10/ We find our measures diverge from related psycholinguistic norms (concreteness and imageability), but this divergence is largely due to our measure's informativity dimension.

9/ While we expect lexical classes to be grounded, we find functional classes—traditionally viewed as “grammatical” or “abstract”—also carry semantic content.

For example, determiners like "der" or "une" still contribute to meaning, challenging common assumptions in linguistics.

8/ Across 30 typologically diverse languages, we find a cline between Nouns > Adjectives > Verbs.

This corroborates ideas from cognitive linguistics that suggest these classes lie in a continuum.

7/ To validate our measure we look at the lexical-functional distinction in word classes:

Lexical classes: nouns, verbs, adjectives—contentful words.
Functional classes: prepositions, determiners—“grammatical” words.

How universal is this distinction? Is there a clear line?

6/ Groundedness turns out to be the *decrease in surprisal* of a word when we see the image it refers to! But ensuring comparability is tricky (see paper for details).

5/ Our use of images makes groundedness straightforward to compute. We need only the log probabilities from:
- a language model p(word | context
- an image captioning model p(word | context, meaning)

3/ / To align function across languages, we use image captions. The linguistic content of a caption aims to express the contents of an image. An image represents the state of the world in a language-neutral way, and so can serve as an imperfect proxy for meaning.

NEW PREPRINT!

Language is not just a formal system—it connects words to the world. But how do we measure this connection in a cross-linguistic, quantitative way?

🧵 Using multimodal models, we introduce a new approach: groundedness ⬇️