Logo: to the web site of Uppsala University

This dissertation aims to enable the reliable Bayesian inference from text data, with the specific focus on word embedding models and data quality. The first topic is investigated through probabilistic word embedding models, with particular emphasis on identifiability, posterior uncertainty, and prior specification. Paper I introduces a new family of informative prior distributions for probabilistic word embeddings, while Paper II develops a novel Gibbs sampler and presents a solution for the non-identifiability of the embedding posterior. Furthermore, the thesis develops and analyzes methods for the creation, curation, and preprocessing of large-scale text data, which are commonly required for word embedding models. Paper III presents a general framework for this task and establishes theoretical guarantees for error reduction, and Paper IV provides a case study applying the framework to a parliamentary dataset. By addressing both model-level and data-level challenges, the thesis aims to ensure that statistical analyses of large and potentially noisy text data yield valid and trustworthy conclusions.