Topic Modeling in Theory and Practice
Johns Hopkins University
Topic models can decompose a large corpus of text into a relatively small set of interpretable themes or topics, potentially enabling a domain expert to explore and analyze a corpus more efficiently. However, in my work, I have found that theories put forth by topic modeling research are not always borne out in practice. In this dissertation, I use case studies to explore four theories of topic modeling. While these theories are not explicitly stated, I show that they are communicated implicitly, some within an individual study and others more diffusely. I show that this implicit knowledge fails to hold in practice in the settings I consider. While my work is confined to topic modeling research and moreover concentrated on the latent Dirichlet allocation topic model, I argue that these kinds of gaps may pervade scientific research and present an obstacle to improving the diversity of the research community.
natural language processing, machine learning, artificial intelligence, topic modeling, reproducibility