A Computational Psycholinguistic Model of Natural

Language Understanding

Jerry T. Ball

Air Force Research Laboratory, 6030 S. Kent Street, Mesa, AZ 85212-6061

Abstract. Double R Model (Referential and Relational Model) is a computational

psycholinguistic model of natural language understanding founded on the linguistic

principles of Cognitive Linguistics and implemented using the ACT-R cognitive

architecture and modeling environment

1 Double R Grammar

Double R Grammar [1] is the Cognitive Linguistic theory [2,3] underlying Double R

Model. In Cognitive Linguistics, all grammatical elements have a semantic basis,

including parts of speech, grammatical markers, phrases and clauses. Our

understanding of language is embodied and based on experience in the world [4].

Categorization is a key element of linguistic knowledge, and categories are seldom

absolute—exhibiting, instead, effects of prototypicality, family resemblance [5], base

level categories [6], fuzzy boundaries, and radial structure [7]. Our linguistic abilities

derive from basic cognitive abilities—there is no autonomous syntactic component

separate from the rest of cognition. Knowledge of language is for the most part

learned and not innate. Abstract linguistic categories (e.g. noun, verb, nominal,

clause) are learned on the basis of experience with multiple instances of words and

expressions which are members of these categories, with the categories being

abstracted and generalized from experience. Also learned are schemas which abstract

away from the relationships between linguistic categories. Over the course of a

lifetime, humans acquire a large stock of schemas at multiple levels of abstraction and

generalization, representing knowledge of language and supporting language

understanding. These schemas constitute what might be called grammatical

semantics [8] in contrast to the lexical semantics of individual lexical items,

although the schemas are, for the most part, associated with specific lexical items.

Two key dimensions of meaning that get grammatically encoded are referential

meaning and relational meaning. Double R Grammar is focused on the representation

and integration of these two dimensions of meaning within the wider scope of

Cognitive Linguistics. Consider the expressions

1. The book on the table

2. The book is on the table

T. Ball J. (2004).

A Computational Psycholinguistic Model of Natural Language Understanding.

In Proceedings of the 1st International Workshop on Natural Language Understanding and Cognitive Science, pages 3-14

DOI: 10.5220/0002667400030014

 SciTePress

These two expressions have essentially the same relational meaning. They both

express the relation “on” existing between “a book” and “a table”. However, their

referential meaning is significantly different. Expression 1, as a whole, refers to an

object and is called an object referring expression. In referring to an object, 1 uses

the determiner “the” to specify that the object is salient in the context of use of the

expression (and may have previously been referred to). Expression 1 also uses the

word “book” to indicate the type of object being referred to, with “book” functioning

as the head of the expression. Further, the phrase “on the table” refers to a location

with respect to which the object can be identified and functions as a modifier in the

expression. In referring to a location, the expression “on the table” refers to a second

object “the table” and indicates the location of the first object with respect to the

second object. Within the modifying expression, the relation “on” functions as the

relational head with the object referring expression “the table” functioning as a

complement. In expression 1, the relational meaning of “on” is subordinated to

referential meaning with the modifying function of “on the table” dominating the

relational meaning of “on”. That is, although “on” is the relational head of the

prepositional phrase “on the table”, it is not the head of the overall expression and

does not determine the semantic type of that expression.

Expression 2 refers to a situation and is called a situation referring expression.

The auxiliary “is” provides a temporal specification for the situation, fulfilling a

referential function similar to that of the determiner “the” in “the book” and “the

table”. The relational meaning of 2 is about “being on” and not just “being”, with

“on” functioning as the relational head of the situation referring expression. The

relational head of a situation referring expression is called a predicate—reflecting the

assertional function of the relational head. Note that “on” in 1 is not functioning as a

predicate, since it is presupposed and not asserted. That is, relational heads of

modifying expressions are not predicates, they are (modifying) functions. In

expression 2, the object referring expression “the book” functions as the subject

(argument) of “being on” with “the table” functioning as the object (argument).

Referentially, there is also a reference to a location “on the table”, which competes

with the expression of the relational meaning of “on” as reflected in the difference

between:

3. What is the book on?

4. Where is the book?

where 3 highlights the relation “on” in asking about the object of that relation and 4

highlights the reference to a location using “where” to do so.

The terms specifier, head, modifier and complement are borrowed from X-Bar

Theory [9]. It is acknowledged that X-Bar Theory captures an important grammatical

generalization, with the distinction between specifiers and modifiers representing a

significant advance, but X-Bar theory is in need of semantic motivation, which, when

provided, necessitates certain modifications to the theory [10]. For example, the

combination of a specifier and a head results in a maximal projection which

corresponds to a referring expression. However, the specifier determines the type of

referring expression, not the head, with referential type corresponding most closely to

the syntactic type of a maximal projection in X-Bar Theory. The head, on the other

hand, determines the relational type of the expression (where “relational”

encompasses objects). Both referential type (from the specifier) and relational type

(from the head) project to the expression as a whole. Consider the referring

expression

5. The kick

in which the specifier “the” determines the expression to be an object referring

expression, whereas, the head “kick” determines the expression to be a type of action.

In 5, the specifier has the effect of objectifying the action expressed by “kick”,

allowing it to be referred to as though it were an object. Note that since the inherent

relational meaning of “kick” is not affected (only its function), there is no need to

assume that the part of speech of “kick” is a noun instead of a verb (i.e. the head does

not project the “syntactic” type of the referring expression, the specifier does). And if

we allow verbs (especially action verbs) to function as heads of object referring

expressions (i.e. noun phrases), then one of the primary syntactic arguments against

the meaning based definition for parts of speech is nullified. That the head of an

object referring expression need not be a noun is further demonstrated by the

following examples:

6. The cheering up (present participle + verb particle) of the crowd

7. She is an Audrey Hepburn (proper noun) in the rough

8. The up and down (conjoined prepositions) of the elevator

9. Your giving money to strangers (participial phrase) is very generous

Besides demonstrating that the head of an object referring expression need not be a

noun, these examples show the importance of distinguishing the form of an

expression from its function in a particular context.

2 Double R Process

Double R Process is the psycholinguistic theory of language processing underlying

Double R Model. It is a highly interactive theory. Representations of referential and

relational meaning are constructed directly from input texts. There is no separate

syntactic analysis that feeds a semantic interpretation component. The processing

mechanism is driven by the input text in a largely bottom-up, lexically driven manner.

There is no top-down assumption that a privileged linguistic constituent like the

sentence will occur. There is no phrase structure grammar and no top-down control

mechanism. How then are representations of input text constructed? Operating on the

text from left to right, schemas corresponding to lexical items are activated. For those

lexical items which are relational or referential, these schemas establish expectations

which both determine the possible structures and drive the processing mechanism. A

short-term working memory (STWM) [11] is available for storing arguments which

have yet to be integrated into a relational or referential structure, partially instantiated

relational and referential structures, and completed structures. If a relational or

referential entity is encountered which expects to find an argument to its left in the

input text then that argument is assumed to be available in STWM. If the relational or

referential entity expects to find an argument to its right in the input text, then the

entity is stored in STWM as a partially completed structure and waits for the

occurrence of the appropriate argument. When that argument is encountered it is

instantiated into the stored relational or referential structure. Instantiated arguments

are not separately available in STWM. This keeps the number of separate linguistic

units which must be maintained in STWM to a minimum.

3 ACT-R

ACT-R is a cognitive architecture and modeling environment for the development of

computational cognitive models [12]. It is a psychologically validated cognitive

architecture which has been used extensively in the modeling of higher-level

cognitive processes (see the ACT-R web site for an extensive list of models and

publications). ACT-R includes symbolic production and declarative memory

systems integrated with subsymbolic production selection and spreading activation

and decay mechanisms. Production selection involves the parallel matching of the

left-hand side of all productions against a collection of buffers (e.g. goal buffer,

retrieval buffer, visual buffer, auditory buffer) which contain the active contents of

memory and perception. Production execution is a serial process—only one

production is executed at a time. The parallel spreading activation and decay

mechanism determines which declarative memory chunk is put into the retrieval

buffer for comparison against productions. The combination of symbolic and

subsymbolic mechanisms makes ACT-R a hybrid system of cognition. The noise

parameter used by these computational mechanisms adds stochasticity to the system.

ACT-R supports single inheritance of declarative memory chunks and limited,

variable-based pattern matching (including a partial-matching capability). ACT-R

incorporates learning mechanisms for learning both declarative and procedural

knowledge. Version 5 of ACT-R adds a perceptual-motor component supporting the

development of embodied cognitive models. With the addition of the perceptual-

motor component, and the use of buffers as the interface between various cognitive

modules (e.g. vision module, declarative memory), ACT-R is referred to as an

“integrated theory of the mind” [13].

4 Double R Model

Double R Model is the ACT-R based computational implementation of Double R

Grammar and Process (together called Double R Theory). Double R Model is

currently capable of processing an interesting range of grammatical constructions

including: 1) intransitive, transitive and ditransitive verbs; 2) verbs taking clausal

complements; 3) predicate nominals, predicate adjectives and predicate prepositions;

4) conjunctions of numerous grammatical types; 5) modification by attributive

adjectives, prepositional phrases and adverbs, etc. Double R Model accepts as input as

little as a single word or as much as an entire chunk of discourse—using the

perceptual component of ACT-R to read words from a text window. Unrecognized

words are simply ignored. Unrecognized grammatical forms result in partially

analyzed text, not failure. The output of the model is a collection of declarative

memory chunks that represent the referential and relational meaning of the input text.

Although Double R Model is essentially a computational psycholinguistic model, it is

intended to be used as the basis for development of large-scale, functional language

understanding systems and the current coverage of the model will need to be extended

significantly to support that objective.

4.1 Inheritance vs. Unification

Unification allows for the unbounded, recursive matching of two logical

representations and is an extremely powerful pattern matching technique used in

many language processing systems. Unfortunately, it is psychologically too powerful.

For example, the following two logical expressions can be unified:

p(a,B,c(d,e,f(g,h(i,j),K),l))

p(X,b,c(Y,e,f(Z,T,U),l))

where capitalized letters are variables and lowercase letters are constants. Humans are

unlikely to be capable of performing such unifications consciously or otherwise

without significant effort and an external scratch pad, since STWM does not have the

capacity to retain more than a few variable bindings simultaneously.

On the other hand, although extremely powerful, unification does not support the

matching of types to subtypes. Thus, if we have a verb type with intransitive and

transitive verb subtypes, unification cannot unify a chunk of type verb with a chunk of

type intransitive verb or transitive verb. Unification’s inability to match types to

subtypes often results in a proliferation of rules (or conditions on rules) to handle the

various combinations. For example, the verb type can be variableized and a test for

the valid types can be used to constrain the variable (e.g. Verb-Type equal verb or

Verb-Type equal intrans-verb or Verb-Type equal trans-verb). With inheritance, a

production that checks for a verb type will also match a transitive verb and an

intransitive verb type (assuming an appropriate inheritance hierarchy). Humans

appear to be able to use types and subtypes in appropriate contexts with little

awareness of the transitions. For example, when processing a verb, all verbs (used

predicatively) expect to be preceeded by a subject, but only transitive verbs expect to

be followed by an object. Thus, humans presumably have available a general

production that applies to all verbs (or even all predicates) which will look for a

subject preceding the verb, but only a more specialized production for transitive verbs

(or transitive predicates) which will look for an object following the verb.

Inheritance supports the matching of two representations without requiring the

recursive matching of their subparts so long as the types of the two representations are

compatible. Types are essentially an abstraction mechanism which makes it possible

to ignore the detailed internal structure of representations when comparing them. For

example, once the model has identified an expression as an object referring

expression, the model can match the object referring expression against productions

without consideration of the internal structure of the expression. Of course, there may

be productions that do consider the internal structure, but types are useful here as

well. Instead of having to fully elaborate the internal structure, types can be used to

partially elaborate that structure. For example, if a production is specifically

concerned with object referring expressions headed by a quantifier (e.g. “some” in

“some of the books”), the production can check to see that the head is of the

appropriate type, providing a (limited) unification like capability where needed. In

sum, inheritance and limited pattern matching provide a psychologically plausible

alternative to a full unification capability.

To take advantage of inheritance, Double R Model incorporates a type hierarchy (a

tangled hierarchy or lattice, with multiple inheritance, is preferred, but ACT-R

currently only supports single inheritance). Representative elements of the top levels

of the current hierarchy of types (below

top-type) are shown below:

Lexical-type

Noun Adjective Verb Preposition Adverb Determiner

Quantifier Auxiliary

Referential-type

Head Specifier Modifier Complement

Referring-expression-type

Object-refer-expr Situation-refer-expr Predicate-refer-

expr

Location-refer-expr Direction-refer-expr

Relation-type

Relation (with subtypes: Predicate Function) Argument

The more specialized a production is, the more specialized the types of the chunks in

the goal and retrieval buffers to which the production matches will need to be. The

most general productions match a goal chunk whose type is

top-type and ignore the

retrieval buffer chunk.

4.2 Default Rules

ACT-R’s inheritance mechanism can be combined with the subsymbolic production

utility parameter—which influences production selection—to establish default rules.

Since all types extend a base type, using the base type as the value of the goal chunk

in a production will cause the production to match any goal chunk. If the production

is assigned a production utility value that is lower than competing productions, it will

only be selected if no other production matches. A sample default production is

shown below:

(p process-default--retrieve-prev-chunk

=goal> ISA top-type

=context> ISA context

state process

chunk-stack =chunk-stack

=chunk-stack> ISA chunk-stack-chunk

this-chunk =chunk

prev-chunk =prev-chunk

==>

=context>

state retrieve-prev-chunk

chunk-stack =prev-chunk

+retrieval> =chunk)

where the parentheses reflect the underlying Lisp implementation, p identifies a

production,

process-default--retrieve-prev-chunk is the name of the

production,

=goal> identifies the goal chunk, ISA top-type is a chunk type,

=context>

identifies a context chunk, state is a chunk slot, process is a slot

value,

==> separates the left-hand side from the right-hand side and variables are

preceded by

= as in =chunk. This default production causes the previous chunk to be

retrieved from declarative memory (using the

+retrieval> form) if no other

production is selected.

4.3 The Context Chunk and Chunk Stack

The current ACT-R environment provides only the goal and retrieval buffers to store

the partial products of language comprehension, although earlier versions of ACT-R

provided a goal stack. The lack of a stack is particularly constraining, since a stack is

the primary data structure for managing the kind of (limited) recursion that occurs in

language. There needs to be some mechanism for retrieving previously processed

words from STWM in last-in/first-out order during processing (subject to various

kinds of error that can occur in the retrieval process). A stack provides this

(essentially error free) capability. It is expected that a capacity to maintain about 5

separate linguistic chunks in STWM is needed to handle most input—supporting at

least one level of recursion (and perhaps two for the more gifted). The goal chunk

could be adapted for this purpose, except that it is also the basis for creation of new

declarative memory chunks and activation spread and these architectural needs would

conflict. Further, it would be difficult to get the kind of stack like behavior needed out

of the slots in the goal chunk.

To overcome these problems, Double R Model introduces a context chunk

containing a bounded, circular stack of links to declarative memory. As chunks are

stacked in the circular stack, if the number of chunks exceeds the limit of the stack,

then new chunks replace the least recently stacked chunks (supporting at least one

type of STWM error). The actual number of chunks allowed in the stack is specified

by a global parameter. This parameter is settable to reflect individual differences in

STWM capacity. Chunks cannot be directly used from the stack. Rather, the chunk on

the stack provides a template for retrieving the chunk from declarative memory. Since

the chunk must be retrieved from declarative memory before use, the spreading

activation and partial matching mechanisms of ACT-R are not circumvented and

retrieval errors are possible—unlike the goal stack of earlier versions of ACT-R

(which was criticized for this reason). Thus, the bounded, circular stack of links to

declarative memory avoids the arguments against the goal stack of earlier versions of

ACT-R, adds the insight of activated pathways to declarative memory, and retains the

insights that motivated the inclusion of a goal stack in the earlier versions.

Besides storing the chunk stack, the context chunk is also used to separate out state

information from the goal chunk. Since the goal chunk is the basis for creating new

declarative memory chunks, storing the chunk stack in it would result in the chunk

stack being stored with each new declarative memory chunk. While this might be

used to support a kind of episodic memory where the context in which a word occurs

is stored with the declarative memory chunk created during the processing of the

word, ACT-R 5.0 does not currently provide a mechanism for transitioning episodic

memory into semantic memory (i.e. abstracting from the context of use), and storing

the context with a chunk has undesirable side-effects within the ACT-R environment

(e.g. it interferes with the spreading activation mechanism). To avoid such problems a

separate context chunk is maintained and made available to all productions. Although,

the existence of a separate context chunk that productions match to violates the ACT-

R 5.0 architecture where only the buffers are supposed to be used for this purpose,

earlier versions of ACT-R allowed multiple chunks to be matched on the left-hand

side of productions and this functionality is still available in ACT-R 5.0 environment.

The context chunk maintains several pieces of information in addition to the chunk

stack. Its definition (as specified by a

chunk-type) in the model is shown below:

(chunk-type context state rel-context sit-context text-

context

word prev-word-1 prev-word-2 repeat chunk chunk-stack)

In this

chunk-type definition, context is the name of the chunk, state is a slot

that provides state information to guide production selection,

rel-context is a slot

that identifies the current relational context,

sit-context is a slot that contains

information about the current situation context,

text-context is a slot that contains

information about the larger discourse context, word contains the lexical item being

processed,

word-prev-1 and word-prev-2 contains the previous two words

processed,

repeat is yes if the word has been attended to previously and no-more if

there are no more words in the input,

chunk contains the most recently processed

chunk, and

chunk-stack contains the entire chunk stack.

4.4 Lexical and Functional Entries

The lexical entries in the model provide a limited amount of information which is

stored in the

word and word-info chunks. The definition of the word and word-

info

chunk types are provided below:

(chunk-type word word-form word-marker)

(chunk-type word-info word-marker word-root word-type word-

subtype word-morph)

The

word-form slot of the word chunk contains the physical form of the word

(represented as a string in ACT-R); the

word-marker slot contains an abstraction of

the physical form. The

word-root slot contains the value of the root form of the

word. The

word-type slot contains the lexical type of the word and is used to

convert a

word-info chunk into a lexical-type chunk for subsequent processing.

word-subtype slot is provided as a workaround for the lack of multiple

inheritance in ACT-R 5.0. The

word-morph slot supports the encoding of

morphological information

Sample lexical entries for a

noun and verb are provided below:

(cow-wf isa word word-form "cow" word-marker cow)

(cow isa word-info word-marker cow word-root cow word-type

noun

word-morph 3d-per-sing))

(running-wf isa word word-form "running" word-marker

running)

(running isa word-info word-marker running word-type verb

word-root run

word-subtype intrans-verb word-morph pres-part)

Note that there is no indication of the functional roles (e.g. head, modifier, specifier,

predicate, argument) that particular lexical items may fulfill. Following conversion of

word-info chunks into lexical-type chunks, functional roles are dynamically

assigned by the productions that are executed during the processing of a piece of text.

Since functional role chunks are dynamically created, only

chunk-type definitions

exist for functional categories prior to that processing. As an example of a

chunk-

type definition for a functional category, consider the category pred-trans-verb

(transitive verb functioning as a predicate) whose definition involves several

hierarchically related

chunk-types as shown below:

(chunk-type top-type head)

(chunk-type (rel-type (:include top-type)))

(chunk-type (pred-type (:include rel-type)) subj spec mod

post-mod)

(chunk-type (pred-trans-verb (:include pred-type)) obj)

The top-type chunk-type contains the single slot head. All types are subtypes of

top-type and inherit the head slot. Rel-type is a subtype of top-type that

doesn’t add any additional slots.

Pred-type is a subtype of rel-type that adds the

slots

subj, spec, mod, and post-mod. It is when a relation is functioning as a

predicate that these slots become relevant.

Pred-trans-verb is a subtype of pred-

type

that adds the slot obj. Summarizing, pred-trans-verb contains the slots

head, subj, spec, mod (pre-head), post-mod (post-head), and obj, all of which

are inherited from parent types except for the

obj slot.

The following production creates an instance of a

pred-trans-verb providing

initial values for the slots:

(p process-verb--pred-trans-verb

=goal> ISA verb

head =verb

subtype trans-verb

=context> ISA context

state convert-verb-to-pred-verb

==>

+goal> ISA pred-trans-verb

subj none

spec none

mod none

head =goal

post-mod none

obj none

=context>

state retrieve-prev-chunk)

In this production, a

verb (subtype of lexical-type) whose subtype slot has the

value

trans-verb is converted into a pred-trans-verb for subsequent

processing. The only slot of

pred-trans-verb that is given a value other than

none is the head slot whose value is set to be the

goal chunk (head =goal). This

production has the effect of assigning a transitive verb the functional role of predicate

(specialized as a transitive verb predicate). Its selection and execution is based on the

previous context which set the value of the

state slot of the context chunk to

convert-verb-to-pred-verb and on having a goal chunk of type verb whose

subtype slot has the value trans-verb.

4.5 Productions

Sample productions were shown above in the discussion of default rules and in the

creation of functional roles. This section discusses the productions used in the

processing of the word “kick” following the word “the” in the expression “the kick”.

The

read-next-word production initiates the find-attend-encode sequence for

reading the next word from the computer screen (using ACT-R’s perceptual

component). Following the

read-next-word production, the retrieve-word-

info production retrieves the word-info chunk. The word-info chunk is then

used by the

convert-word-to-verb production to create a verb lexical-type

chunk which becomes the goal. Then, the

process-verb--obj-context--

convert-to-rel-head

production matches a verb goal chunk and in the context

of an

obj (object referring expression) converts the verb type into a rel-head type

(using the

+goal> form)

(p process-verb--obj-context--convert-to-rel-head

=goal> ISA verb

head =verb

=context> ISA context

state retrieve-prev-chunk

rel-context obj

==>

+goal> ISA rel-head

mod none

head =goal

post-mod none)

Rel-head (relational head) is a subtype of head. The process-head--prev-chunk-is-obj-spec

production matches a

head goal chunk (which could be a rel-head) and an obj-spec

(object specifier) retrieval chunk and creates a new obj-refer-expr (object referring

expression) which becomes the goal. Together, these two productions support to use

of verbs as (relational) heads of object referring expressions following an object

specifier.

(p process-head--prev-chunk-is-obj-spec

=goal> ISA head

=context> ISA context

state retrieve-prev-chunk

=retrieval> ISA obj-spec

==>

+goal> ISA obj-refer-expr

spec =retrieval

mod none

head =goal

post-mod none

referent none-for-now

=context> state process

rel-context none)

The creation of an object referring expression causes the value of the rel-context

(relational context) slot to be set to none indicating the end of the object referring

expression context.

4.6 Context Accommodation vs. Backtracking

Context accommodation is a mechanism for changing the function of an expression

based on the context without backtracking. For example, when an auxiliary verb like

“did” occurs it is likely functioning as a predicate specifier as in “he did not run”

where the predicate is “run” and “did not” provides the specification for that

predicate. However, auxiliary verbs may also function as predicates when they are

followed by an object referring expression as in “he did it”. Determining the ultimate

function of an auxiliary can only be made when the expression following the auxiliary

is processed. In a backtracking system, if the auxiliary is initially determined to be

functioning as a predicate specifier, then when the noun phrase “it” occurs, the system

will backtrack and reanalyze the auxiliary, perhaps selecting the predicate function on

backtracking. However, note that backtracking mechanisms typically lose the context

that forced the backtracking. Thus, on backtracking to the auxiliary, the system has no

knowledge of the subsequent occurrence of a noun phrase to indicate the use of the

auxiliary as a predicate. Thus, the system can only randomly select a new function for

the auxiliary which may or may not be that of a predicate.

A better alternative is to accommodate the function of the auxiliary in the context

which forces that accommodation. In this approach, when the noun phrase “it” is

processed and the auxiliary functioning as a predicate specifier is retrieved, the

function of the auxiliary can be accommodated in the context of a subsequent noun

phrase to be a predicate. Context accommodation avoids the need to backtrack and

allows the context to adjust the function of an expression just where that

accommodation is supported by the context. Of course, there may cases where the

context accommodation mechanism breaks down and some form of backtracking is

needed (e.g. garden-path sentences), but in such cases backtracking is likely to

involve a jump back to the beginning of a major constituent (e.g. clause) and some

contextual information will be carried back with the jump. In any case, a reverse-

depth-first, context-unraveling backtracking mechanism like that provided in Prolog

is psychologically implausible.

Context accommodation assumes the activation, selection and integration of the

most appropriate schema given the current context, subject to accommodation based

on the subsequent context. Context accommodation is highly compatible with

Preference Semantics [14], and naturally handles the cases where the initially

preferred choice turns out not to be appropriate in the wider context.

5 Summary

Double R Model may be the first attempt at the development of an NLU system

founded on the principles of Cognitive Linguistics and implemented using the ACT-R

cognitive architecture and modeling environment. In its current state it demonstrates

the possibility of building such a system. Much work remains to be done before the

feasibility of building functional NLU systems using this approach can be fully

demonstrated. For more details of the theory and the full source code, see the Double

R Theory web site at www.DoubleRTheory.com.

References

1. Ball, J. (2003). “Double R Grammar.” http://www.

DoubleRTheory.com/DoubleRGrammar.pdf

2. Langacker, R. (1987, 1991). Foundations of Cognitive Grammar, Vols 1 and 2. Stanford,

CA: Stanford University Press.

3. Talmy, L. (2003). Toward a Cognitive Semantics, Vols I and II. Cambridge, MA: The MIT

Press

4. Lakoff, G., & M. Johnson (1980). Metaphors We Live By. Chicago: The University of

Chicago Press.

5. Wittgenstein, L. (1953). Philosophical Investigations. New York: MacMillan.

6. Rosch, E. (1978). “Principles of Categorization.” In Cognition and Categorization. Edited by

E. Rosch & B. Lloyd. Hillsdale, NJ: LEA.

7. Lakoff, G. (1987). Women, Fire and Dangerous Things. Chicago: The University of Chicago

Press.

8. Nirenburg, S. & Raskin, V. (1995). “Ten Choices for Lexical Semantics”. Memoranda in

Computer and Cognitive Science, MCCS-96-304. CRL, NMSU, Las Cruces, NM.

9. Chomsky, N. (1970). “Remarks on Nominalization.” In R. Jacobs & P. Rosembaum, eds.,

Readings in English Transformational Grammar. Waltham, MA: Ginn.

10. Ball, J (2003). “Towards a Semantics of X-Bar Theory.”

http://www.DoubleRTheory.com/SemanticsOfXBarTheory.pdf

11. Kintsch, W. (1998). Comprehension, a Paradigm for Cognition. New York, NY:

Cambridge Univ Press.

12. Anderson, J. & Lebiere, C. (1998). The Atomic Components of Thought. Mahway, MJ:

LEA.

13. Anderson, J., Bothell, D., Byrne, M., LeBiere, C (in press). “An Integrated Theory of the

Mind” Psychological Review

14. Wilks, Y. (1975). “A preferential, pattern-seeking semantics for natural language

inference.” Artificial Intelligence, 6:53-74