| Koncar and Guthrie (1994) used NN to translate sentences from English to Serbo-Croatian. Each sentence in English consisted of two determiners, two nouns, a verb, an adverb, a
preposition and an adjective. Each Serbo-Croatian sentence consisted of two nouns, a verb, and
adverb, a preposition and an adjective. The position of the verb was randomly chosen among
various valid positions. Of particular interest is the fact that certain words changed translation
depending on context. In addition, the translation was not a simple, direct mapping because of
the difference in number of words and the context sensitivity of translating adjectives and nouns
between these two languages. Regarding the structure of their network, they found that the
effectiveness of the NN was very susceptible to the initial weight values, that too many nodes in
the hidden layers caused the network to loose its ability to generalize, and that the first hidden
layer served as a feature detector, while successive layers tried to combine these features in order
to achieve the desired output. These findings corroborate those of other researchers, such as
Caudill (1990), Lang and Witbrock (1988), and Mirchandani and Cao (1989).
They also found that traiinng with the Backpropagation algorithm would not produce satisfactory
results. Because of this, they created a training algorithm that used Backpropagationas one of ots
steps. This algorithms added hidden layers and gidden nodes until the error fell below an
acceptable value. The error tended to decrease with an increase of hidden layers and nodes, but
only to a certain point. |
Back to the previous topic
Back to the Table of Content
To the next topic