How do people make sense of a sentence like “The car agreed”? I built an NLP metric for meaning change, validated it against human judgment and designed out the obvious alternative explanation. The answer: people keep the noun and bend the verb.
Read the full paper (opens in a new tab)Adapted from King & Gentner (2022), Cognitive Science, co-authored with Dedre Gentner. Charts are redrawn from the published results; the fitted lines in Finding 1 are read from the paper's figure and are approximate.
Three experiments and 2,700+ paraphrases, each scored for how much the noun's and the verb's meanings changed.
Nouns survived a second round of paraphrasing twice as often as verbs (21% vs. 10%). The verb is the part of the sentence that changes.
Having more meanings didn't explain it. Words with many dictionary senses changed more overall, but didn't bend any faster under strain.
Verbs changed meaning through metaphor six times as often as nouns (165 vs. 27 coded paraphrases). Nouns shifted to related or broader terms instead.
To make sense of “The car agreed”, a reader has to reinterpret the noun, the verb or both. Earlier work pointed to the verb, but its verbs also had more meanings than its nouns, so it couldn't rule out a simpler explanation.
Verbs stretch their meaning on the spot to fit the noun.
Verbs just have more stored meanings to choose from, so they only look more flexible.
430 people wrote paraphrases of sentences like “The bell complained”, screened with attention checks and two-coder exclusion rules (κ = .63–.66).
A new word2vec metric measured how much of each original word's meaning survived in the paraphrase.
A factorial design against the rival explanation, a human-judged replication and qualitative coding of how words changed.
Earlier methods had judges split each paraphrase into noun and verb parts. That was slow, judges often disagreed, and the most hands-off version lost most of the data. I scored the whole paraphrase at once instead.
How close does the paraphrase sit to the original word in word2vec's 300-dimensional meaning space? Lower means the word's meaning changed more.
The noun stayed close. The verb moved further.
No splitting, and nothing lost to scoring. Comparing each word with the whole paraphrase sidesteps the step that tripped up human judges, and it scales to thousands of responses.
Embeddings blur similarity with association and are less accurate for words with many meanings. That's why Finding 2 checks the metric against people.
From literal (left) to highly strained (right), noun meaning barely moves. Strain had no significant effect on nouns (p = .65).
The stranger the sentence, the further the verb drifts from its original meaning (p = .001). The verb is where readers resolve the conflict.
Verb similarity falls about ⅓If verbs only look flexible because they have more stored meanings, words with many meanings should bend faster under strain, nouns included. The lines would fan out.
Many-meaning words sit lower overall, so some sense selection does happen. But they don't bend any faster (interaction p = .78), and nouns stay flat either way. On-the-spot adjustment drives the effect.
Rival explanation ruled out| Series | Least strained | Most strained | Strain effect |
|---|---|---|---|
| Noun · few meanings | 0.29 | 0.30 | β = −0.03, p = .65 |
| Noun · many meanings | 0.27 | 0.24 | |
| Verb · few meanings | 0.31 | 0.20 | β = −0.29, p = .001 |
| Verb · many meanings | 0.26 | 0.16 |
Fitted values read from the paper's figure (approximate). Polysemy main effect: verbs β = −0.22, nouns β = −0.16 (both p = .01). Polysemy × strain: verbs p = .78, nouns p = .40.
To check the metric, I measured Experiment 2 again with no algorithm at all. 71 new people paraphrased the paraphrases, and I counted how often the original words came back.
Beyond how much words change, I looked at how. Two coders classified the changes in 576 paraphrases using a seven-code scheme, then resolved every disagreement (initial κ = .58).
An abstract parallel, not a literal one.
A broader or narrower category.
Something closely linked, like a part for its whole.
| Change type | Verbs | Nouns |
|---|---|---|
| Metaphoric (analogous) | 165 | 27 |
| Associative (metonymic) | 110 | 146 |
| Taxonomic | 101 | 251 |
Remaining codes (synonym, describes the situation, other) not shown.
The subject was language, but the problems are the ones people analytics teams face: open text, new metrics that need validating, confounds and nested data.
Score open-ended engagement, exit-survey and review comments at scale, and track how themes shift between survey waves.
Validate AI-generated sentiment, theme or attrition-risk scores against human-rated samples before anyone acts on them.
Separate composition from real effects: does a pay or attrition gap persist within role, level and tenure, or disappear?
Model employees nested in teams and managers, and repeated survey waves, without overstating certainty.
Keep performance, engagement and risk measures continuous. Binary flags like “high performer” or “at risk” throw away signal.
When wording is ambiguous, people keep the noun and reinterpret the verb. Survey items and competency models are full of metaphorical verbs: drive results, own outcomes, champion growth. I'd expect respondents' readings to diverge most on those verbs, so they're the wording I'd pretest first.