RDF-to-Text using Generative Adverserial Networks

We envisage the usage of GANs for this problem statement will result in fluent and adequate verbalization of the input RDF triples from DBpedia, given the use of human evaluated text to train the discriminator, as well as reinforcement learning approaches to guide the convergence of the generator.

The discriminator is trained to distinguish real text from synthetically generated text, using the generator’s output and a human-evaluated dataset. The generator is updated by employing a policy gradient and Monte Carlo (MC) search on the basis of the expected end reward received from the discriminator model. The reward is estimated by the likelihood that it would fool the discriminator model. We envisage that this approach will result in more fluent and adequate verbalizations. It is also noteworthy that this would be a novel approach to solve the problem statement which has never been previously explored in literature. Closest available research is that of (Yu et al., 2017), however they do not rely on graphical input representations for their model, and employ an RNN based encoding module within their generator network.