Skip to main content

NLP helps Arbetsförmedlingen understand the labor market

Monday, May 10, 2021

Can AI and natural language processing help more job seekers find a new job? If you ask Felix Stollenwerk that question, the answer is "yes". As data scientist at Arbetsförmedlingen, The Swedish Public Employment Service, he builds applications that both help individuals and provides a better understanding of the labor market at large.

For most people, Arbetsförmedlingen's role is to help job seekers get a new job. But Arbetsförmedlingen is also the expert authority on everything related to the labor market. In both areas, language models play an increasingly important role. 

Helping job seekers finding a job is about making 1-to-1 matching, finding jobs that fit the job seeker's skill set. By doing a 1-to-many analysis, comparing an individual to the job market at large, it's also possible to estimate the 'distance' between a job seeker and the labor market. Based on that analysis it's possible to suggest what courses or other supporting activities a job seeker could benefit from. 

Finally, by doing many-to-many comparisons between both job seekers and employers, Arbetsförmedlingen gets the insights needed to understand the labor market as a system. 

For all of these, the 1-to-1, the 1-to-many, and the many-to-many, the data available is unstructured text. 

"The problem with unstructured text is that different people use different words to describe the same thing. Or that they use the same word, but in a different context. This means that simple keyword matching isn't enough," says Felix Stollenwerk, data scientist at Arbetsförmedlingen. 

This is where NLP, natural language processing, becomes interesting for Felix and his colleagues. With the help of NLP, it's possible to go beyond the word-by-word comparisons between different texts and instead have software that in some sense also understands the context for how a word is used. 

"We use NLP to create what we call Annotated Job Ads, where we add a metadata layer to the unstructured text. For this, we use what's called named entity recognition," says Felix Stollenwerk.

Named entity recognition, NER, is a variant of NLP where the software can decide what words in a text represents a job title, the required skills etcetera. 

"Take 'mathematics' as one example. In a job ad, it can be used to describe different things. It could be a required skill, it could be a part of the company description, or that some of the colleagues are mathematicians. This is why words must be considered in the context to be meaningful, and a match between mathematics as a skill and mathematics as a requirement hint at a job opening suitable for someone who has a background in math." 

To Felix Stollenwerk and his colleagues, the AI Sweden-led project Swedish Language Models for Authorities plays an important role. 

"This field is moving extremely fast at the moment, the evolution over only the last few years is huge. In this fast-moving world, cooperation between organizations is essential. What everyone needs are data and competence. But not every organization has enough of both. By coming together, we can pool resources in a way that benefits everyone. Thanks to Swedish Language Models for Authorities, we have access to leading researchers at Peltarion and RISE and access to early releases of state-of-the-art models, as two examples," says Felix Stollenwerk.