CV Retrieval System based on job description matching using hybrid word embeddings

摘要

We use the Average Word Embeddings (AWE) model for retrieving relevant CVs based on a job description. We designed experiments to demonstrate that the trained vectors, obtained from a balanced domain corpus, are better than using pre-trained word embeddings. We also present some experiments to show that different combinations of both word embeddings spaces increase the overall accuracy of the retrieval task compared to using only the pre-trained vectors. However an issue arised when both embeddings spaces are not sharing the same dimensions and terms, as shown in our case. In order to handle this situation, we suggest to use a method to reduce dimensions of pre-trained vectors (e.g. PCA), and combine them with our trained vectors. This improves the accuracy of the retrieval task for unseen CVs. Our main contribution is to create a model that detects which embeddings need to be used in order to maximize the relevant retrieval results. (C) 2019 Elsevier Ltd. All rights reserved.