Names Generation with Generative Adversarial Networks (GANs)[go to overview]
Text generation represents an active research field, especially with generative models. This thesis uses the Generative Adversarial Networks architecture for the purpose of names generation. This architecture was not initially designed to work on discrete data, but thanks to its high performance in other fields (e.g. computer vision) and the advantage of easy data creation, the ongoing research of the community tries to answer the question if the above mentioned architecture is suitable for text (here: name) generation.The discrete nature of names makes it difficult to leverage the potential of this framework. The Gumbel-Softmax reparameterization technique proposed in the literature is used in this thesis to overcome this drawback.Different training scenarios are carried out in order to compare the performance of the mentioned reparameterization technique. In this context, the impact of using different loss functions is analyzed and assessed to obtain a good name generator. Each training scenario is represented by a different model. Every model is interpreted with the help of a data set containing real names, which is also used for the training. The analysis is being conducted based on the characteristics of the data set: distribution of n-grams, length distribution and percentage of overlapping names. Based on its results, the best models are being selected manually for the final assessment of the quality of the generated names - a crowdsourcing survey. The human appreciation offers valuable feedback about the acceptance of the generated names as being real (or not).
11.07.19 - 10:15