A biologically plausible and computationally efficient architecture and algorithm for a connectionist natural language processor

Nowadays, most connectionist models are more oriented to computational efficiency instead of neurophysiological inspiration. Classical learning algorithms, like the largely employed back propagation, is argued to be biologically implausible. This paper aims to prove that a biologically inspired connectionist architecture and algorithm is not only capable of dealing with a high level cognitive task, like a natural language processing application, but also be more computationally efficient. It is presented a comparison between a standard simple recurrent network using back propagation with a physiologically inspired system. Symbolic data, extracted from connectionist architectures, show that the physiologically plausible model displays more expectable semantic features about thematic relations between words than the conventional one.