The AI
The AI
Artificial intelligence
The AI
Artificial intelligence in a nutshell.
THE PRINCIPLE OF THE CAREERCOMPASS
The development
The CareerCompass (formerly known as the interactive career atlas through the career landscape) is a web application that can be used on a PC or mobile phone. The principle is very simple.
In the first step, the user writes a free text with statements about themselves, such as ‘I suffer from back pain’, ‘I have completed an apprenticeship as an electrician’ or ‘I like being around people’. The user can then select individual filter options, such as ‘only education’, to further refine the display of results.
THE PRINCIPLE
The development
The CareerCompass (formerly known as the interactive career atlas through the career landscape) is a web application that can be used on a PC or mobile phone. The principle is very simple.
In the first step, the user writes a free text with statements about themselves, such as ‘I suffer from back pain’, ‘I have completed an apprenticeship as an electrician’ or ‘I like being around people’. The user can then select individual filter options, such as ‘only education’, to further refine the display of results.
What happens in the background?
A language model (‘AI = Artificial Intelligence’) in the background matches the free text with possible occupations. As a result, the user is presented with a map that represents their so-called ‘career landscape’. On this map, they can see various occupations that might be suitable for them, each filtered according to their individual interests, knowledge and possible restrictions. Professions that appear close to the user on the map are more likely to match the individual's description of the career they are aiming for and are suitable for them, while professions that appear further away on the map require additional training or education. Users can explore the career landscape and save professions that sound interesting for later.
The web application provides a visually appealing and realistic impression of the nearby career and training opportunities. This look is not only user-friendly, but also creates a well-organised overview of possible career opportunities and career paths.
What happens in the background?
A language model (‘AI = Artificial Intelligence’) in the background matches the free text with possible occupations. As a result, the user is presented with a map that represents their so-called ‘career landscape’. On this map, they can see various occupations that might be suitable for them, each filtered according to their individual interests, knowledge and possible restrictions. Professions that appear close to the user on the map are more likely to match the individual's description of the career they are aiming for and are suitable for them, while professions that appear further away on the map require additional training or education. Users can explore the career landscape and save professions that sound interesting for later.
The web application provides a visually appealing and realistic impression of the nearby career and training opportunities. This look is not only user-friendly, but also creates a well-organised overview of possible career opportunities and career paths.
MORE ABOUT THE AI
How does the AI work?
In the background, AI-supported natural language processing is used to calculate a spatial representation of the positions of individuals in the career landscape based on users' free-text self-descriptions.
So-called ‘sentence transformer models’ [1] are used to represent the textual self-descriptions of the users as numerical vectors. These machine learning methods transform a sequence of characters (e.g. text data) into a numerical representation (i.e. vectors). The vectors are used to perform mathematical operations, whereby in the case of the CareerCompass, a comparison between job descriptions and user information is calculated using the cosine similarity based on mathematical representations.
Cosine similarity is a measure of the similarity between two vectors and can be conceptually compared to a correlation (in the range of -1 to 1). This comparison takes place in a high-dimensional representation, often referred to as a semantic space. Here, this space consists of 768 dimensions, which is too much for a meaningful visual representation. Therefore, the results are projected onto a two-dimensional representation in the form of a graphically appealing landscape using dimension-reducing methods (e.g. UMAP).
[1]. N. Reimers and I. Gurevych. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. 2019.
Pictures: AI model / method at the top and first mock-up drafts for the CareerCompass below
MORE ABOUT ARTIFICIAL INTELLIGENCE
How exactly does AI work?
In the background, AI-supported natural language processing is used to calculate a spatial representation of the positions of individuals in the career landscape based on users' free-text self-descriptions.
So-called ‘sentence transformer models’ [1] are used to represent the textual self-descriptions of the users as numerical vectors. These machine learning methods transform a sequence of characters (e.g. text data) into a numerical representation (i.e. vectors). The vectors are used to perform mathematical operations, whereby in the case of the
CareerCompass, a comparison between job descriptions and user information is calculated using the cosine similarity based on mathematical representations.
Cosine similarity is a measure of the similarity between two vectors and can be conceptually compared to a correlation (in the range of -1 to 1). This comparison takes place in a high-dimensional representation, often referred to as a semantic space. Here, this space consists of 768 dimensions, which is too much for a meaningful visual representation. Therefore, the results are projected onto a two-dimensional representation in the form of a graphically appealing landscape using dimension-reducing methods (e.g. UMAP).
[[1]. N. Reimers and I. Gurevych. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. 2019.
MORE ABOUT ARTIFICIAL INTELLIGENCE
How exactly does AI work?
In the background, AI-supported natural language processing is used to calculate a spatial representation of the positions of individuals in the career landscape based on users' free-text self-descriptions.
So-called ‘sentence transformer models’ [1] are used to represent the textual self-descriptions of the users as numerical vectors. These machine learning methods transform a sequence of characters (e.g. text data) into a numerical representation (i.e. vectors). The vectors are used to perform mathematical operations, whereby in the case of the CareerCompass, a comparison between job descriptions and user information is calculated using the cosine similarity based on mathematical representations.
Cosine similarity is a measure of the similarity between two vectors and can be conceptually compared to a correlation (in the range of -1 to 1). This comparison takes place in a high-dimensional representation, often referred to as a semantic space. Here, this space consists of 768 dimensions, which is too much for a meaningful visual representation. Therefore, the results are projected onto a two-dimensional representation in the form of a graphically appealing landscape using dimension-reducing methods (e.g. UMAP).
[1]. N. Reimers and I. Gurevych. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. 2019.
Pictures on the right: AI model / method and below first mock-up drafts for the CareerCompass
Even MORE ABOUT THE AI
AI Training Method
We started with a pre-trained sentence transformation model that was fine-tuned to train the model for this specific application. This was done in two steps [2]:
Step 1: Polarity calibration
During polarity calibration, the model was trained to distinguish between semantically opposite concepts. This was necessary because, although the cosine similarity ranges from -1 to 1, negative coefficients rarely occur when comparing vector representations of sentences [2]. This restriction arises mainly from the fact that the high-dimensional vector representation of sentences encodes a range of abstract linguistic features, many of which typically correlate positively across text sequences. The goal of polarity calibration is to maximise the cosine distance between the vector representations of the respective concepts.
Step 2: Domain adaptation
Domain adaptation helps the model to focus on the text segments that convey relevant information. In the last step, the occupation-related data (source: official website of the Federal Employment Agency) is coded into the semantic space. Together with the coded self-reports of the users, the distances between the vectors can finally be calculated and displayed in an appealing two-dimensional space by displaying the dimensions in a procedurally generated map.
[2]. Dr. B. E. Hommel and R.C. Arslan. Language models accurately infer correlations between psychological items and scales from text alone. Center for Open Science, 2024. doi: 10.31234/osf.io/kjuce.
EVEN MORE ABOUT ARTIFICIAL INTELLIGENCE
The AI training method
We started with a pre-trained sentence transformation model that was fine-tuned to train the model for this specific application. This was done in two steps [2]:
Step 1: Polarity calibration
During polarity calibration, the model was trained to distinguish between semantically opposite concepts. This was necessary because, although the cosine similarity ranges from -1 to 1, negative coefficients rarely occur when comparing vector representations of sentences [2]. This restriction arises mainly from the fact that the high-dimensional vector representation of sentences encodes a range of abstract linguistic features, many of which typically correlate positively across text sequences. The goal of polarity calibration is to maximise the cosine distance between the vector representations of the respective concepts.
Step 2: Domain adaptation
Domain adaptation helps the model to focus on the text segments that convey relevant information. In the last step, the occupation-related data (source: official website of the Federal Employment Agency) is coded into the semantic space. Together with the coded self-reports of the users, the distances between the vectors can finally be calculated and displayed in an appealing two-dimensional space by displaying the dimensions in a procedurally generated map.
[2]. Dr. B. E. Hommel and R.C. Arslan. Language models accurately infer correlations between psychological items and scales from text alone. Center for Open Science, 2024. doi: 10.31234/osf.io/kjuce.
A GUARANTEE FOR DMARKETABILITY
The evaluation study(s)
The aim of the evaluation study is to examine the effectiveness and efficiency of the application and to ensure that it meets the needs of its users. First, the obstacles to finding a job were analysed and a tailor-made solution developed to significantly improve career guidance and development for people at all stages of their lives. The evaluation of the web application was planned in two steps: an evaluation with the original version to check the user-friendliness and to correct small errors. If the user-friendliness was satisfactory, the second step was to evaluate the application with regard to psychologically relevant concepts.
The methodological design of the study corresponds to the classic design with an intervention and a control group. The intervention group used the
CareerCompass, while the control group either did not use it or received conventional career counselling or information about vocational training from conventional websites. The effects of the above hypotheses were tested before and after the intervention. Collecting and analysing this data provided a comprehensive insight into the app's impact and identified potential for future improvements and enhancements.
MARKET READINESS?
Evaluation study(s)
The aim of the evaluation study is to examine the effectiveness and efficiency of the application and to ensure that it meets the needs of its users. First, the obstacles to finding a job were analysed and a tailor-made solution developed to significantly improve career guidance and development for people at all stages of their lives. The evaluation of the web application was planned in two steps: an evaluation with the original version to check the user-friendliness and to correct small errors. If the user-friendliness was satisfactory, the second step was to evaluate the application with regard to psychologically relevant concepts.
The methodological design of the study corresponds to the classic design with an intervention and a control group. The intervention group used the
CareerCompass, while the control group either did not use it or received conventional career counselling or information about vocational training from conventional websites. The effects of the above hypotheses were tested before and after the intervention. Collecting and analysing this data provided a comprehensive insight into the app's impact and identified potential for future improvements and enhancements.
A GUARANTEE FOR MARKETABILITY
The evaluation study(s)
The aim of the evaluation study is to examine the effectiveness and efficiency of the application and to ensure that it meets the needs of its users. First, the obstacles to finding a job were analysed and a tailor-made solution developed to significantly improve career guidance and development for people at all stages of their lives. The evaluation of the web application was planned in two steps: an evaluation with the original version to check the user-friendliness and to correct small errors. If the user-friendliness was satisfactory, the second step was to evaluate the application with regard to psychologically relevant concepts.
The methodological design of the study corresponds to the classic design with an intervention and a control group. The intervention group used the
CareerCompass, while the control group either did not use it or received conventional career counselling or information about vocational training from conventional websites. The effects of the above hypotheses were tested before and after the intervention. Collecting and analysing this data provided a comprehensive insight into the app's impact and identified potential for future improvements and enhancements.
READ THE FULL ARTICLE
The Interactive Career Atlas - an AI-based Web Application for Informed Career Decisions
Karoline Schubert, Nina Mader, Melina S. Welt, Björn E. Hommel, Franz JMWollang and Judith Volmer
READ THE FULL ARTICLE
More about the CareerCompass
The Interactive Career Atlas - an AI-based Web Application
for Informed Career Decisions
