Natural language processing with recurrent neural networks

Various experiments have been done on natural language using neural networks. Qualities such as distributed representation, the ability to generalize and efficient parallel computation make neural networks very attractive for natural language processing. In this work, a particular model is considered: recurrent neural networks. Their main advantage over feed-forward models is that they require fewer units to represent the input, and this results in a smaller network. Their main disadvantage is that they are harder to train. Some solutions to this problem are presented along with a comparison in performance of recurrent neural networks versus feed-forward neural networks in three different tasks: syntactic feature extraction, sentence comparison and translation between English and Spanish.