A Connectionist Network That Learns To Process Some (Very) Simple Sentences.

The application of the "back-propagation" learning algorithm to the task of determining the right set of features corresponding to the words in an input sentence is described. Features that are specific to particular nouns and verbs, that indicate whether a nominal constituent is singular or plural, definite or indefinite, and that furnish case-frame information, are discussed. On examination, it appears that the network has learned concepts appropriate to the domain of natural language processing. The learning also generalizes well to novel sentences. Three related experiments are described. The shortcomings of the network are discussed, and ideas are suggested for an alternative model that should overcome some of these shortcomings. (Author/DJD)