“ ¶ § ”

Putting a number on how meaning bends

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)
King & Gentner (2022). Verb metaphoric extension under semantic strain. Cognitive Science.
Role
First author
Published
Cognitive Science · 2022
Scale
3 experiments · 430 participants
Methods
NLP · mixed-effects models
Tools
R · Python · word2vec
Participants read
The motor complained
↓ and wrote what it meant
Their paraphrase
The engine made strange noises
noun: kept close verb: reinterpreted
Real responses from the study
How I work

What this case study shows about how I work

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.

“ At a glance

When a sentence strains, the verb gives way

430 participants

Three experiments and 2,700+ paraphrases, each scored for how much the noun's and the verb's meanings changed.

2×

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.

Rival ruled out

Having more meanings didn't explain it. Words with many dictionary senses changed more overall, but didn't bend any faster under strain.

6×

Verbs changed meaning through metaphor six times as often as nouns (165 vs. 27 coded paraphrases). Nouns shifted to related or broader terms instead.

§ The question

A sentence that can't be literally true. Which word gives?

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.

My hypothesis

Online adjustment

Verbs stretch their meaning on the spot to fit the noun.

The rival explanation

Sense selection (polysemy)

Verbs just have more stored meanings to choose from, so they only look more flexible.

The design: every noun meets every verb Experiment 2 · 36 sentences · hover or tap a cell
Crossing few-meaning and many-meaning words keeps polysemy and strain separable.
Rated strain: leastmost (quartiles) Bold = many meanings (7–15 senses) · regular = few (1–2)
” Approach

Collect interpretations, score them and try to break the result

“

Collect

430 people wrote paraphrases of sentences like “The bell complained”, screened with attention checks and two-coder exclusion rules (κ = .63–.66).

Score

A new word2vec metric measured how much of each original word's meaning survived in the paraphrase.

§

Stress-test

A factorial design against the rival explanation, a human-judged replication and qualitative coding of how words changed.

A new metric

A ruler for meaning change

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.

Meaning retained=cos(original word,whole paraphrase)

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.

Worked example from the paper
The lantern limped → “The candle flickered”
lantern ↔ paraphrase .47
limped ↔ paraphrase .22

The noun stayed close. The verb moved further.

Why it works

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.

Known blind spot

Embeddings blur similarity with association and are less accurate for words with many meanings. That's why Finding 2 checks the metric against people.

↓ The next chart builds one layer at a time as you scroll, the way I'd present it to a room.
“ Finding 1 · Experiment 2

The verb bends, the noun holds, and polysemy isn't the reason

Meaning retained as sentences get stranger
word2vec similarity to the original word · fitted mixed-model lines · 250 people, 1,216 paraphrases
Noun Verb few meanings many meanings 95% CI Predicted if polysemy drove it
Build 1

Nouns hold their meaning

From literal (left) to highly strained (right), noun meaning barely moves. Strain had no significant effect on nouns (p = .65).

Build 2

Verbs bend

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 ⅓
Build 3

Test the obvious explanation

If 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.

Build 4

The lines stay parallel

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
View data
SeriesLeast strainedMost strainedStrain effect
Noun · few meanings0.290.30β = −0.03, p = .65
Noun · many meanings0.270.24
Verb · few meanings0.310.20β = −0.29, p = .001
Verb · many meanings0.260.16

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.

§ Finding 2 · Validation

A human-judged measure reached the same verdict

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.

OriginalThe motor complainedStimulus
ParaphraseThe engine did not work wellExperiment 2 participant
Re-paraphraseThe motor functioned badlyNew Experiment 3 participant
✓motor came back: meaning kept ✕complained didn't: meaning changed
How often the original word came back
Share of 1,012 re-paraphrases · overall: nouns 21%, verbs 10%
Noun Verb solid = few meanings · faded = many
Prediction
word2vec (Exp 1–2)
Human task (Exp 3)
Verbs change more than nouns
✓Confirmed
✓Confirmed 10% vs. 21% came back
Strain drives verb change, not noun change
✓Confirmed
✓Confirmed p = .02 vs. .69
Many-meaning words change more at every level of strain
✓Confirmed
~Verbs yes nouns marginal, p = .06
” Finding 3 · Qualitative coding

Verbs change by metaphor. Nouns swap in a neighbor.

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).

Type of meaning change, by word class
Count of coded paraphrases (576 sentences, noun and verb coded separately)
Verb Noun
Metaphoric · typical of verbs
The queen dried → The monarch aged

An abstract parallel, not a literal one.

Taxonomic · typical of nouns
motor → machine

A broader or narrower category.

Associative · typical of nouns
motor → car

Something closely linked, like a part for its whole.

View data
Change typeVerbsNouns
Metaphoric (analogous)16527
Associative (metonymic)110146
Taxonomic101251

Remaining codes (synonym, describes the situation, other) not shown.

From the lab to people analytics

The same toolkit, pointed at the workforce

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.

In this studyScored 1,700+ free-text responses with word embeddings
In people analytics

Score open-ended engagement, exit-survey and review comments at scale, and track how themes shift between survey waves.

In this studyChecked the NLP metric against an independent, human-judged measure
In people analytics

Validate AI-generated sentiment, theme or attrition-risk scores against human-rated samples before anyone acts on them.

In this studyCrossed polysemy with strain to rule out the rival explanation
In people analytics

Separate composition from real effects: does a pay or attrition gap persist within role, level and tenure, or disappear?

In this studyMixed-effects models with random effects for people and items
In people analytics

Model employees nested in teams and managers, and repeated survey waves, without overstating certainty.

In this studyReplaced a yes/no “strained” label with continuous ratings, and the effect grew (β −.26 → −.38)
In people analytics

Keep performance, engagement and risk measures continuous. Binary flags like “high performer” or “at risk” throw away signal.

One more implication · a hypothesis worth testing

Survey items bend at the verb

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.