A Connectionist Automatic Encoder and Translator for Natural Languages

One of the problems of the connectionist translator RECONTRA is the representation of the vocabularies of the languages implied in the task to be translated. In previous work, a simple connectionist model was used to provide automatic codifications for RECONTRA, but sometimes these codifications have shown not to be adequate for the translation task. In this paper we aim to extend the RECONTRA topology in order to integrate the creation of the codifications (for the languages to be translated) and the translation task in an unique connectionist architecture. To do that, a new hidden layer is added to the network, as it’s known how a neural network develops its own internal representation of its input.