Words matter: AI can predict salaries based on the text of online job postings

The work landscape in the United States is drastically shifting: The COVID-19 pandemic has redefined important function and moved personnel out of the office. New technologies are reworking the character of quite a few occupations. Globalization carries on to drive careers to new places. And climate change worries are adding positions in the alternative electricity sector even though slicing them from the fossil gas market. 

Amid this workplace turmoil, staff, as perfectly as employers and policymakers, could profit from knowing which position qualities lead to bigger wages and mobility, claims Sarah Bana, a postdoctoral fellow at Stanford’s Digital Economy Lab, component of the Stanford Institute for Human-Centered Artificial Intelligence. And, she notes, there now exists a huge dataset that could possibly assistance deliver that comprehension: the textual content of tens of millions of on line work postings. 

“Online facts gives us with a tremendous prospect to measure what matters,” she suggests.

In fact, utilizing artificial intelligence (AI) and device learning, Bana recently showed that the phrases made use of in a dataset of extra than one million online job postings demonstrate 87% of the variation in salaries throughout a wide proportion of the labor industry. It is the first function to use this kind of a massive dataset of postings and to seem at the partnership between postings and salaries. 

Bana also experimented with injecting new textual content – adding a talent certification, for case in point – into relevant task listings to see how these words changed the salary prediction.

“It turns out that we can use the textual content of career listings to evaluate the wage-applicable attributes of work opportunities in close-to actual time,” Bana says. “This information could make implementing for work opportunities extra transparent and strengthen our approach to workforce instruction and training.”

An AI dataset of 1 million position postings 

To assess how the text of on the net job postings relates to salaries, Bana attained extra than one particular million pre-pandemic career postings from Greenwich.HR, which aggregates hundreds of thousands of occupation postings from on the web position board platforms. 

She then made use of BERT, 1 of the most highly developed all-natural language processing (NLP) versions out there, to educate an NLP design using the text of extra than 800,000 of the work postings and their related wage details. When she tested the model making use of the remaining 200,000 job listings, it correctly predicted the connected salaries 87% of the time. By comparison, making use of only the position postings’ work titles and geographic areas yielded precise predictions just 69% of the time.

In follow-up operate, Bana will try to characterize the contribution of a variety of words and phrases to the wage prediction. “Ideally, we will colour words in just postings from red to eco-friendly, in which the darker purple words and phrases are connected with reduced salary and the darker inexperienced are joined with larger wage,” she says. 

The price of upskilling: A textual content-injection experiment

To identify which techniques make a difference for income prediction, Bana made use of a textual content-injection technique: To sure relevant career postings, she extra short phrases indicating the occupation demands a unique occupation certification, this sort of as those people stated in In truth.com’s 10 In-Demand Occupation Certifications (And How To Accomplish Them). Getting these certifications can be high priced, with prices ranging from about $225 to about $2,000. But, right until now, there has been no way to determine no matter whether the financial investment is worthwhile from a income level of view. 

Bana’s experiment unveiled that some certifications (these kinds of as the IIBA Agile Investigation Certification) generate significant income gains swiftly while other individuals (these kinds of as the Cisco Accredited Internetwork Professional) do so far more slowly – valuable data for employees who would like to have superior information about how an investment in competencies education will have an impact on their salaries and potential customers, Bana claims.

Personnel are not the only ones to reward from this info, Bana notes. Businesses can use these results to greater invest in human capital, she says. If, for instance, equipment learning types reveal a gradual change away from some duties and toward other people, businesses would have advance warning and could retrain specified staff.

And policymakers thinking about what occupation schooling plans to encourage would similarly reward from being familiar with which abilities are waxing or waning in financial value.

To that close, Bana and her colleagues are currently doing the job on a companion paper that identifies what duties are disappearing from task listings above time and what new duties are showing. 

In the foreseeable future, Bana hopes that textual evaluation of work postings could yield a website-based application in which staff or providers could research the price added by upskilling or by going to a new geographic area. 

“Currently there is not a ton of clarity close to a route to greater earnings,” Bana states. “Tools like these could assist work seekers increase their career potential clients, companies establish their workforces, and policymakers react to speedy adjustments in the economic system.”

Katharine Miller is a contributing writer for the Stanford Institute for Human-Centered AI.

This story at first appeared on Hai.stanford.edu. Copyright 2022