Learning to Match Job Candidates Using Multilingual Bi-Encoder BERT
This work addresses the challenge of scalable and maintainable job matching for recruitment agencies, though it appears incremental as it applies an existing method to a new domain-specific dataset.
The paper tackled the problem of matching job candidates to vacancies by fine-tuning a multilingual BERT model with a bi-encoder structure on a labeled dataset of CV-vacancy pairs, resulting in improved semantic understanding and the ability to bridge vocabulary and language barriers.
In this talk, we will show how we used Randstad history of candidate placements to generate labeled CV-vacancy pairs dataset. Afterwards we fine-tune a multilingual BERT with bi encoder structure over this dataset, by adding a cosine similarity log loss layer. We will explain how using the mentioned structure helps us overcome most of the challenges described above, and how it enables us to build a maintainable and scalable pipeline to match CVs and vacancies. In addition, we show how we gain a better semantic understanding, and learn to bridge the vocabulary gap. Finally, we highlight how multilingual transformers help us handle cross language barrier and might reduce discrimination.