A major project has been the development of the Hyperspace
Analogue to Language (or HAL), which is a computer simulation of
human memory. HAL has a lexicon of 70,000 items and learns its representations
as a function of the contexts in which words occur. This is accomplished
with a concept-acquisition process that requires no supervision
using an input of 320 million words of text. Word meanings (broadly
based) are represented in a 140,000 dimensional space (thus, Hyperspace
Analogue to Language). The model accounts for a wide range of semantic,
language, grammatical, and syntactic phenomena. New areas of exploration
for the model involve commercial and forensic applications as well
as memory disorders in deep dyslexia, schizophrenia, AlzheimerÕs
and normal aging.
Our most recent work involves incorporating the HAL
semantic representations into a connectionist style processing model.
In this way we can begin to explore the interaction between processing
and representational issues.
|
|