Switching GRUs on a deep sequence-to-sequence neural model: an application to Natural Language Generation

In the last few years, many different methods focus on a Recurrent Neural Network (RNN) approach to the generation of natural language from tabular information. The most widely used word based sequence-to-sequence neural methods resort to a preprocessing step called delexicalization (conversely, relexicalization) to deal with the rare word problem. In this work, we present a character-level sequence-to-sequence model with attention mechanism which does not requires delexicalization, tokenization nor even lowercasing. Moreover, our model leverages two major features: it switches between a generation and a copying mechanism when required, showing the ability to directly copy input facts to produce outputs, and it is trained in such a way that significantly reduces recall issues, i.e. the possible omission of input information, a well-known issue in the context of Recurrent Neural Networks.