|
|
Abstract
This paper presents a neural network approach to document semantic
indexing. A hopfield net algorithm was used to simulate human associative
memory for concept exploration in the domain of computer science and
engineering. INSPEC, a collection of more than 320,000 document abstracts
from leading journals, was used as the document testbed. Benchmark tests
confirmed that three parameters (maximum number of activated nodes,
maximum allowable error, and maximum number of iterations) were useful
in positively influencing network convergence behavior without negatively
impacting central processing unit performance. Another series of benchmark
tests was performed to determine the effectiveness of various filtering
techniques in reducing the negative impact of noisy input terms.
Preliminary user tests confirmed the expectation that the Hopfield net
algorithm is potentially useful as an associative memory technique to
improve document recall and precision by solving discrepancies between
indexer vocabularies and end-user vocabularies.
|
