Cody Arlie ReedResearch + data
WorkCan two descriptions measure occupational knowledge?

Natural language processing · measurement

Can two descriptions measure occupational knowledge?

ProblemA one-word job title and a detailed account do not provide the same evidence, even when both describe similar work.

What I builtI combined corpus-frequency and occupation-specific weights with GloVe word vectors, used an unnormalized dot product to retain shared descriptive detail, and standardized scores within occupational class.

OwnershipSole researcher and developer: index design, text processing, R and Python implementation, modeling, visualization, and writing

Result

In 60 synthetic pairs, detailed aligned descriptions scored above both similarly detailed descriptions of different work and short aligned descriptions.

Synthetic illustration showing an occupational-description similarity index and 60 artificial pairs: aligned detailed descriptions score above similarly detailed descriptions of different work and short aligned descriptions
Synthetic illustration of an occupational-description similarity index

Research question

Does occupational knowledge relate to class reproduction?

I developed the index to compare a young person’s description of a parent’s job with the parent’s own description. I then used survey-weighted multinomial models to study whether that knowledge was associated with later occupational destinations. I presented the study at the 2024 Population Association of America Annual Meeting.

Index

Measure shared meaning and shared descriptive detail.

After normalizing each GloVe word vector to unit length, I use corpus-frequency weights to reduce the influence of common terms and occupation-specific weights to emphasize words concentrated within an occupational class. I sum the weighted vectors for each description, take their unnormalized dot product, and standardize the scores within occupational class.

01Description vector
ed,k=∑w∈dgwcw,kvw‖vw‖2
02Raw paired score
sa,b,k=ea,kTeb,k
03Within-class comparison
za,b,k=sa,b,k−μkσk
gw
corpus-frequency penalty
cw,k
occupation-specific term weight
μk, σk
within-class mean and standard deviation

The dot product retains description-vector magnitude, allowing semantically aligned detail to carry more weight than a matching one-word title.

Synthetic illustration

Aligned detail produces the highest scores.

I generated 60 artificial pairs across four invented job families. Detailed aligned pairs had a median score of +1.30 within-family standard deviations, compared with −0.75 for similarly detailed pairs describing different work and −0.63 for short aligned pairs.

All descriptions and scores are synthetic and use public 50-dimensional GloVe vectors.

Download the synthetic descriptions and scores ↓