Matching Job Applicants to Free Text Job Ads Using Semantic Networks and Natural Language Inference

Sammanfattning: Automated e-recruitment systems have been a focus for research in the past decade due to the amount of work required to screen suitable applicants to a job post, whose resumés and job ads are commonly submitted as free text. While recruitment organizations have data concerning applicant resumés and job ad descriptions, resumés are often confidential, limiting the use of direct deep learning methods. This presents an issue where traditional data-agnostic methods have greater potential to achieve good results in determining applicant suitability for a given job post. However, with the advent of transfer learning methods, it is possible to train language models independently of the task at hand, and thus independently of the available data. In this report, a language model fine-tuned on Natural Language Inference (NLI) via cross-lingual transfer learning is used for the job matching task. This is compared to a semantic method using Swedish taxonomies to construct networks with hierarchical and synonymy relations. As NLI may apply to arbitrary sentence pairs, the use of text segmentation to enhance the methods’ performance is also examined. The results show that the NLI approach is significantly better than a random suitability classifier, but is outperformed by the semantic method that achieved 34% better performance on the used dataset. The use of text segmentation had negligible effect on overall performance, but was shown to improve ranking of the top-most suitable applicants in relation to manual expert relevance scores.