New on the arXiv:

“Ousiometrics
and
Telegnomics:

The essence of meaning
conforms to
a two-dimensional powerful-weak and dangerous-safe framework

with

diverse corpora presenting a safety bias”https://t.co/qGikE0l8EK pic.twitter.com/L7AeUuyRuW

— Computational Story Lab (@compstorylab) October 14, 2021

We measure essential meaning using language as the map.

Some definitions:

1. Ousiometrics: The quantitative study of the essential meaningful components of an entity, however perceived. ⁰
2. Telegnomics: The far sensing of knowledge (~ distant reading)

— Computational Story Lab (@compstorylab) October 14, 2021

Fair warning: This is something of a long paper (for us). The conclusion has 12 sections. We know, we know. pic.twitter.com/RKRrOOQBMQ

— Computational Story Lab (@compstorylab) October 14, 2021

We’ll be briefer here. As we would for revenge purposes.

Please see the paper for non-brevity. pic.twitter.com/dLbqqIKKWq

— Computational Story Lab (@compstorylab) October 14, 2021

Okay, enough with the giffery.

Once upon a time, there was a … No, no, no.

The middle of last century saw work emerge on measuring meaning using semantic differentials.

Here’s a page of semantic differentials from Osgood et al.’s “The Measurement of Meaning”, 1957: pic.twitter.com/l2ttC31uoF

— Computational Story Lab (@compstorylab) October 14, 2021

Roughly how a measurement of meaning study might progress:

Raters
(undergraduate students)
are asked to score entities
(often words or phrases but also sounds, images, etc.)
using semantic differentials.

Scoring is mostly done on discrete scales (like Likert). pic.twitter.com/vTlT60ozW4

— Computational Story Lab (@compstorylab) October 14, 2021

So why are we revisiting this work now?

1. In the last 10 years, the scale of studies have moved to scoring more than 10,000 words.

2. Better scoring systems for semantic differentials are now being used (best-worst scaling).

We work with Mohommad’s 2018 NRC VAD lexicon.

— Computational Story Lab (@compstorylab) October 14, 2021

We build ‘ousiograms’ to help us understand essential meaning:

Automatically annotated 2-d histograms for categorical data.

Here’s one for dominance versus valence.

(There’s an askew SVD ellipse in the middle that indicates non-orthogonality.) pic.twitter.com/SQx0XOI2XN

— Computational Story Lab (@compstorylab) October 14, 2021

Main things we find:

1. VAD is indeed far from orthogonal.

2. Reanalysis through SVD uncovers a power-danger-structure framework.

3. Powerful-weak and dangerous-safe span the main compass-like plane of essential meaning.

4. Words used in real corpora have a safety bias.

— Computational Story Lab (@compstorylab) October 14, 2021

We show analytic steps that move through different frameworks, making tableaus of ousiograms (explained in the paper).

These are crucial figures. pic.twitter.com/rrPhDsVwly

— Computational Story Lab (@compstorylab) October 14, 2021

We show analytic steps that move through different frameworks, making tableaus of ousiograms (explained in the paper).

These are crucial figures. pic.twitter.com/rrPhDsVwly

— Computational Story Lab (@compstorylab) October 14, 2021

Here’s the full power-danger ousiogram.

Track the words around the edges and through the middle.

Now this is only for a lexicon—we don’t know yet how words are used across power-danger-structure space in real corpora. pic.twitter.com/6nvRlV1stV

— Computational Story Lab (@compstorylab) October 14, 2021

Types to tokens:

When we look at real corpora and take into account how often words are used, we see evidence of a ‘safety bias’.

Here’s a reflection of English fiction over 120 years from Google Books: pic.twitter.com/7HwYivMhb6

— Computational Story Lab (@compstorylab) October 14, 2021

And it’s the power-danger framework that best fits this revision of the Pollyanna Principle.

We see the ‘safety bias’ across diverse corpora:

- Jane Austen
- Sherlock Homes
- The New York Times
- Wikipedia
- Talk radio
- Twitter pic.twitter.com/MzRF3dyQjf

— Computational Story Lab (@compstorylab) October 14, 2021

There are many other pieces (read the conclusion!) and we’ll add just a few more here:

We made a prototype ousiometer to measure power and danger and other dimensions.

As an example, here are ousiometric time series for Twitter in 2020 and the start of 2021. pic.twitter.com/cW2qjT3XH2

— Computational Story Lab (@compstorylab) October 14, 2021

We focus in on January 6, 2021.

The attack on the Capitol registers most strongly on the danger time series. pic.twitter.com/LBiiYr3Ey4

— Computational Story Lab (@compstorylab) October 14, 2021

We also look at synousionyms and antousionyms, the ousiometric counterparts of synonyms and antonyms.

These help us understand and describe the end points of dimensions (a difficult task that we work hard on in the paper).

Some examples: pic.twitter.com/BgM4rI0tx4

— Computational Story Lab (@compstorylab) October 14, 2021

More things in the paper:

- Russell’s Circumplex Model of Affect
- Personality measures (OCEAN and Cipolla)
- Telegnomics: Measuring stories
- Information theory and ousiometrics
- Power-danger sensing in the context of survival
- Play ~ Unstructure

and

- D&D alignment charts pic.twitter.com/tOitVvpZ6N

— Computational Story Lab (@compstorylab) October 14, 2021

One last thing (for those searching for, or in pursuit of):

Happiness = Power + Safety. pic.twitter.com/5mbsAGsZCS

— Computational Story Lab (@compstorylab) October 14, 2021

(This is also happiness) pic.twitter.com/cdrK8Z6Tes

— Computational Story Lab (@compstorylab) October 14, 2021

Team:
P. S. Dodds, T. Alshaabi, M. I. Fudolig, J. W. Zimmerman, J. Lovato, S. Beaulieu, J. R. Minot, M. V. Arnold, A. J. Reagan, C. M. Danforth

Online appendix (some code, some data, etc.):https://t.co/Kky65HJRlo

And again, the preprint is here:https://t.co/qGikE0l8EK

— Computational Story Lab (@compstorylab) October 14, 2021