Computational Linguistic Models for Personal Values and Human Activities

Personal values are theorized to influence thought and decision making patterns, which often manifest themselves in the things that people say and do. We explore the degree to which we can employ computational models to infer people's values from the text that they write and the everyday activities that they perform. In addition to investigating how personal values are expressed in language, we use natural language processing methods to automatically discover relationships between a person's values, behaviors, and cultural background. To this end, we show that the automatic analysis of less constrained, open-ended essay questions leads to a model of personal values that is more strongly connected to behaviors than traditional forced-choice value surveys, and that cultural background has a significant influence these connections. To help measure personal values in textual data, we use a novel crowd-powered sorting algorithm to construct a hierarchical lexicon of words and phrases related to human values. Additionally, we develop semantic representations of human activities that capture a variety of useful dimensions such the motivation for which they are typically done. We leverage these representations to build deep neural models that are able to make predictions about a person's activities based on their observed linguistic patterns and inferred values.