A TWO-STAGE FRAMEWORK FOR ACCURATE JOB TITLE CLASSIFICATION IN ONLINE ADVERTISEMENTS

Abstract

Large databases may be mined for knowledge using data science approaches. Recently, there has been a lot of interest in categorising online job advertisements (ads) in order to analyse the labour market. To determine the occupation from a job advertising, many multi-label classification techniques (such as self-supervised learning and clustering) have been developed and have shown satisfactory results. Nevertheless, these methods rely on specialised databases like the Occupational Information Network (O*NET) that are more suited to the US labour market and need for labelled datasets with hundreds of thousands of samples. To solve the issue of tiny datasets, we introduce a two-stage job title identification mechanism in this study. First, we categorise the job advertising by sector (e.g., Agriculture, Information Technology) using Bidirectional Encoder Representations from Transformers (BERT). The closest matching job title is then identified from the list of jobs within the anticipated sector using unsupervised machine learning methods and a few similarity metrics. In order to solve the problems of processing and categorising employment advertisements, we also suggest a unique document embedding technique. According to our experimental findings, the suggested two-stage method increases the accuracy of job title detection by 14%, reaching over 85% in some industries. Furthermore, we discovered that, in comparison to methods based on the Bag of Words model, integrating document embedding-based techniques such weighting schemes and noise reduction increases the classification accuracy by 23.5%. Additional assessments confirm that the suggested methodology either surpasses or performs on par with the state-of-the-art techniques. Finding new and in-demand jobs in Morocco has been made easier by applying the suggested technique to data from the Moroccan labour market.