A Semi-supervised Model for Automated Classification of AI-related Job Tasks using Bloom’s Taxonomy

Recent developments in artificial intelligence have been giving rise to new job profiles comprising several tasks which require different technical and cognitive skills. These tasks, most of which are new to human resources staff and job seekers, make it difficult to assess the complexity of a job and therefore challenging to select the right candidate profile. In this paper, we present an automated classification scheme based on Bloom’s taxonomy, a hierarchical model of educational objectives, which is applied on online AI-related job postings. The main goals are an improvement of the prediction accuracy of the classification model as well as an analysis of requirements for AI-related jobs. The modeling relies on a pre-trained BERT model which is fine-tuned on our dataset. In a two-step process, a semi-supervised approach is used in order to benefit from a large amount of unlabeled data. The model-generated pseudo-labels have been evaluated by the experts. Taking advantage of the now available larger correctly labeled dataset, a fully supervised training is done on the enhanced dataset and compared to the semi-supervised approach. Our results show that both models can classify AI-related tasks with good performance, supporting the use of semi-supervised training. The performance limitation rather lies in the subjectivity of expert labeling which is addressed in more detail in the paper. Moreover, we observe the model being more accurate at classifying tasks at higher levels of Bloom’s taxonomy than at lower levels.