Psychological states measured from social media text by projecting words
Measuring Psychological Constructs from Social Media Text Using the Word Embedding Projection Approach
Social and Information Networks
Summary
Measuring people's thoughts and feelings through their social media posts can be tricky because the methods need to be clear and consistent over time. The authors developed a new way called Word Embedding Projection Approach (WEPA) that uses special word maps and important example words to measure psychological ideas in social media language. They tested this method on a large Chinese fitness social media platform and found it matched well with human ratings and could predict future exercise behavior. This approach covers more text and works well over time, offering a new tool to track psychological trends in social media posts.
What this means in practice
- •For social media analysts: Measure psychological traits and changes from user posts in specific social media domains using interpretable word projection methods.
- •For health app developers: Integrate domain-specific psychological measurements from user text to better predict and support exercise habits and motivation.
Authors
Xudong Deng, Yijun Li
Abstract
Social media text offers new opportunities for tracking psychological constructs over time, but such measurement requires methods that are interpretable and stable in longitudinal settings. Grounded in Conceptual Space Theory and semantic projection, this study develops the Word Embedding Projection Approach (WEPA), which represents psychological constructs as semantic axes defined by theory-driven anchor words in a domain-specific word embedding space. WEPA scores are interpreted as indicators of expressed psychological salience in user-generated text. Using panel data from Keep, a Chinese fitness-oriented social media platform (177,829 users; 2,668,298 weekly observations), we evaluate WEPA across four studies. Study I examines the measurement foundation through anchor-word validation, human-coded benchmarks, and cross-temporal semantic-axis stability checks. Studies II and III show theoretically interpretable predictive associations between WEPA-derived scores and subsequent exercise duration for goal-setting constructs and the four source dimensions of self-efficacy. Study IV provides exploratory relative-week trajectories of aggregate construct dynamics. Overall, WEPA shows strong agreement with human-coded benchmarks for goal specificity and physiological states and achieves broader text coverage than a dictionary-based baseline. These findings suggest that WEPA offers a promising approach to theory-driven psychological measurement from domain-specific social media text.