Lynellen Perry

PRELIMINARY EXAM

QUESTION THREE

Lynellen D. S. Perry

A Preliminary Exam Response
Submitted to the Faculty of
Mississippi State University
in Partial Fulfillment of the Requirements
for the Degree of Doctor of Philosophy
in Computer Science
in the Department of Computer Science

Mississippi State, Mississippi

February 199

QUESTION 3

Psychological studies have indicated that people can comprehend metaphors just as quickly as
they can comprehend equivalent literal constructs [Lytinen, Burridge, and Kirtner, 1992]. As you
pointed out in your dissertation topic description, researchers are now using a wide range of tools and
techniques in natural language processing and understanding. Do you think that the incorporation of
fractals, chaos, dynamical systems, and/or neural networks into a natural language understanding system
can improve the ability of the system to handle metaphors? If your answer is “no,” explain why these
approaches to not offer any benefits to such a system. If your answer is “yes,” illustrate how this could
be done by selecting one technique (or a combination of techniques) and describing how that technique
could provide useful information to an NLU system in the understanding of metaphors.

I believe it is possible that the incorporation of dynamical systems and neural
networks would improve the ability of a natural language understanding (NLU) system to
handle metaphors. Metaphor is the comparison of two objects, with the suggestion of a
likeness between them, or “an expression that applies one domain to another” (Jones and
McCoy, 1992).
Metaphors can be divided into three basic categories: dead, conventional, and
novel. Dead metaphors are basically frozen phrases. They may have once been
novel metaphors, but the tracing of their meaning has been lost. Conventional
metaphors have been connoted wit the more transparent metaphors. They
generally follow common patterns and associations. Finally, new metaphors that
are not easily realted to specific patterns are considered fresh, novel, or creative
metaphors (Jones and McCoy, 1992).

Even if there is no direct or practical way of using dynamical systems for
comprehending metaphors, a change in the way language researchers view language
processing would be a useful on its own. For example, instead of viewing language as
statistical or combinatorial, Poston (1987) says that a continuously dynamical approach is
useful.
A metaphor works when the concepts involved `feel the same’ to the
visual/aural/kinesthetic imagination: is this combinatorial in character? . . . The
heart of metaphor being recognition of similarity, those workers in pattern
recognition who seek to find a `grammar’ of images, modeled on few-finite-state
grammars of language, may be importing a fundamental error about language into
a field where it will be even less fruitful. The task for a theorist, or a machine
analyst, of language is to predict the results of the human tumbling of superficially
discrete words in a system of continuous dynamics (Poston, 1987).

There are several ways in which the concepts of dynamical systems are similar to
word sense disambiguation (or metaphor comprehension). First, Lytinen et. al (1992)
explain that mapping rules are used “to transform the literal meaning to the intended
meaning”. If we view our mental representation of word senses in the way that Elman
(1995) suggests we view the lexicon in a language processing system, then the word
senses are regions of state space. Grammar is then the “dynamics (attractors and
repellers) which constrain movement in that space” (Elman, 1995). The constraint rules
presented by Lytinen et. al (1992) would be analogous to Elman’s (1995) concept of
grammar because they constrain “the interpretation which the parser can build in some
way, by limiting, for a class of nodes, the set of arcs that can lead from a node of that
class, as well as the types of nodes that arcs can lead to” in the same way that syntactic
grammar rules work with natural languages.
The mapping rules are similar to a grammar, but the term and process of expanding
a grammar into concrete examples also reminds me of the transformations that are found
in Iterated Function Systems (IFSs). In an IFS, a set of rules is used to transform a point
(or set of points) using the operators of sheering, scaling, rotation, and reflection (Peitgen
et. al, 1992). The mapping rules use their own operators to move the language
understanding system from one location in the space of word meaning to another (or
several other) location(s) which represents a separate word meaning. Again, Elman
(1995) says “representations are not abstract symbols but rather regions of state space.
Rules are not operations on symbols but rather embedded in the dynamics of the system, a
dynamics which permits movement from certain regions to others while making other
transitions difficult.”
Poston (1987) is, among other things, examining the way that word meanings
change over the decades and centuries. However, his ideas could be applied to the way
that a word changes meaning in the mind of the reader while they are processing a
metaphor. Perhaps the reader slowly changes the meaning of the word as they process
other context cues, or perhaps they jump directly to a particular meaning. Poston (1987)
argues that “meanings can change continuously over time, contrary to combinatorics, and
when they do shift discontinuously the jumps may be better modeled by continuous
dynamics with bifurcating attractors than by discrete models.” As the NLU system works
to decide word meaning, “a natural approach, dynamically, is to think of each final
decision as an `attractor’” (Poston, 1987).
Both Elman (1995) and Resnik (1991) use recurrent neural networks as dynamical
systems to automatically discover lexical classes from a corpora. By correlating words
into groups “whose similar behavior indicates that they have attributes or features in
common” (Resnik, 1991), they show that “representations which are near one another in
representational space form classes, and higher-level categories correspond to larger and
more general regions” of the representation space. Perhaps a similar method could be
used to correlate the literal and metaphorical senses of words. This correlation
information could then be given to the NLU system to aid it in understanding metaphors.

REFERENCES

Elman, Jeffrey L. 1995. Language as a dynamical system. In Robert F. Port and T. van
Gelder (Eds.) Mind as motion: Explorations in the dynamics of cognition.
Cambridge, MA: MIT Press, pp. 195-223.

Jones, Mark A. and Kathleen F. McCoy. 1992. Transparently-motivated metaphor
generation. Aspects of Automated Natural Language Generation (Proceedings of the
6th International Workshop on Natural Language Generation, April 5-7, 1992, in
Trento, Italy). Berlin: Springer-Verlag. 231-246.

Lytinen, Steven L., Robert R. Burridge, and Jeffrey D. Kirtner. 1992. Literal meaning and
the comprehension of metaphors. Proceedings of the Tenth National Conference on
Artificial Intelligence, July 12-16, 1992. AAAI Press/The MIT Press. 309-314.

Peitgen, Heinz-Otto, Hartmut Jurgens, and Dietmar Saupe. 1992. Fractals for the
Classroom, Part One: Introduction to Fractals and Chaos. New York: Springer-
Verlag.

Poston, Tim. 1987. Mister! Your back wheel’s going round! In Thomas T. Ballmer and
Wolfgang Wildgen (Eds.) Process Linguistics. Tubingen: Max Niemeyer Verlag, pp.
11-36.

Resnik, Philip. 1991. An investigation of lexical class acquisition using a recurrent neural
network.