






(d)  John  told  Mary  the  story
MTRANS
    agent  =  John
    object  =  story
    from  =  John
    to  =  Mary
    instrument  =
        SPEAK
            agent  =  John
            object  =  story
            from  =  mind  of  John
            to  =?
        and
        ATTEND
            agent  =  Mary  object  =  ears
            from  =  John
            to  =  mind  of  Mary



14.9  Some  Thoughts  on  CD  Theory



It is easy to be critical of CD theory. The description provided here is necessarily brief and inadequate. It is intended only to give a flavour of the system, and does not do justice to the very real success which the approach has achieved in producing, for example, good paraphrases of the input text. It is not recommended, however, that the reader try to reproduce a working system for CD theory.

The criticisms which have been levelled at CD theory tend to be generated by a distaste for the clumsy representations it produces and the almost complete absence of mathematical elegance. Two things need to be said in its defence, however. Firstly it represents a frontal attack on the real problem of NL processing. It asks the right questions of itself. It tries to produce a representation of the actual meaning of a statement. It recognises the importance of motivation (or 'mental states'). It recognises that events are linked causally and can be chained together to produce an overall result. Secondly by demonstrating its own limitations it has drawn attention to the need for contextual information and knowledge about the knowledge of others. When the system was put through its paces and was asked questions about the sentences it was processing, it tended to give answers which were much more detailed than a human would give. For example, if such a system processed the statement 'John fell in the water and was drowned', and was then asked why John died, it tended to give a string of answers such as:

because  he  could  not  breath
because  his  air  supply  was  blocked
because  his  mouth  was  under  water  because  he  was  immersed  in  water
and
because  he  could  not  float
and
because  the  water  was  deeper  than  his  height  because  ...


In other words it would print a very detailed list of cause and effect in terms of primitive acts. A human, on the other hand, would know that the list was of little interest to another human, because it was common knowledge. The interesting information is that which cannot be assumed to be common knowledge - that John fell in the water. The rest can be taken for granted. This clearly points up the need for an NL processing system to have some idea of what others (as well as itself) know about. It is suggested that we should ignore the clumsy notation and terminology and the crude numerical scale representing mental states and the like. Whatever method we use to represent the meaning of a statement, it will require a notation or symbolism of some kind. What form it takes is of less importance than the things it attempts to represent. The important aspect of CD theory is the fact that it tackles the representation of the important factors, and correctly relegates syntactical analysis to the role of an aid in the achievement of understanding. The function of language is, after all, to transmit understanding from one person to another. The grammatical elegance of our words (or lack of it) is of secondary importance.