Master thesis: Transfer Learning for Text Classification with Convolutional Neural Networks (Glavaš, Ponzetto)

Convolutional neural networks have been shown to be very successful to various text classification tasks. The main shortcoming of CNNs used for text classification is that they, like most neural models, require a large number of annotated instances (i.e., training examples) in order to achieve solid classification performance. The goal of this thesis would be to explore and experiment with several transfer learning techniques at different network layers that would allow for a smaller number of examples in each particular classification task. Transfer learning means that some of the parameters trained on one datasets can be set as initial parameter values for the CNN trained for another classification task. The underlying assumption is that the early layer parameters (e.g., such as semantic vectors of words as input) of the CNN are general and transferable across domains. The thesis should includes the development of custom CNNs, transfer learning implementation and proper evaluation. All of these steps should be extensively described and documented in the thesis itself.