Sophie Burkhardt(Data Mining Group, Johannes Gutenberg University of Mainz)
hosted by Department of Computer Science
"Generative Topic Models and Disentanglement of Style and Topic in Text Data"
Large amounts of text data increase the need for scalable andinterpretable models to analyze such data. The main drawback of unsupervisedtraining methods is the lack of control one has on the learned representations.It would be desirable for the learned topics to capture relevant informationthat the user is interested in as well as for the topics to be disentangled,meaning that they should not contain redundant information. This is a difficultyof current methods for topic modeling that often exhibit component collapsing.Another goal is to also take the word order into account to be able to analyzeand control the style of text in addition to topics and content. This talk willsummarize past research on generative topic models and discuss future work ontext generation and the learning of disentangled representations.
|Time:||Monday, 02.03.2020, 11:00|