Building an automated text analysis metric that scales

The existing methodology for open-ended text analysis used human raters and was slow, costly, and unreliable. I developed and validated a novel metric using NLP techniques that automated and scaled the assessment process.

Read the full paper (PDF) (opens in a new tab)
King & Gentner (2022). Verb metaphoric extension under semantic strain. Cognitive Science, 46, e13141.
Role
First author
Metric
word2vec cosine similarity
Validation
5 checks · 3 experiments
Scale
430 participants · 2,700+ paraphrases
Tools
R · Python · word2vec
Key strengths

What this case study shows about how I work

Adapted from King & Gentner (2022), Cognitive Science, co-authored with Dedre Gentner. All statistics are from the published paper. Fitted lines in the validation chart are redrawn from its figures and are approximate. For the language findings themselves, see the companion case study.

At a glance

A new measure, tested five ways

1,742 paraphrases

Scored automatically across two experiments, with no splitting and no human raters needed to score them.

3 of 3

An independent human benchmark reproduced every pattern the metric found, one of them marginally.

13 → 69

Blind-coded metaphoric verb changes rose steadily across the metric's score quartiles, from least to most changed.

Length-proof

Paraphrase length didn't track strain (p = .82 and .49), and controlling for it left the estimates nearly unchanged.

The measurement problem

How do you measure how far a word's meaning moved?

The research needed to know how much the noun and the verb each changed when people reinterpreted odd sentences. But paraphrases blend the two, and there's no ground truth to check against.

An example from the paper
The car limped→The badly functioning vehicle struggled to drive
from the nounfrom the verbfrom both? Judges couldn't agree

Which part came from car and which from limped? Every earlier method depended on human judges answering questions like this.

Three human-rater approaches, the same problems

Divide and rateRetraceDouble paraphrase
✕SlowTime-consuming, labor-intensive judging
✕ExpensiveNew judges or participants for every study
✕Judges can't agreeSplitting a paraphrase is a judgment call
✕Low reliabilitySubjective ratings vary from judge to judge
✕Doesn't scaleEffort and cost grow with every response
✕Loses most of the dataIn one method, only 4% of verbs resurfaced
So the new metric had to be
✓No splitting ✓Graded, not yes/no ✓Every response scored ✓No judgment calls ✓Scalable and replicable
Definition and construction

Meaning retained: how close the paraphrase stays to the original word

Word embedding models from NLP offered a solution to the human-rater problem: reliable estimates of meaning change that can be calculated instantly, scaled to as many datapoints as necessary, for free.

1

I developed a method for producing noun- and verb-change scores for each participant's response using cosine similarity as a metric.

2

Lower scores = more meaning change

3

If the verb-change score is lower than the noun-change score, then the verb changed more

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

The verb's score (.22) is lower than the noun's (.47): the verb changed meaning more.

Validation

Five ways I tried to break the metric

A new metric earns trust by passing tests it could have failed. Each check below uses evidence that doesn't depend on the embeddings themselves.

1Reproduces a known human-judged resultReplicationSame verb-vs-noun pattern human judges found in 1988Passed
2Responds to a finer-grained inputSensitivityEffect grows when strain is rated, not labeled yes/noPassed
3Agrees with an independent human measureConvergent validityHuman benchmark reproduced all three patternsSee below ↓Passed
4Low scores mean what they shouldConstruct validityBlind coding: bigger scored change, more metaphorPassed
5Isn't driven by a nuisance variable or exclusionsConfounds & robustnessLength and data exclusions ruled outPassed
3Convergent validity · Experiment 3 · 71 people, 1,012 re-paraphrases

Does an independent human measure agree?

Embeddings have known blind spots, so I re-measured Experiment 2 with no algorithm at all. New people paraphrased the paraphrases, and I counted how often the original words came back: lossy, but objective.

Same data, two instruments
Fitted lines as strain rises · left: the metric (Exp 2) · right: human benchmark (Exp 3) · each panel on its own scale
Noun Verb few meanings many meanings
Build 1

What the metric says

As sentences get stranger, verb similarity falls (p = .001) while noun similarity stays flat (p = .65).

Build 2

What people say, with no algorithm

In the human task, verbs came back less often as strain rose (p = .02). Nouns didn't move (p = .69). Different instrument, same shape.

Nouns came back 2× as often: 21% vs. 10%
Build 3

Even where embeddings are weakest

Embeddings are known to be less accurate for words with many meanings. Humans showed the same polysemy effect: many-meaning words changed more in both (humans: verbs p < .01, nouns marginal at p = .06).

3 of 3 patterns reproduced
View data
EffectMetric (Exp 2)Human benchmark (Exp 3)
Verbs change more than nounsβ = 0.04, p = .01β = 1.24 (logit), p < .001
Strain → verb changeβ = −0.29, p = .001β = −0.25, p = .02
Strain → noun changeβ = −0.03, p = .65β = −0.02, p = .69
Polysemy → verb changeβ = −0.22, p = .01β = −0.63, p < .01
Polysemy → noun changeβ = −0.16, p = .01β = −0.20, p = .06

Human resurfacing rates: nouns 24.2% (few meanings), 18.0% (many); verbs 13.9% (few), 6.9% (many). Fitted lines are read from the paper's Figs. 4 and 5 and are approximate.

Applied benefits

More reliable, more scalable and far cheaper than human raters

Once validated, the metric changed what the research could afford to measure.

Reliability
Same response, same score

Scoring is deterministic, so there's no judge-to-judge disagreement. Even the human coders in this study started at only moderate agreement (κ = .58–.66).

Scalability
Every response, instantly

1,742 paraphrases scored with no raters and none lost to the method. One human method got a signal from as few as 4% of verb responses.

Cost
Free to run

The pretrained vectors cost nothing to use. Re-measuring one experiment with the human method took 77 paid online participants.

For what the metric revealed about language, see the findings case study
Where it applies

Turning people's words into trustworthy data

The same approach, scoring text with embeddings and then validating the scores against people, carries over wherever open-ended responses need to become numbers.

People analyticsUX research

Checking how survey items are understood

Ask respondents to restate an item in their own words, then score how far their reading drifts from the intended meaning: cognitive pretesting at scale.

People analytics

Open-text comments at scale

Score engagement, pulse and exit-survey comments against theme descriptions instead of hand-coding thousands of responses.

UX research

Mental-model alignment

Compare how users describe a feature or task with the product's own language to find where understanding breaks down.

People analyticsUX research

Tracking change over time

Measure how the language around a topic, like career growth, onboarding or a new feature, shifts between survey waves or releases.

The common thread

Whether the scorer is an embedding model or an LLM, the validation steps here (replicate a known result, benchmark against human judgment, rule out confounds) are what make a new measure safe to act on.