A Neural Network Model for the Representation of Natural Language

impulseverseΤεχνίτη Νοημοσύνη και Ρομποτική

24 Οκτ 2013 (πριν από 3 χρόνια και 7 μήνες)

73 εμφανίσεις

A Neural Network Model for the Representation of Natural Language


Eleni Koutsomitopoulou

Georgetown University (Washington, DC) and LexisNexis (London, Great Britain)

Eleni.Koutsomitopoulou@lexisnexis.com



Current research in natural language processing
demonstrates the importance of analyzing
syntactic relationships, such as word order, topicalization, passivization, dative movement,
particle movement, pronominalization as dynamic resonant patterns of neuronal activation
(Loritz, 1999). Following this li
ne of research this study demonstrates the importance of also
analyzing conceptual relationships, such as polysemy, homonymy, ambiguity, metaphor,
neologism, coreference, as dynamic resonant patterns represented in terms of neuronal
activation. This view
has implications for the representation of natural language (NL).
Alternatively, formal representation methods abstract away from the actual properties of real
-
time natural language input and rule
-
based systems are of limited representational power.


Since

NL is a human neurocognitive phenomenon we presume that it can be best
represented in a neural network model. This study focuses on a neural network simulation, the
Cognitive Linguistic Adaptive Resonant Network (CLAR
-
NET) model of online and real
-
time as
sociations among concepts. The CLAR
-
NET model is a simulated Adaptive
Resonance Theory (ART, Grossberg 1972 et seq.) model. Through a series of experiments, I
address particular linguistic problems such as homonymy, neologism, polysemy, metaphor,
construct
ional polysemy, contextual coreference, subject
-
object control, event
-
structure
metaphor and negation. The aim of this study is to infer NL specific mappings of concepts in
the human neurocognitive system on the basis of known facts and observations provi
ded
within the realms of conceptual metaphor theory (CMT), and adaptive grammar (AG, Loritz
1999), theories of linguistic analysis, and known variables drawn from the brain and cognitive
sciences as well as previous neural network systems built for similar

purposes.

My basic hypothesis is that the association among concepts is primarily an expression of
domain
-
general cognitive mechanisms that depend on continuous learning of both previously
presented linguistic input and everyday, direct experiential (i.e
. sensory
-
physical) behaviors
represented in natural language as "common knowledge" (or "common sense"). According to
this hypothesis, complex conceptual representations are not actually associated with pre
-
postulated feature structures, but with time
-
sens
itive dynamic patterns of activation. These
patterns can reinforce previous learning and/or create new "place
-
holders" in the conceptual
system for future value binding. This line of investigation holds implications for language
learning, neurolinguistics,

metaphor theory, information retrieval, knowledge engineering,
case
-
based reasoning, knowledge
-
based machine translation systems and related ontologies.

This study finds that although short
-
term memory (STM) effects in ART
-
like networks are
significant,
most of the time long
-
term memory (LTM) calculation yields better semantic
discrimination. It is suggested that the internal structure of lexical frames that correspond to
clusters of congenial associations (in fact, neuronal subnetworks), is maintained as

long as it
resonates with new input patterns or learned in long
-
term memory traces. Different degrees
of similarity (or deviation) from previously acquired knowledge clusters are computed as
activation levels of the corresponding neuronal nodes and may b
e measured via differential
equations of neuronal activity. The overall conclusion is that ART
-
like networks can model
interesting linguistic phenomena in a neurocognitively plausible way.