Similarity-Based Resume Matching via Triplet Loss with BERT Models

Automatic resume matching for the recruitment engines is an important task because of the vast volume and varying types of applicants. We propose a resume matching method to be used as a recommendation engine for recruiters. Our approach combines cutting-edge transformerbased natural language processing technology with the triplet loss, a training method originally developed for the computer vision domain. By treating the output embeddings of a transformer model similarly to those of a convolutional neural network, we develop a model for the document retrieval task. The paper also investigates a clustering based pretraining method before fine-tuning with the triplet loss. The method is applied on the data extracted from an online recruitment website, where real users actively create their own resumes. Measured by the precision at k score, the method yields an accuracy boost of %12 compared to a base model.