| The main focus of this dissertation is new and improved ways of bringing high quality content to the users by leveraging the power of machine learning. Starting with a large amount of data we want to condense it into an easily digestible form by removing redundant and irrelevant parts and retaining only important information that is of interest to the user. Learning how to perform this from data allows us to use more complex models that better capture the notion of good content.;Starting with supervised learning, this thesis proposes using structured prediction in conjunction with support vector machines to learn how to produce extractive summaries of textual documents. Representing summaries as a multivariate objects allows for modeling the dependencies between the summary components. An efficient approach to learning and predicting summaries is still possible by using a submodular objective/scoring function despite complex output space.;The discussed approach can also be adapted to unsupervised setting and used to condense information in novel ways while retaining the same efficient submodular framework. Incorporating temporal dimension into summarization objective lead to a new way of visualizing flow of ideas and identifying novel contributions in a time-stamped corpus, which in turn help users gain a high level insight into evolution of it.;Lastly, instead of trying to explicitly define an automated function used to condense information, one can leverage crowdsourcing. In particular, this thesis considers user feedback on online user-generated content to construct and improve content rankings. An analysis of a real-world dataset is presented and results suggest more accurate models of actual user voting patterns. Based on this new knowledge, an improved content ranking algorithm is proposed that delivers good content to the users in a shorter timeframe. |
