Leveraging AI for Skills Extraction & Research (LAiSER)

Skills are the language of education, training, and employment. From course and program descriptions to job postings, skills serve as the language in which we talk to one another about the learning outcomes of education and training programs, and the capabilities sought by firms. Skills have increasingly become the focus of government policy, as leaders grapple with skill-biased technological change and aging working populations and aim to promote equity and inclusion through skill-based hiring.

Educators, employers, learners, and policymakers often struggle to communicate in mutually intelligible ways across the diverse skill languages of their different disciplines, occupations and industries, and regions. This contributes to a wide range of inefficiencies, ranging from inadequate recognition of prior learning, work experience, and “alternative credentials” to poor alignment between the skills cultivated by educators and those prioritized in labor markets.

To address this problem, we aim to make the language of skills more fully intelligible within education and training systems, and between them and firms. We recognize that other initiatives also have this aim, and note that they often create a single, authoritative, and standardized definition of skills that are set within a master taxonomy – a “single source of truth for skills and learning data,” as one EdTech company puts it. Unfortunately, this strategy risks competition among the creators of different master taxonomies, succeeds only with a high degree of commitment among educators and firms to take up this new language, and may require many changes to existing skills descriptors and organizational practices.

Our solution is to develop a skills translation tool that works across data environments by creating a flexible architecture for translating skills among data sets. We use artificial intelligence (AI) and machine learning (ML) algorithms to link skills across multiple taxonomies, creating in effect a “taxonomy of taxonomies.” These algorithms permit us to link one set of skills data to another with all the benefits that having a static crosswalk implies, but with the flexibility to link different types of skills data as they evolve. This creates a dynamic system of skill translation using the similarities between taxonomies to link the conceptual clustering of a skill between definitions.

This process is designed to be both dynamic for process-oriented updating and flexible enough to incorporate new data as it becomes available. The ML algorithm underlying our design allows for regular adjustment as data sets evolve without requiring redesign of the core systems, while our AI-based implementation facilitates flexibility and rapid updating of existing linkages. These systems also benefit from additional work being done in this field allowing for new capabilities and features to be added without significant R&D expenditures on underlying algorithms/processes. Rather, we will be able to focus on learning from verified outputs to increase scope and reliability as well as adding additional language capabilities as new systems come online.

Our solution is designed to serve employees, ranging from those just entering the workforce to those at the end of their careers, by streamlining the coordination of data between employers and skills certification organizations (e.g. schools, trainers, etc.). Currently these workers are required to navigate a new hiring process for virtually every application and thus must provide a new interpretation the skills for which they are being considered each time or risk being excluded from the application pool. This process is costly in time and effort and is typically not even recognized as a step in the application process by the majority of applicants. Rather, businesses and schools attempt to compensate by producing more and more convoluted solutions which are often part of walled gardens of technical infrastructure, such as coordination agreements, micro-credentials, or even pre-employment testing.

LAiSER operates as an external translator for skills which all of these stakeholders and more can utilize to extract skills information and contextualize it in the taxonomy of the desired use case. Doing so is free/inexpensive (depending on the volume of data to be processed) and thus accessible at scales ranging from individuals to governments. It is also simple to understand and intuitively easy to use. We intend to partner with many education and HR management technology companies to integrate this translator into the back-end of their platforms to allow for seamless skills tagging for all of the learning and experience. Once tagged, the system allows for direct translation between any incorporated skills standard giving applicants a quick and efficient method for conveying their skills to a new potential employer.

GWIPP is a research institute within George Washington University with a long tradition of developing policy and technology projects for public good. The university has relationships with stakeholders across the world and often serves as experts in public policy and technology. The University also attracts some of the best and brightest mind from across the world as both students and faculty which allows us to quickly incorporate top experts from a variety of fields into any project we pursue.

We have also been building a core team of interested stakeholders who support the development of this project. Higher education institutions including Northern Virginia Community College, Northern Arizona University, George Washington University have pledged direct support for the project. Similarly, education technology company Territorium is supporting our integration into EdTech platforms. Internationally we have partnered with University de Technologico de Monterrey to develop a global version of the software that will also support multi-lingual processing.

We currently have a demonstration model of the skills extraction technology that can be used to tag human readable data with skills concepts. We are in the process of coding initial ML algorithms for de-duplicating and combining these skills into functional skills domains. Our partners are currently preparing for a summer development push which will combine the local data of 5 different stakeholders and 21 researchers (14 from GWIPP) to test the process using real world data.

We are actively looking for financial support from a multitude of partners but unlike most funders, SOLVE is a platform investor that also provides a relationship model which we fell is extremely important at this stage in LAiSER's development cycle. We have committed significant sweat equity in building relationships with project partners, but to scale to a national or international level we will need support creating new connections with stakeholders. Further, winning a MIT solve is a prestigious award which we believe will help unlock future partnerships and funding opportunities.