| |
Abstract:
Semantic ambiguity offers a challenge to computational models
of memory. Traditional approaches have either used binary vector
representations to insure that the meanings are distinct or have
used hand-coded semantic features that are intuitively plausible
but have no empirical basis. The drawback to these approaches is
that they do not scale up to real language issues (large
vocabularies or large corpra of languages), nor are they based in
the actual statistical properties of language use. The HAL model of
memory uses a large sample of text (320 million words) and a simple
learning algorithm to acquire semantic representations for 70,000
words. Distributed high-dimensional memory models, such as HAL,
have a limitation (like most distributed models) in that ambiguous
word representations carry information related to both meanings
thus leaving the disambiguation process to a "selection mechanism."
Selection mechanisms have been underspecified, at best, in
computational models. In our approach, we show - at the
representational level - how a system can refine ambiguous meaning
representation using a bootstrapping process in the high
dimensional meaning space. We argue that this bootstrapping process
is an important part of the representational acquisition of
ambiguous word meanings.
|
