Supervised Topic Models with Multiple Annotators

Abstract

We live in an era where information over ows. Yet, for this information to become knowledge, it has to be given meaning. This thesis focuses on a machine learning approach that evolved from probabilistic graphical models, which automatically extracts knowledge from vast amounts of data by assigning themes to documents: topic modeling. Topic models are an emergent technique used for both descriptive and predictive tasks. As a result, it was soon extended to other goals that do not only model topics, but also target variables. This work presents a supervised topic model that is able to learn from crowds. That is, we consider the case where the label set of the data was provided by multiple annotators. In the multi-annotator setting, the ground truth labels need to be modeled from several noisy versions of them given by the di erent annotators. To address this sort of problems, it is often assumed that all labelers are equally reliable through the use of voting techniques, which was proven to be an unrealistic conjecture. On the contrary, the proposed model takes into account the di erent levels of expertise and biases of annotators, by jointly modeling them together with the topics and the true labels. In order to make this process computationally tractable, a variational inference algorithm was developed, which provides an e cient approximate inference method. We nalize by showing how general supervised topic models can be used to predict demand in special events by correlating internet search query data with real measurements of transport usage, thus, motivating the usage of the topic models in real-world applications.