29 Stronger knowledge representation methods: Conceptual Dependency

Bhushan Trivedi

Introduction

In this module we will look another method for knowledge representation, Conceptual Dependency. Though it is quite old method, it is still useful in many ways. Also, CD is an ideal example of how one can have distinguished types of links and nodes and give them specific meaning and provide a far better way to represent a natural language statement than simpler methods like semantic nets. CD is designed in a way that, when a statement is presented, the meaning of the statement is derived from the statement and the meaning and not the words of the statement derives the representation. The derived representation is used for inference subsequently. As the inference is derived from meaning and not from the bare words, it is more human like. Links with specific type (single line and double line, arrows on one or both sides etc.) and specific names are used for typical purpose in CD. A link also has restriction on what endpoint types it can have. Not all objects can have all types of links. The restriction on types of links and endpoints enable to represent knowledge in much stronger way than the methods we described in the previous modules. The hard and fast rules about the structure, links and objects help the stronger knowledge representation which is the theme of this module.

Conceptual Dependency

CD was developed by Roger Schank and his teammates from Yale University.Conceptual Dependency or CD is a way of representing knowledge in a stricter form than semantic net. In semantic net we have no strict laws for naming links and endpoints while CD has laws for both of them. In fact CD is designed especially for representing natural language statements so it is designed keeping that purpose in mind. The idea of CD can be characterized by following points.

•  CD should act as an Interlingua; a language which represent meaning of statement in a form which can be used to convert it to a different language easily. The idea of Interlingua eliminates the need of conversion routings from each possible language to every other possible language. For each language, we need to have two routines, first to convert from that language to an Interlingua and second from Interlingua to that language.

•  The knowledge representation provided by CD is generic and contain information not directly provided in the statement itself. This it helps the system reason with and answer based on knowledge which are otherwise not explicitly mentioned in the statement.

• The CD, as being an Interlingua, is independent of any specific language. In this module we will see some examples of converting English statements to CD, but once the conversion is done to CD, the representation is free from English. That means, we can use CD to accept information in any language, and produce response to query in any other language.

The idea of having a language independent structure drove the designers of CD (Shank and others in 1975) to work on meaning and not the words themselves. The designers worked on designing a construct which represent the meaning of the words used in the statements in a conceptual form. The design of CD models the concepts the statement is trying to convey. Thus the structure is not only independent of the language but also independent of the words used in the statement. The semantic net was a general idea that represents a statement using the words used in the statement. Semantic net connects the words by some relations apparent from the statement. Thus semantic net describes the statement itself. Unlike that, the CD represents the meaning of the statement and not the statement itself. Different statements with same meaning will be converted to same representation. The CD provides specific primitives to act as building blocks. Whenever a statement is encountered, it is broken into those specific primitives and specific values are assigned to the links and nodes accordingly. Thus similar statements has similar structures with different values. Schank has provided some primitive acts to describe each statement. They are shown in figure 29.1 to 29.4.

The idea of conceptual primitives is derived from the requirement of gathering meaning from the statement. Conceptual primitives can be considered to be basic meaning structures that cause the words that are part of the statement. Though these conceptual primitives sound simple, they can be combined to produce structures to represent complex meaning. The first step is to get the idea of the statement. The goal of CD is to represent that idea in form of basic primitives. Second step is to present the idea using CD primitives. Once the statement’s idea is represented by those primitives, we can infer from that further1. Thus small sets of conceptual primitives (according to Schank) can represent large number of statements. The conceptual primitive actions are listed in table 29.1. Any physical or mental action is conceptualized using these small set of primitives. These are types of links we were discussing earlier. The table 29.2 describes types of endpoints which can be used to describe a statement in CD form. Which object can do what, i.e. the role of an object is also specified in CD. Thus endpoints and links together describe the complete meaning of a given statement.

When we write statements and convert to their predicate logic form, we call them propositions; they describe events. For example Anand beaten Gary is an event and represented by beaten (Anand, Gary) . In CD they are called conceptualization and represented as a combination of an Actor, an Action and set of cases that are dependent on that Action.

The idea of CD is to provide canonical form for representing meaning of the sentence. The conceptual roles do not correspond to syntactical roles. A word is broken down into smaller primitives (Conceptual Primitives) and placed in the structure based on their conceptual role rather than syntactical role. For example word ‘love’ has a nominal form, as well as a verbal form ‘love’, adjectival form ‘lovely’ and adverbial form ‘lovingly’. The meaning of love does not change when a person chooses one or the other lexical form of the word ‘love’ to represent his idea.

While describing love in CD, a typical type of structure is used. One meaning of love is something which improves somebody’s mental state, has another object associated with it which is the reason for that improved mental state. One can describe the statement containing love is that somebody’s mental state is improved.

There is a possibility of the complete meaning cannot be derived from the statement like “Ram is in love” as the object is unknown. Still, it is perfectly alright to construct a structure that conceptually describe mental state of Ram improved as a reason of some object2.

1 The idea is quite similar to system calls of UNIX. They are about 60 but can represent any action that require kernel access.

2 My representation of word love may look quite crude but however hard you try, it is impossible to describe love into words. This is not the problem of CD, this is the problem of the language itself. This problem might have occurred to you before during the predicate logic description. Brother(X, Y): – Parent (X, A, B) and Parent (Y, A, B) crudely describes a relation called Brother when the objects involved have same parents. Can one describe a relation in words like this? Many agree that it is not. Being brother is much more than having same

If we encounter a statement “Ram is in love with Sita” or “Ram loves Sita” or “Ram found Sita loving” we can see that in all cases the actor is Ram and object is Sita while the action involved is ‘love’. Though there are some differences in the meaning, for a common man, the statements are quite identical and CD representation of them, which is considered to be a strength of it, is also the same.

Another thing the CD attempts to address is to get more information about the statement than it is apparent from the statement. For example “Purochan endangered Pandavas” might be one of the statements encountered by us. That means that Purochan is likely to do something which cause Pandavas some harm. In fact one may go further (looking at the context) and determine that Purochan is likely to set fire to the LakshaGruha where Pandavas are residing. A CD representation in that case, might be like, Purochanis likely to set fire to LakshaGruha which in turn is likely to reduce Pandavas’ health to -103.

This setting the value of variable health to -10 is a typical way to indicate that the person involved is dead.

Before we proceed further, let us quickly look at primitive actions used in CD.

Primitives actions for CD

We have three tables which summarises the types of links and endpoints used in CD. Table 29.1 describes eleven basic primitives. Let us take each one of the primitives and understand what they are trying to convey.

ATRANS: – this action describes transfer of an abstract relationship, such as control, ownership or position of something. The actor is instrumental in initialling the process, the object’s ownership is being transferred while one of the object is losing the ownership and the other one is earning the ownership. For example if we encounter verb ‘give’, it means the actor ATRANS an object to the recipient. Instead, if we encounter word ‘take’, the actor ATRANS an object from some other animate object to himself. If we find a verb ‘buy’ instead, we have two ATRANS, one moves the position of money to seller from buyer and another moves the ownership of the object being sold from the seller to buyer.

PTRANS: – while ATRANS describes abstract transfer, the PTRANS describes physical transfer of an object. For example the verb ‘throw’ describes an actor PTRANS an object from one location to another and ‘catch’ describes actor is a recipient of somebody else PTRANS an object 4. PTRANS can also be described for movement. For example if we want to describe a verb ‘run’, it is about actor transferring himself to other location. In this way, the meaning of ‘run’ and ‘walk’ is represented in similar fashion in CD. Sometimes the additional action called instrument is provided to indicate the finer meaning. For example if we want to represent ‘fly’, the actor PTRANS himself to another location but the instrument is the journey through plane. When we find ‘drive’ instead, the same representation has an instrument is the journey through a car.

PROPEL: – propel describes physical force is being applied to an object. Many a times PROPEL also results into changing the location of the object, in that case PTRANS also is applied. For example when we encounter verb Push it means PROPEL in direction of the actor, while Pull means PROPEL in parents. However CD being derived from the language, it is impossible for it to improve the linguistic representation of the idea.

3  How can CD learn about context, what is the indication of health = -10 will soon be clear. 44Throw also involves moving a body part and physical force and thus MOVE and PROPEL also are involved. opposite direction. Similarly throw or kick involves PROPEL the object in a typical direction using an instrument using another primitive MOVE (moving a body part).

MTRANS: – this is about transfer of mental information between animate objects or within an animate object. The conventional discussion related to memory contain three parts according to Schank’s model. CP or conscious processor is the first component. When something is going to CP, it is being thought. The IM or immediate memory is the place where the relevant information is kept while being processed. The LIM or long term memory is where the complete information is stored. For MTRANS, the information comes from various senses and the recipients are any one of the three elements mentioned above. The object is the mental information being transmitted.

When we encounter verb ‘remember’ for example, we MTRANS information from LTM to IM. When we encounter ‘forget’ it is negation of the same. In the verb ‘hear’ the actor MTRANS information from sensory organ ear to the CP and ‘tell’ an actor MTRANS information from IM to another actor by means of SPEAK. Similarly read can be represented as MTRANS information from the newspaper to IM by means of ATTEND eye to the newspaper.

MBUILD: – MBUILD is about constructing information using some mental process. The inference that we have discussed so far is a process which generates new information from an old set of information. This is represented by primitive MBUILD. This process takes place in IM and the newly generated idea is placed in CP. For example when we encounter word ‘imagine’ is MBUILD a new idea, word ‘consider’ is MBUILD from current information, word ‘answer’ is MBUILD a response to some input from some other actor received recently in CP (and then MTRANS it to that actor).

INGEST: – Taking in an object by animal. The actor INGEST object into the inner part of the actor. When we encounter verb ‘eat’ it is about INGEST solid into body through mouth, ‘drink’ is similar for liquid, breath is INGEST of air into body by nose, ‘smoke’ is INGEST smoke from cigarette into the mouth using the instrument as mouth INGEST air through cigarette.

GRASP: – taking an object into the actor’s hand; grasping it. Verbs like ‘pickup’ is GRASP that object.

‘hold’ is another verb which is same as GRASP that object.

ATTEND: – this is about focusing a sense organ to an object. Verbs like ’listen’ , ‘hear’, ‘see’, ‘experience’ etc uses MTRANS as major primitive but it also use ATTEND as instrument.

EXPEL: – this primitive is about pushing something out of the body. The verb ‘cry’ is about EXPEL tears from the eyes5.

SPEAK: – this primitive is about producing voice. In fact sing, talk, shout all of them are crudely represented as SPEAK.

Table 29.1 Conceptual primitives for actions and their meaning

	Primitives	meaning
1	ATRANS	Initiate a change in abstract relationship of a physical object
2	ATTEND	Focus or direct a sense organ towards a stimulus.
3	INGEST	Take something inside an animate object
4	EXPEL	Something is coming out of animate object forcefully
5	GRASP	grasp an object

55Again, you might indicate that crying is not about just tears coming out of eyes and much beyond, but that, as discussed before, limitations of the language.

6	MBUILD	Contrive or combine thoughts
7	MTRANS	Mental information transfer
8	MOVE	Move a body part
9	PROPEL	Apply a force to
10	PTRANS	Change the location of a physical object.
11	SPEAK	Produce a sound6

Conceptual categories

The action primitives describe action as links. We will soon see how these links are depicted using examples. The endpoints of those links describe another set of building blocks for CDs. The links are also called dependencies. The endpoints of dependencies indicate who depends on what, the entities. The link also indicate the type of dependency. The endpoints are of any conceptual categories described in table 29.2. Actors of the actions are known as PP. PP or Picture Producer are of two different types; animate objects and natural forces. Action is described by ACT. ACT is conceptual category containing all 11 primitives (12 if you count PLAN). LOC indicates location of an object while T indicates time. Action Aiders (Or action modifiers) provides modification to objects by given ACT (for example run modifies the location, poison modifies the health state, grow modifies the size etc.) PA or picture aiders are attributes of the object described in the form of state (value) form.

Table 29.2 Conceptual categories

Categories	Meaning
PP (picture producer) :	Physical object.
	Actors are either animate pp, or a natural force
ACT :	One of eleven primitive actions.
LOC :	Location
T :	Time
AA (action aider) :	Modifications of features of an ACT
PA :	Object’s attributes.Are described asstate (value). E.g., voice
	(melodious).

Conceptual Roles and Tenses

Table 29.3 and 29.4 describes two more important information. The single conceptual entity representing an event is known as a conceptualization. A natural language statement sometimes contain multiple conceptualizations. Actor is acting on Object using ACT. ACT is performed on object by actor. Recipient receives the outcome of the process. Direction suggest the direction of the process. For example if one is giving a book to another, the Direction indicates the direction in which the possession is transferred. Taking a book is described by the same conceptualization except the direction. State indicates some state of an object. For example when we encounter verb like ‘grow’, it is indicated by state of the object changing to a bigger size.

Additionally conceptual tenses can be represented using a single letter indicator as depicted in table 29.4. Normal tenses like present, past and future (with continuous) are included with negation,

6   There was a 12th primitive introduced by Schank later called PLAN. The idea of PLAN was to provide the difference between MBUILD which conceives something general and PLAN which conceives something which itself is a conceptualization. Unlike MBUILD, PLAN must have another conceptualization which indicates what exactly the actor is planning to do, go somewhere (PTRANS), eat something (INGEST), whatever. PLAN was used in specific circumstances where plan goal and actions are related to each other. We will discuss about that in brief later in module 31. interrogation, for additional information. Conditional tense indicates if-this than that kind of a relation. Timeless tense is also available to indicate information which has nothing to do with time. Transitions commencement and conclusion are also indicated by specific tenses decided for the purpose.

Table 29.3 Conceptual Roles

Roles	meaning
Conceptualization:	Representation of a single concept. It is a basic unit of the conceptual level of understanding. Multiple such units are combined to represent a natural statement.
Actor	Whoever is performing the ACT is an actor
ACT:	An action (performed on the object)
Object	A thing that is acted upon
Recipient	Whoever receives the object as an outcome of the ACT
Direction	The direction or location towards the ACT is heading to
State	State of an object

Conceptual Tenses
past	p
future	f
negation	/
start of a transition	ts
end of a transition	tf
conditional	c
continuous	k
interrogative	?
timeless	∞
present	nil

Syntactical Rules

CD, as mentioned before, is quite stringent in its expression and there are quite a few rules for it to work. The links describe dependencies. As the representation is based on conceptual and not verbatim representation of a statement, these dependencies are between concepts and thus this structure is known as CD.

Let us take an example to illustrate the point. Suppose we want to represent a statement, I watched a match.

It is represented as follows with meaning described next.

•  Arrows indicate direction of dependency

•  Double arrows indicate two way dependency between actor and action

•  The ‘p’ indicates past tense

•  The ‘o’ indicates object

•  MTRANS is one of the primitive ACTS, it indicates transfer of mental information

•  R indicates recipient case relationship, eyes are initiator while CP is the receiver. This process indicates that the mental information is coming from eyes and delivered to CP

•  Both structure and primitive act are provided by the CD unlike semantic net. The user just need to provide information for the placeholder.

•  The construct is designed based on one or more rules described later. For every rule, there is a typical construct which must be used. Apart from primitive acts, the information can also be conveyed in form of state change. For example, when we encounter following statement Ravan is dead, It is indicated as follows. The health (like many other possible states) is described in quantitative terms of -10 to 10. When it is assigned value -10, the actor is dead.

Also note the state (value) format to indicate the value to a state. Also note the three prong structure to indicate that Ravan is affected by the change of state from other than -10 (not dead i.e. alive) to dead (making it equal to -10). Another important issue is to represent multiple statements with similar meaning using same structure. For example following statements will have same CD representation

Sachin gave a bat to Sarfaraz

Sarfaraz was given a bat by Sachin.

Sachin gifted a bat to Sarfaraz. …….

You can see that the primitive act here is ATRANS which indicate transfer of ownership in this case from Sachin to Sarfaraz, the object is bat and the recipient case indicates that the object is changing its ownership from Sachin to Sarfaraz.

In addition, a dependency structure itself can act as a conceptualization and can serve as a component of another larger structure. Thus it is possible to represent complex statements as collection of simpler statements and thus a complex CD representation comprises of smaller CD representations of smaller components of that statement.

We will look at number of syntax rules for CD in the next module. We will also see an example for each rule and also look at construction of some complex statements.

Summary

CD or conceptual dependency is a method of representing statements in a language and word independent form. The CD provide primitive acts which represents the English (or any other language) verbs in simpler structure. There are total 11 primitive acts and whenever an English verb is observed, it is converted to a form containing one or more primitive acts. The representation of a single idea is called conceptualization. The conceptualization connects actors, recipients, objects and instruments by different dependencies. Those dependencies or links are specifically designed to indicate the purpose. Single or double lines indicate strength of the relationship, arrows and their directions indicate dependencies while indicators over the link provides further information.

Indicator include conceptual tenses and other names. For constructing a CD representation out of a given statement, some syntax rules are designed which we will discuss in the next module.