Using neural networks to infer grammatical structures in natural language

The work described takes natural language text and shows how grammatical structure can be mapped onto it, with a hybrid system that uses both neural networks and rules. Sentences are decomposed, with markers inserted in all possible positions to define the boundaries of grammatical features. These possibilities are reduced by applying rules that prohibit certain combinations. Then the pattern matching capabilities of the neural network are used to select from the remaining possibilities the correct grammatical mapping. A single layer, feed forward, higher order network is used. The work describes how the processing principles can be applied. An unrestricted natural language vocabulary is allowed. The particular task undertaken is to decompose sentences into subject and predicate, then to detect the head of the subject