Scalable and Efficient Probabilistic Topic Model Inference for Textual Data
Måns Magnusson · 2018
There are many people that I need to thank for their direct and indirect contributions to this thesis.People who have given their support and personal contributions, and also some that just put up with me through these five, very intensive, years.First and foremost I want to thank my main supervisor Mattias Villani.It has been a privilege to be his student and I really want to thank him for all the ideas, time, and effort he put into me throughout the years.He has always pushed me to go further, accepting nothing less than high quality research from me.But he also helped me focus on the right things when so many exciting research projects were possible.My co-supervisor Marco Kuhlmann has also been important during these years, helping me through the difficulties of Natural language processing and computational linguistics.Marco's advice and counseling has been invaluable to me.I am also very grateful to David Mimno, who welcomed me to Cornell University and acted as my supervisor during the fall 2016.Doing research at Cornell for one semester really helped me to get different perspectives on the latent semantic analysis research field.The way I try to present the different parts of latent semantic analysis in this thesis is heavily influenced by discussions with David and David's course on advanced topic models.The most important part of my graduates studies has been learning to be a researcher.I entered graduate school, knowing very little about how to do statistical research, especially in the field of probabilistic text modeling and natural language processing.But thanks to my many collaborators I now feel like I can actually do real research.My research collaborators on different projects have been extremely important.One of the longest collaborations has been with Leif Jonsson, who taught me all of the hidden knowledge of computer science and programming.Alexander Terenin contributed a lot to my deepened interest in the more theoretical parts of Bayesian inference.David Broman pushed my knowledge on computational concurrence and complexity.With Alexandra Schofield I enjoyed a lot of discussions on topic modeling in general and corpus curation specifically.Alexandra also made me feel very welcome to the Cornell group during my visit to Cornell.Finally, I want to thank my collaborators Richard Öhrvall and Katarina Barrling for the long and interesting discussions on how to combine probabilistic modeling with social science theory.No man is an island.My fellow doctoral students at Linköping University, both at IDA and at other departments have been a great support.I want to give a special thanks to Josef Wilzén, with whom I have shared office and teaching duties for more than five years.I have also had the privilege of spending time with Per Sidén and Sarah Alsaadi, with whom I have had many exciting discussions.The discussions I've had