Computational linguistic models for understanding attitude and behavior

ABSTRACT

Attitudes are often expressed in what people say and write, as well as the content they choose to interact with. With the proliferation of social media and other online content, we are able to understand how people express their attitudes through large-scale linguistic analyses. Further, people's attitudes and behaviors are often intertwined attitude signals can be predictive of future behaviors, and conversely behavioral patterns can reveal underlying attitudes. This thesis explores the development of computational linguistic models to understand attitudes and behaviors. We surface the attitudes that people hold with respect to social roles (e.g., "professor," "mother") and compare them across different cultures using corpus-statistics models and dependency-based embedding models. Next, we look at how personal traits are predictive of behavior. To this end, we explore how we can incorporate implicit world knowledge into language models by predicting attitudes towards charitable giving. In this same direction, we examine traits, as expressed on social media, that are indicative of people likely to persist in pursuing self-improvement. We leverage linguistic characteristics such as expressed affect, writing style, and latent topics. Finally, we gain insight into how attitude and behavior give insight to each other by predicting attitudes towards philanthropic causes based on engagement behavior with newsletters and personal background information, using text-aware graph representation models. We also show how behavioral traits present in online communities are predictive of resilient attitude during the COVID-19 pandemic. (PsycInfo Database Record (c) 2022 APA, all rights reserved)