Skip to main content

In a previous article, we illustrated how maintenance and enrichment of the ESCO Occupations Pillar can be supported by Artificial Intelligence (AI). The ability to connect an external job title to an ESCO occupation is a significant component of this process.


Connecting a job title with an ESCO occupation is often not a trivial task. It can sometimes be done in a straightforward way: the job title Senior Android developer can be related to Mobile application developer with sufficient confidence. In other cases, this can be a more challenging exercise. A job title does not always reflect this same level of detail as the ESCO occupations and in some circumstances a mix of different occupations might apply. Therefore, contextual information could be needed to connect external content in a more accurate way to the Occupations Pillar.


In this article, we discuss the use of supporting information, such as tasks, knowledge, and skills to connect data to ESCO occupations. First, we complement a job title with a single piece of additional information. We then examine the outcomes of entirely discarding the job title from the input, only leaving the context with a limited set of skills and tasks. Finally, we experiment with more raw input extracted from a job vacancy.


Using context when suggesting occupations


We start by combining a job title with different duties or knowledge and compare the ESCO occupations suggested by the machine learning model when changing this contextual information. Test it yourself! Drag and drop one of the options under its job title.


You may notice that, as for architect, the same job title can have a very different meaning in the labour market. This is reflected in the result of the machine learning approach, and it shows the importance of providing additional context to support the task of connecting a job title to an ESCO occupation.

Relation between tasks, knowledge, skills and occupations


We will now try to connect the contextual information directly to ESCO, omitting the job title. The following four examples refer to different management occupations that can be distinguished by the role of the personnel that is being managed. Note that these examples are largely syntactically identical to show the partial effect of changing the staff occupation. Test it yourself! Complete the sentence by choosing the type of employees and compare the different ESCO occupations suggested by the model.


As the results show, the two key parts of the input (i.e., manage and the staff occupation) are reflected in the top suggestions of the machine learning model.

Vacancy text as context


The input phrases that were used for the examples in the previous sections can typically be found in CVs, user profiles and vacancy descriptions. However, real-world data usually contain a certain amount of content which does not, or to lower degree, contribute to the task of connecting it to an ESCO occupation. For example, a vacancy might contain a general paragraph about the employer, contact details, the application procedure, benefits and salary details. In some circumstances it could even have a negative impact on the task: terminology as medical insurance or wellness programmes as parts of work benefits could be harmful if not treated carefully.


The consequence is that, while contextual information is extremely relevant to provide accurate suggestions, a machine learning model should be robust to handle noise. But how does a prediction model successfully distinguish between relevant and irrelevant information when suggesting an ESCO occupation? Even in case of a black box machine learning model, we can still mathematically explain its behaviour by deriving what drives its suggestions.


To this end, we compute the parts of the input that have the largest contribution to the predicted suggestion. For ease of presentation, these continuous score contributions are mapped to a discrete four-level scale (white, yellow, orange, red) ranging from minimal impact to maximal impact. The following visualisation shows the density function (left) for all the input word scores of the vacancy description (right). The suggested ESCO occupation for this input text is Bakery specialised seller and the coloured scoring results illustrate the parts of the input that lead to this suggestion.


Test it yourself! Hover over the density function below and see how words in the vacancy description change colour based on their relevance score, as assigned by the machine learning model.

This article explained how ESCO is leveraging contextual information through occupation mapping for improving the Occupations Pillar and support implementers. Developing the methodology is typically an iterative process. We started with a self-supervised learning approach for which the results were presented in this series of news articles. Next, this approach has been extended to the classical supervised learning approach for which we expect to publish more results and findings in the coming months.  

Learn more: