A natural language interface to the Unix operating system using recurrent neural networks

This paper presents an experiment that was done in natural language processing using a recurrent neural network, specifically attempting to construct a natural language interface to the Unix operating system. Most natural language experimentation with neural networks has been done using a small vocabulary. The experiments in this paper were done to try to address what would happen using a larger vocabulary. Several different trials were run using binary inputs, bipolar inputs and a variety of different network architecture. When the input to the network was binary the network was able to correctly classify 85 percent of the sentences using the training set; when using bipolar input, the net was only able to classify 62 percent of the sentences correctly. Neither network did well with generalization.