Observations of complex data in the real world might be caused by several underlying generative factors that account for specific perceptual features. However, these factors are usually entangled and can not be underlined directly from the data. Modeling such factors could generalize the learning of complex concepts through compositions of simpler abstractions. This enables us to understand the inner structure of the data, to process it efficiently and to control meaningful generative processes which may eventually open upon artificial creativity and machine intelligence. Deep neural networks have been very successful at the automatic extraction of features from various data distributions, making manual feature extraction obsolete. However, due to the complexity of these neural networks, the extracted features often themselves are highly complex. Which hinders reusing them for downstream tasks as humans can not interpret the extracted meaning of these features due to their entanglement. Often, neural networks are rather treated as a black box and we have to trust external evaluation metrics such as train and test error. It would be beneficial to understand what kinds of hidden representations the model has learned. Interestingly, several methods exist that are particularly suited for learning meaningful hidden representations. In the image domain, an extensive body of research has been carried. Through various deep generative models such as Generative Adversarial Networks (GAN) and Variational Autoencoders (VAE). For example, when being fed with images of faces, a VAE might automatically learn to encode a person's gender and beard length/existence into two separate hidden variables. These disentangled features we could then use to generate new images which are similar to the underlying image distribution of the images the network was trained with. The goal of this research is to extend these promising results into the natural language text-domain. By performing experiments with different neural network architectures for the feature extraction and sample generation to find a possible disentangled hidden representation of sentences. State of the art for representation learning in NLP is limited, therefore we are constructing our new dataset, the "dSentences" dataset. Thus, we are performing research from scratch by transferring and adapting knowledge and approaches from the image domain into the natural language domain.     «

Observations of complex data in the real world might be caused by several underlying generative factors that account for specific perceptual features. However, these factors are usually entangled and can not be underlined directly from the data. Modeling such factors could generalize the learning of complex concepts through compositions of simpler abstractions. This enables us to understand the inner structure of the data, to process it efficiently and to control meaningful generative processes...     »

Observations of complex data in the real world might be caused by several underlying generative factors that account for specific perceptual features. However, these factors are usually entangled and can not be underlined directly from the data. Modeling such factors could generalize the learning of complex concepts through compositions of simpler abstractions. This enables us to understand the inner structure of the data, to process it efficiently and to control meaningful generative processes which may eventually open upon artificial creativity and machine intelligence. Deep neural networks have been very successful at the automatic extraction of features from various data distributions, making manual feature extraction obsolete. However, due to the complexity of these neural networks, the extracted features often themselves are highly complex. Which hinders reusing them for downstream tasks as humans can not interpret the extracted meaning of these features due to their entanglement. Often, neural networks are rather treated as a black box and we have to trust external evaluation metrics such as train and test error. It would be beneficial to understand what kinds of hidden representations the model has learned. Interestingly, several methods exist that are particularly suited for learning meaningful hidden representations. In the image domain, an extensive body of research has been carried. Through various deep generative models such as Generative Adversarial Networks (GAN) and Variational Autoencoders (VAE). For example, when being fed with images of faces, a VAE might automatically learn to encode a person's gender and beard length/existence into two separate hidden variables. These disentangled features we could then use to generate new images which are similar to the underlying image distribution of the images the network was trained with. The goal of this research is to extend these promising results into the natural language text-domain. By performing experiments with different neural network architectures for the feature extraction and sample generation to find a possible disentangled hidden representation of sentences. State of the art for representation learning in NLP is limited, therefore we are constructing our new dataset, the "dSentences" dataset. Thus, we are performing research from scratch by transferring and adapting knowledge and approaches from the image domain into the natural language domain.     «

Observations of complex data in the real world might be caused by several underlying generative factors that account for specific perceptual features. However, these factors are usually entangled and can not be underlined directly from the data. Modeling such factors could generalize the learning of complex concepts through compositions of simpler abstractions. This enables us to understand the inner structure of the data, to process it efficiently and to control meaningful generative processe...     »