Recurrent neural network to acquire the grammatical competence

The author examines whether a fully recurrent neural network can be trained to acquire grammatical competence. To simplify the experiments, grammatical competence means the ability to infer the well-formedness of sentences. These abilities must be applied not only to familiar sentences but also to sentences never heard before. To simulate the process of acquisition of grammatical competence with the property of creativity, training sentences were made by using restricted sentence formulas. Then, the performance of the recurrent neural network was evaluated as to the inference of well-formedness of sentences, when the new sentences have sentence formulas, equivalent to or different from the formula of training sentences. All experimental results seemed to suggest that the recurrent neural network had the ability to acquire grammatical competence with the property of creativity