Published online by Cambridge University Press: 14 October 2010
A connectionist back-propagation net is trained to learn a mapping between verb forms (infinitive, imperative, present tense) and past tense forms. A back-propagation network requires that a one-to-one mapping exists between input and output. Since Swedish past tense is characterized by choice between weak and strong forms this proves a hard task for the net. Examples are: nypte/nöp ‘pinched’, tvang/tvingade ‘forced’, where there is no obvious reason for preferring one form over the other. The task is still worth doing since regularities emerge in the net over time. The performance of the connectionist model is analysed and compared to previous results.