“Thick data” captures the qualitative richness of human experience including emotions, behaviors, and perceptions, and while it is central to the design process, synthesizing it risks losing nuance as researchers cluster experiences into discrete themes.

This data is often identified as unquantifiable, but through modern natural language processing (NLP) techniques, the bridge between qualitative and quantitative is shortened. This research introduces a methodology for identifying sentiment between speakers in an interview study. The process includes (1) partitioning a sentence sequence into subsequences representing each speaker (2) embedding sentences into vectors representing sentiment polarity using a BERT-based embedding model architecture, and (3) softmax normalization to produce probability distributions used for visualization.

This process is then expressed through new visualizations and explores various visual encoding characteristics to create effective visuals, defined by three principles: consistency, accuracy, and meaningfulness. The goal is to provide a new synthesis methodology for real-world research operations.