|
Munro, Cosic, and Tabasko (1991) worked with NN that
translated prepositions from English to German. The
input to the network consisted of two nouns, a locative
preposition, and a representation of the spatial
relationship between the two nouns. The output consisted
of a copy of the preposition, and its semantics for the
specific phrase it was used on. The prepositions
{in, at, on, under, above} thus got mapped to the
spatial relationships {over, at the edge of, embedded
in, contains, within border of, touching, near, far
from, supports}. By doing this the semantics of the
preposition was grounded based on the nouns it was
used with. After using this method to train one network
in English and one in German, the two networks were
connected together, and translation from one language
to the other was performed via the semantic
representation associated with the preposition.
Their networks obtained error rates of less than 0.5%,
but no network managed to respond correctly to all 137
sentences in the training set.
|
Back to the previous topic
Back to the Table of Content
To the next topic