Subjectivity Is All You Get: Automatic Extraction and Application of Population Subjective Views

Abstract

People’s subjective views play an important role in complex systems. For instance, during COVID, people’s political views often guided their behaviors regarding masking and social distancing, which, in turn, impacted disease spread. In other words, people’s beliefs in this case lead to real world consequences. Including this type of relations into scientific modeling is a non-trivial task, which requires expensive manual data collection and analysis. In the three papers that comprise this dissertation, we take first steps towards automating this process. In the first paper, we define and provide an initial analysis for a new task—belief-consequence pair extraction—which we see as a special case of causal relation extraction, with the cause being subjective. We then focus on the first component of the belief-consequence link: beliefs, or subjective views, as expressed through natural language. In the second paper, we analyze and model subjective views through data collection and model training for the task of identifying sentences containing beliefs in natural language texts. We then further the analysis through the application of extracted beliefs to a downstream task in the third paper. In addition to analyzing subjective views themselves, we provide a discussion on subjectivity that occurs in other aspects of the work, particularly in data annotation and system output evaluation.