Distant supervision and knowledge transfer for domain-oriented text classification in online social networks

Social networks are known to reflect users preferences and traits, for instance, in the form of posts and comments. This data can be used to solve various user profiling tasks that involve classifiers construction. State-of-the-art NLP models, despite their superior performance, still require manually labelled train and test datasets of significant, however, reduced size. The labelling with the help of user assessors may be impeded by the task complexity or lack of financial or time resources. Thus, we address an alternative distant supervision approach which helps to relatively easy acquire data, while producing noisy observations and imbalanced classes. In this work, we propose an approach to diminish the downsides of distant supervision by introducing how the subsets should be formed and how to optimize the training dataset structure.