A Natural Language Translation Neural Network

proper translation by a user without any expert knowledge of how the computer stores and represents rules. This paper demonstrates the utility of neural networks in precisely this area on a small scale translation problem. We have tested the ability of neural networks to perform natural language translation. Our results have shown a greatly improved translation accuracy in comparison to the work of R.B. Allen (1987) in translating English into Spanish. A neural network was trained on a set of 10,000 sentences from a total of 24,750 sentences using a novel training algorithm. On a test set of 100 sentences the neural network showed a 98% sentence accuracy. The neural network had 48 input nodes, 70 nodes in the first hidden layer, 1 node in the third hidden layer, and 36 nodes in the output layer (48-701-36). A fully connected architecture was used. Connectionist NLP Research has already shown the usefulness of neural networks in various natural language processing tasks: (Allen, 1987), (Jain, 1991), (Waibel, 1988) and (Waibel et al, 1991).