Aligning Multidimensional Worldviews and Discovering Ideological Differences

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 4832–4845 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 4832 Aligning Multidimensional Worldviews and Discovering Ideological Differences Jeremiah Milbauer Language Technologies Institute, Carnegie Mellon University jmilbaue@cs.cmu.edu Adarsh Mathew Knowledge Lab, University of Chicago adarshm@uchicago.edu James Evans Knowledge Lab / Sociology, University of Chicago jevans@uchicago.edu Abstract The Internet is home to thousands of com- munities, each with their own unique world- view and associated ideological differences. With new communities constantly emerging and serving as ideological birthplaces, battle- grounds, and bunkers, it is critical to develop a framework for understanding worldviews and ideological distinction. Most existing work, however, takes a predetermined view based on political polarization: the “right vs. left” dichotomy of U.S. politics. In reality, both political polarization – and worldviews more broadly – transcend one-dimensional differ- ence, and deserve a more complete analysis. Extending the ability of word embedding mod- els to capture the semantic and cultural char- acteristics of their training corpora, we pro- pose a novel method for discovering the mul- tifaceted ideological and worldview character- istics of communities. Using over 1B com- ments collected from the largest communities on Reddit.com representing 40% of Reddit ac- tivity, we demonstrate the efficacy of this ap- proach to uncover complex ideological differ- ences across multiple axes of polarization. 1 Introduction and Motivation “The limits of my language mean the limits of my world” Tractatus Logico-Philosophicus, 1921, Ludwig Wittgenstein Media choice, social networking platforms, and collaborative filtering on the internet have enabled individuals to enter “echo chambers” that reflect shared worldviews (Sunstein, 2018; Mutz, 2006; Bishop, 2009). The internet also publicly reveals these communities and their communication for analysts of language, culture and interaction at un- precedented scale. Despite the abundance of such data, however, analysis of worldviews and ideolog- ical difference has been dominated by considera- tions of “polarization” (Boxell et al., 2017; Bail politics candidate trump corrupt Community 1 clinton trump Alignment candidate clinton politics corrupt Community 2 Figure 1: By training the model to align “candi- date”, “politics”, and “corrupt,” a hypothesis alignment f(“trump”|C1) ≈ f(“clinton”|C2) emerges. et al., 2018), which impoverishes the comparison of ideologies by reducing them to pairs separated along a singular dimension. Here, we draw inspiration from the approach of interpretive anthropology and the focus of cog- nitive anthropology to represent, investigate and compare worldviews from community discourse. In the Interpretation of Cultures, Geertz rendered culture as “a system of inherited conceptions ex- pressed in symbolic forms by means of which men communicate, perpetuate, and develop their knowl- edge about and attitudes toward life” (Geertz et al., 1973). Combined with cognitive anthropology’s concern with how implicit knowledge changes the way people perceive and relate to the world (d’Andrade, 1995), this motivates assessment of worldviews through modern pre-trained natural lan- guage models that render words (Mikolov et al., 2013b; Pennington et al., 2014) and phrases (De- vlin et al., 2019; Radford et al., 2019) in relation to one another as a function …

Content truncated to 500 words