Essence of meaning
« previous post | next post »
Below is a guest post by Robert Shackleton:
Peter Dodds and coauthors have recently published research that proposes a significant shift in the essence-of-meaning framework, which traces its lineage back to Charles Osgood’s initial efforts to use dimension reduction to quantify human meaning. The paper, “Ousiometrics: The essence of meaning aligns with a power-danger-structure framework instead of valence-arousal-dominance,” appeared in Science Advances. The abstract:
From work emerging through the middle of the 20th century, the essence of meaning has become widely accepted as being described by the three orthogonal dimensions of valence, arousal, and dominance. These essential dimensions have become the cornerstone of sentiment analysis across many fields. By reexamining first types and then tokens for the English language, and through the use of automatically annotated histograms—“ousiograms”—we find here that the essence of meaning conveyed by words is instead best described by a goodness-power-aggression-danger-structure (GPADS) circumplex framework; that large-scale English language corpora reveal a systematic bias toward safe, low-danger words; and that the power-danger-structure framework is the minimal framework that represents essential meaning. We find remarkable congruences between the GPADS framework and other spaces including mental states and fictional archetypes, and we construct and demonstrate a prototype ousiometer.
Extensions of Osgood’s original work led to a shift in the 1970s from his foundational dimensions of evaluation, potency, and activity (EPA) to a VAD framework that substitutes valence (or pleasure) for evaluation, arousal for activity, and dominance for potency. Dodds et al. discuss problems with the data, methods, and results in that previous work and offer significant improvements. Their main conclusion is that essence of meaning categories are more accurately characterizd by a five-dimensional framework involving goodness, power, aggression, danger, and structure (GPADS), but they also provide a three-dimensional PDS “minimal framework” emphasizing power, dominance, and structure.
A power-danger-structure framework for essential meaning seems to fit rather nicely into Lakoff and Narayanan’s model of cognitive schemas and frames, and it might also have intuitive appeal from an evolutionary standpoint. Members of a primate group inevitably have a profound interest both in potential dangers and in patterns of social dominance. But I wonder whether Dodds et al.'s third dimension, structured versus unstructured, might be better framed as animacy-inanimacy. Animacy appears to be at least as consistent with their analogy with thermodynamics as structure does, and early humans may well have interpreted much of the world through a frame of animacy, as evidenced not only by animacy as a fundamental grammatical category in many languages but also in the persistence of belief in non-human agency in most cultures, including ours.
Above is a guest post by Robert Shackleton.
Note that early work on learning semantic distances by projecting words into a meaning space based on orthogonalizing a term-by-document matric, e.g. latent semantic anaysis, was inspired by Osgood's "Semantic differential" method as well as by Gerard Salton's "vector space model".
AntC said,
June 2, 2026 @ 9:41 pm
Oh dear. So the paper is generalising from English to "human language". Isn't Linguistics rather over that sort of cultural imperialism? I'm already inclined to think the exercise is worthless unless validated against languages from dramatically different cultural milieu.
The emphasis on "word" worries me: "(each word is of equal importance)". Since the dataset is English, I see no discussion of our famous propensity for homonyms — which can get so extreme the same lexeme can carry almost opposite meanings.
"perceived meaning in isolation" is no meaning at all. On a quick skim of the illustrated plots, homonyms seem to be avoided. I see no commentary on why/how they've been excluded. (Since sailors and golf were mentioned, the plots don't include 'starboard' but neither 'right'; there is 'weaponry' but not 'club', nor 'society'.
The use of lexical-based dimension analysis that I'm familiar with is the Five Factor Model of personality traits
I'd have to say this has never resonated with me, as evidenced by my never being able to remember all five. (At different times, I can remember four, usually a different four each time.) I guess the explanation may be I'm not a professional psyche-anything. OTOH those five key words seem to be a linguistic hodge-podge: from different registers, variously made-up, abstruse Greek/Latin, everyday (Germanic) words with abstract nominalising suffixes stuck on awkwardly.
I'm not sure the GPADS choices fare much better. 'Goodness' sticks out like a sore thumb as Old English/Germanic, compared to the others being (Norman) French/(low/late) Latin. Why not (say) 'Virtue'?
Seeing as both systems happen to have five dimensions, is there any work on matching the GPADS to the Five Factors? If not, I'm left wondering whether either (or both) ways of carving up the semantic space are really more than arbitrary. ("Russell’s circumplex model of emotion" isn't it.)
I could ask one of those nit-picky Philosophy-of-Science questions: are any of these frameworks falsifiable?
Jonathan Smith said,
June 2, 2026 @ 11:20 pm
So devise a descriptive — here coordinate-system-esque — framework, find it to kinda "work", then confuse your descriptive labels with something constitutive or elemental? And re: "the essence of meaning has become widely accepted as being described by the three orthogonal dimensions of valence, arousal, and dominance" — not around here, but small town.
Kenny Easwaran said,
June 3, 2026 @ 11:33 am
What is at stake in calling some particular small number of dimensions the "essence" of meaning? I don't expect that anyone is claiming that this small number of dimensions is sufficient to distinguish the meanings of most words – Large Language Models need hundreds of dimensions in order to string together words in ways that people find meaningful.
Is the claim that some different set of dimensions is better for automated sentiment analysis than the set of dimensions that have historically been used? Why think this is the "essence" of meaning?
Peter Grubtal said,
June 3, 2026 @ 12:25 pm
Trying to convince myself that this isn't a sokal or an attempt at the Ig Nobel Prize, I googled "ousiometrics" to find that it seems to be a nonce or an ad hoc word, apparently only coined in connection with this paper.
It all seems like metaphysics, grasping at will-o-the wisps.
Brett said,
June 3, 2026 @ 1:55 pm
It's interesting that this does not appear to be particularly new work. The first preprint version of this paper is over four and a half years old. In the past, I normally would have pointed out that Peter Dodds is a really smart guy, but the computational linguistics work coming out of his Computational Story Lab at the University of Vermont seems to be getting increasingly peculiar.
@Peter Grubtal: The abstract for the preprint linked above starts with:
In the subsequent journal version, these definitions are moved to the body of the paper. Near the end, they also suggest (but do not yet have a rigorous definition for) another term:
Julian said,
June 3, 2026 @ 7:10 pm
@peter grubtal
"Sokal…."
Thank you for clarifying the analogy that I was subconsciously reaching for.