Measuring meaning sounds like trying to weigh fog. You can point at it, you can feel it, but the moment you grab it, it slips between your fingers. In this talk from the Cultural Analytics Series, Peter Dodds, a professor in the Department of Computer Science at the University of Vermont, starts with the classic “semantic differential” approach, where he rates things on opposite adjective pairs, critiques the modern VAD habit, and then rebuilds a cleaner compass of meaning that scales from single words to entire stories.
The Old Map: Semantic Differentials, EPA, and the Rise of VAD
The talk opens with a familiar classroom ritual: a professor hands students a list of adjective opposites (good–bad, weak–strong, active–passive), students rate concepts on little scales, and then the class compresses all those ratings into a few underlying dimensions. When you collect enough ratings, you can stack the results into a big matrix and use dimensionality reduction (like SVD or PCA) which often collapse into three big themes:
- Evaluation (good vs bad)
- Potency (weak vs strong)
- Activity (active vs passive)
Later, in emotion research, those were repackaged as VAD:
- Valence (pleasant vs unpleasant)
- Arousal (calm vs excited)
- Dominance (in control vs powerless)
Dodds argues something important: these sliders aren’t as independent as we pretend. In real survey data, the three dimensions tend to “lean” into each other. The reason for this isn’t that people are careless; it’s that the labels themselves are a little blurry. For example, when people rate “awful,” they’re not only saying “bad.” They’re often also feeling “threatening,” “intense,” “out of control,” and these emotions get mixed together. So when researchers treat VAD as a perfect 3D coordinate system, the points don’t fill a nice cube. They bunch into a tilted, lopsided shape. So instead of forcing the old labels onto the data, they need to re-check what the data is really doing.
Photo: Professor Dodds presenting his work at the I School
Part 1: Telegnomics: A New Compass For Words
To help the audience follow the rest of the presentation, Dodds begins by defining three key terms that shape his framework:
- Ousiometrics: measuring the essence of meaning
- Telegnomics: “distance-sensing” knowledge, meaning you can learn about attitudes and stories by reading large-scale text, without interviewing everyone directly.
- Ousiograms: tools and charts that show how meaning is distributed and how it moves over time.
Looking again at vast collections of word evaluations, Dodds suggests that arousal, often described as high versus low energy, typically reflects a narrower idea: perceived risk instead of security. Rather than only signaling intensity, terms such as battle, weapon, murder, and nuke group together due to their shared link with harm. Through this perspective, an alternative framework emerges that is built around three axes:
- Power (strong vs weak)
- Danger (threatening vs safe),
- Structure (orderly vs chaotic).
This shift also reframes the Pollyanna principle, the idea that everyday language is biased toward positivity. Dodds suggests it may be less about constant “positivity” and more about a low-danger baseline. Most of the time, we’re using language to coordinate, describe, joke, and make plans, which means safer words naturally dominate. They also highlight a key comparison: counting word types (each unique word once) is like giving one vote to every tree species in a forest, while counting word tokens (how often words appear) is like counting how many trees there actually are. Once real-world spread enters the picture, from novels to tweets, headlines to transcripts, a shape appears repeatedly. For much of our linguistic lives, threats rarely show up and that quiet backdrop slips past basic mood meters.
Photo: Peter Dodds and Tim Tangherlini meet with a group of students and faculty members at AI Futures Lab.
Part 2: Fictional Characters and Archetypometrics
Now comes Part 2, where the data changes entirely. Rather than using word scores, Dodds turns to massive ratings of made-up figures. People evaluated more than two thousand characters across over three hundred narratives, focusing on contradictory features - brave↔cowardly, warm↔cold, and selfish↔selfless. After compressing that immense trait spectrum into just a handful of dominant axes, something clear emerges.
What stands out is how the character types match up with deep story patterns found in the "meaning compass”. The meaning compass describes major character poles like Heroes ↔ Fools, Angels ↔ Demons, and Traditionalists ↔ Adventurers. Surprisingly, these contrasts appear even when the material used to map them has never touched the same sources before. In other words, even though the dataset is totally different, stories keep rebuilding the same building blocks. The names change, but the underlying structure repeats.
Final Thoughts
In the end, we learn from Dodds that it is not merely redefining terms on emotional charts but uncovering a livelier lens for spotting what speech and tales quietly carry underneath. Switching out vague concepts like ‘negative’ for something sharper like danger, patterns start to make sense across different worlds. The results feel less like a measurement project but more like a map of how people navigate life through language. Watch the I School event recording for the full account!
To stay in touch and join these far-ranging conversations of critical cultural importance, please join our Cultural Analytics mailing list by visiting this page or emailing bids-cultural-analytics+subscribe@lists.berkeley.edu.