A student-professor team from Duke Kunshan University has seen their research into skill-categorization methods published by an international science journal.
Led by Gergely Horvath, assistant professor of economics, the team used wage regressions to compare different methods of extracting skill demand from the text of job descriptions. Their research, published by Information Processing and Management, which focuses on the intersection of computing and information science, could be useful to analysts at institutions such as universities and government departments looking at changes in the job market.
“Various methods exist for extracting skills from job advertisements in order to assess changes in the labor market,” said Horvath. “These vary widely and face a variety of issues such as lagging behind the development of new skills or fluctuations in skill demand.”
“We have sought to solve this issue by finding a way to determine which skill extraction method works best,” he added.
Job advertisements provide important information on the state of the labor market and are useful for assessing changes to it, such as geographical shifts in demand. In order to do this a number of systems for extracting skill information from job advertisement have been developed.
Gergely’s team measured the accuracy of these extraction methods using wage regressions. This meant first giving skills a value related to wages, and then using this to determine how well extraction methods had performed. If an extraction method had performed well, the skills it pulled from a job advertisement should be able to accurately capture wage variations across jobs.
Gergely worked on the project with four Duke Kunshan students – Ziqiao Ao, Chunyuan Sheng, Yifan Song and Yutong Sun. Initially started as a two-month Summer Research Scholars (SRS) program in 2021, the research expanded over the course of a year. DKU’s SRS program supports undergraduate students working on collaborative research with a member of the university’s faculty.
They applied their wage regression system to six skill extraction methods – three of them word counting and three topic modelling – across a data set of 3,577,509 online job advertisements posted in the United Kingdom in 2018. Word counting methods work by creating a list of skill-related keywords and then looking for them in the text of job advertisements. In topic modelling, a statistical model of text mining examines job advertisements and determines which words relate to job skills, before picking them out.
Using their wage regression system they found topic modelling methods to be better at extracting skill information. Word counting methods explained around 20% of wage variation across jobs, while the best performing topic modelling method, LDA, explained 48.3%.
The research could be useful for firms analysing labor market changes and institutions such as business schools, which may want to tailor teaching to skills required in the job market, said Horvath.
The next step would be to look at how skills included in job advertisements are related to each other, he added, which could aid education establishments in deciding what to teach together.
