Mining for Meaning: using computational text analysis for social inquiry

People interpret their surroundings through associations, determining what they perceive as belonging or not belonging together. For instance, one individual may view immigrants as a beneficial addition to the domestic labor market, while another may perceive them as a threat to job opportunities for native citizens. Despite differing viewpoints on immigration, these individuals share a similar economic interpretation of immigration as a concept. Explaining how these interpretations develop and evolve is a fundamental and open question related to the social world. For a long time, people’s interpretations of the world have been hidden away in their minds, and researchers have primarily relied on surveys to try to measure them. However, individuals and groups leave behind traces of their understandings of the world in their communication and written expressions. Consequently, textual data hold immense potential for sociological research. This thesis pursues three primary objectives. First, to discuss the use of text data for social inquiry. Second, to introduce and explore intrinsically interpretable text models for sociological inquiry. Third, to explore rigorous ways of studying meaning and meaning-making in the Swedish immigration discourse using computational text analysis. The introductory chapter and four research articles presented in this thesis all speak to at least one of these aims. Essay I addresses the question of how researchers can assess the data quality of a corpus to determine its suitability for addressing research questions. Drawing inspiration from survey research, this essay presents a general approach to evaluating the scientific value of a given text dataset. The framework outlined in this essay delineates potential errors that could affect the reliability and validity of any measures derived from a corpus, and offers methods for quantifying some of them. Essay II presents a novel extension to standard word embedding models. Our extension gives researchers the ability to study how the meaning of words relates to pre-specified binary dimensions. We find that our proposed intrinsically interpretable model outperforms current standard approaches on classification tasks related to sentiment and gender. The methodology presented in Essay II will thus help sociologists to measure and test theories pertaining to binary concepts. Essay III contributes to the ongoing discussions in sociology regarding the identification of more formal ways to