Deep learning and word embeddings for HR processes

HireGround是一家加拿大公司，专门提供将简历与招聘广告相匹配的工具，从而减轻了招聘人员的负担，否则他们就无法处理单个招聘广告的大量求职者。他们的客户来自加拿大经济的各个领域。我们力求提高招聘流程的多个方面，从而做出更好的招聘决策，让雇主和员工都更满意。对招聘广告候选人的简历进行匹配和排名是人工智能领域的一个长期挑战。一个趋势是使用定制的推荐系统，它通过对职位描述和简历进行矢量化，使资格、经验和证书与矢量的维度相对应。然后，通过比较向量（工作和简历）来完成匹配。因此，推荐系统无法轻松应对IT等快速变化的行业，因为在这些行业中，新技能总是不断出现（每次都需要重新计算所有向量）。克服这一问题的信息检索技术已经存在。例如，如果简历和职位描述有相当多的重叠术语，可以使用语言建模对简历（文档）和单个职位广告（查询）进行排序。在这个项目中，我们将扩展基于ir的解决方案，使用一种新的算法来对给定职位描述的简历进行评分，该算法基于深度学习技术产生的词嵌入。这种嵌入是单词在上下文中的向量表示。例如，由于“聚类”和“分类”这两个词经常用于描述机器学习任务，因此它们将具有相似的嵌入。因此，一个人可以在简历中提到其中一个术语，而在工作描述中提到另一个术语，即使是近似的。类似地，合并术语“机器”和“学习”的词嵌入将得到一个类似于“聚类”和“分类”嵌入的向量，允许在之前的方法失败的地方进行匹配。

HireGround is a Canadian company that specializes in tools for matching resumes to job ads, thus relieving the burden on recruiters who could not otherwise process the large volume of applicants for a single job ad. They have clients from different and large segments of the Canadian economy. We seek to enhance multiple aspects of the process, leading to better recruitment decisions and happier employers and employees. Matching and ranking resumes of candidates for job ads is a longstanding challenge in AI. One trend is to use custom-built recommender systems, which work by vectorizing job descriptions and resumes in such a way that qualifications, experience, and certifications correspond to the dimensions of the vectors. Then, matching is done by comparing vectors (of jobs and resumes). As such, recommender systems cannot easily deal with fast-changing sectors such as IT, where new skills appear all the time (requiring all vectors to be recomputed each time).Information Retrieval (IR) techniques to overcome this problem exist. For example, language modeling can be used to rank resumes (documents) against a single job ad (query), if the resumes and job descriptions have considerable overlapping terminology. In this project we will extend IR-based solutions with a new algorithms for scoring resumes given job descriptions based on word embeddings produced by Deep Learning techniques. Such embeddings are vectorial representations of the words within context. For example, because the words "clustering" and "classification" are used frequently to describe machine learning tasks, they would have similar embeddings. Thus, one can match a resume that mentions one of the terms to a job description mentioning the other, even if approximately. Similarly, merging the word embeddings for the terms "machine" and "learning" will result in a vector that is similar to the embeddings of "clustering" and "classification", allowing for a match where the previous methods would fail.