Citation-augmented text reuse detection

Abstract

Firstly, I would like to thank my advisor David Smith for all of his guidance and advice over the past seven years. I am so lucky to have been able to learn from him and take advantage of the many opportunities for collaboration beyond Northeastern working with him allowed me to have. Without his patience and willingness to support my own research interests, this would certainly have not been written. David, it has truly been a pleasure being your student.My work, and my PhD experience as a whole, would not be what it is without a great deal of collaboration and help from members of the KITAB team. To single out a few people in particular, I would like to thank Sarah Savant and Kevin Jacques not only for their guidance when I was initially joining the team and was getting up to speed on the very interdisciplinary field I am now a part of, but also for their friendship. I would also like to thank Mathew Barber for his contributions to the final main chapter of the thesis. It was awesome working together on it. I would also like to thank my undergraduate advisors Joyce Madancy and Nick Webb, without whose guidance I likely wouldn't have gone to graduate school in the first place.To my friends, you all know who you are, you made this possible too. Lastly I would like to thank my family. This would have been equally impossible without your unwavering support and I am so lucky to have you all. 1 work focuses on these individual strategies, either at training or test time; as Juba and Le (2019) demonstrate, the effectiveness of strategies to mitigate class imbalance depend heavily on the base rate of the minority class. In this chapter, we explore in a unified framework the tradeoffs in end-to-end performance among different sampling strategies for the training data, data filtering with retrieval at training and test time, and different training methods for the final classifier.