Centre for Language Technology – Department of Nordic Studies and Linguistics – KU

We are interested in automatically generating rap lyrics, using a dataset with over 30,000 rap songs. Past approaches in poetry/verse generation typically involve analytical approaches with hand-crafted grammars and rules. In rap lyrics generation, there are hybrid approaches, where the authors devise an architecture to structurally analyze rap texts and suggest the closest matching next line from a database. On the other hand, we have pure unsupervised approaches using the Long Short Term Memory (LSTM) Recurrent neural networks (RNN) variant with promising results, albeit using a limited corpus. Recently, Generative Adversarial Networks (GANs) and similar methods have been proved to model well specific styles.

Faculty of Science – Department of Mathematics and Computer Science (IMADA) – SDU

For an evaluation study on unsupervised outlier detection and ensemble techniques, we run various methods with various parameterizations on many datasets, resulting in ca. 4.5 mio. individual experiments which are to a certain extent independent of each other and try to see what the benefits of parallelization are on this effort.