Research question
Does occupational knowledge contribute to class reproduction?
Children often enter the same occupational class—a broad grouping of occupations—as their parents. I asked whether knowing more about a parent’s work helps explain that pattern.
In the Education Longitudinal Study, students and parent respondents described the same parent’s job. I compared these descriptions using content-weighted semantic similarity, then tested whether the score predicted class reproduction using survey-weighted multinomial models of students’ occupations reported in 2012.
Index
Content-weighted semantic similarity.
I use word embeddings to compare meaning even when a student and parent respondent use different words. I give common terms less weight and terms characteristic of the parent’s occupational class more weight, then calculate a dot product between the two weighted description vectors. The score captures both agreement in meaning and the amount of relevant detail.
Technical details and equations
The research uses 300-dimensional GloVe vectors. After normalizing each 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 and take their unnormalized dot product. This sums the weighted similarities between words across the two descriptions.
- gw
- corpus token-frequency weight
- cw,k
- occupation-specific term weight
- a, b
- the parent-survey and student descriptions of the same job
- k
- the parent’s occupational class
Because the description vectors are sums, adding words can increase the score when they align with the other account. The illustration below shows this using concrete job duties.
A worked example
One parent’s job, two student descriptions.
Suppose a parent teaches mathematics. The text-similarity score compares a student’s description of that job with the parent’s, accounting for both agreement about the work and the amount of related detail.
Parent’s description · held fixed
“Teaches mathematics; plans lessons; grades assignments.”
A brief account
Student’s description of the same parent’s job
“Explains arithmetic.”
“Explains arithmetic” describes part of the work conveyed by “teaches mathematics.” Word embeddings capture that overlap even though the wording differs.
An account with more task detail
Student’s description of the same parent’s job
“Explains arithmetic; prepares lessons; checks assignments.”
The student also describes preparing lessons and checking assignments. These added duties agree with the parent’s account and raise the score.
How these examples are calculated
I apply the same word embeddings and weights to both comparisons. The only change is the additional duties in the student’s description.
| Student’s account | Score |
|---|---|
| A brief account | 1,524.79 |
| An account with more task detail | 5,840.78 |
The score sums weighted similarities between words in the two accounts. Its scale depends on the word weights and the amount of related detail. Here, adding “prepares lessons” and “checks assignments” increases agreement with the parent’s description.
These calculations use public 50-dimensional GloVe vectors and simplified tokenization; the research uses 300-dimensional vectors. Word weights come from a fixed reference corpus of 60 hypothetical description pairs across education, health care, electrical trades, and logistics.
g_w = ln(D / F_w)c_w,k = 1 − log₂(1.01 − f_w,k / F_w) + ln(f_w,k / 2)D is the number of descriptions in the reference corpus, F_w is a word’s token count across that corpus, and f_w,k is its count in job family k. Terms with nonpositive class weights are omitted. The word weights are held fixed for both student accounts.
All descriptions and scores are synthetic. Public GloVe vectors supply the word representations.
Download this worked exampleDownload the reference corpus and scores