|
Elman (1993) has found that networks are able to learn
grammars better when the complexity of the sentences is
increased gradually, as opposed to presenting exemplars
of the complete language from the beginning. He
defined five levels of complexity, and divided the
sentences in the training set into two levels of
complexity. In the first phase 10,000 simple sentences
were presented. After this phase, a new group of 10,000
sentences was presented, but in this case 2,500 of
these sentences belong to the "difficult" group. The
third phase had 5,000 difficult and 5,000 simple
sentences. The fourth phase had 7,500 difficult
sentences, and 2,500 simple. The final phase consisted
of 10,000 complex sentences. With this setup, the
networks managed to learn the grammar represented by
these 50,000 sentences better.
|
Back to the previous subject
To the next subject