ANNP: a neural network parser for real world texts

A neural parser is described that computes sentence structure and achieves compositionality in a simple and effective way. The model is compositional in the sense that it is able to analyze new structures which are recursive combinations of known structures. The model´s performance is compared to a recently proposed neural parser in terms of efficiency and computational capacity. To test the efficiency of the model we ran two groups of experiments. In the first group, we used the same training and test sentences as did Mikkulainen (1996). In the second group of experiments, we used real texts and integrated the parser with a syntactic disambiguation system as well as a semantic disambiguation system. The objective of the second group of experiments was to maximize the compositionality; from the simplest training set possible, the maximum number of complex sentences could be analyzed in the test phase. The results obtained are very promising