Matching the best profiles to a job description may be a difficult and time-consuming task. In fact, the traditional way that recruiters are using to select candidates doesn't take into consideration all important details. Recruiters have to screen all the applications manually and then calculate the similarity in an efficient way.
The idea is to calculate the similarity between the resume and the job description and then return the resumes with the highest similarity score.
Information Extraction is the task of automatically extracting structured information such as entities, relationships between entities, and attributes describing entities from unstructured sources. Our system uses spacy PhraseMatcher to extract the information from job descriptions. We prepared a dictionary that has all education degrees categories, all majors and skills categories related to computer engineering field. We fed that dictionary to the Spacy rule-based PhraseMatcher in order to detect and recognize entities in our job description. The job information extraction would look like:
Our structured job descriptions dataset :
We implemented matching rules to calculate the similarity between the resume and the job description. Those matching rules don't only use simple keywords matching but also ontology matching techniques.
In this part, we will use semantic similarity-based approach to match resumes' skills and jobs' skills. Semantic similarity approach is the task of searching for documents or sentences (resumes) which contain semantically similar content to a search document (the job description).
We tried different pre-trained models to embed our words ('all-mpnet-base-v2', 'paraphrase-MiniLM-L6-v2', 'gpt3','all-MiniLM-L12-v1','all-roberta-large-v1','bert-base-nli-mean-tokens'). Embeddings are a key part of modern NLP, they encode the meaning of words or other linguistic units into vectors of numbers.
The steps that we followed later on to calculate the similarity between the job and the resume with each model:
-
We derive semantically meaningful word embeddings using the different models mentioned above
-
We compare those embeddings with cosine similarity to find the nearest resumes to the job description Cosine similarity is defined as the inner product of two vectors divided by the product of their length. Cosine similarity is defined as:
where vectors a and b have the same number of dimensions N. Cosine similarity can be used to compare similarity between document vectors.
We tried that approach on skills words of 15 resumes and 5 jobs : The resumes information were extracted by a colleague and the jobs were extracted from a public dataset on Kaggle.
- Jobs:
- Resumes:
These are some of the scores' results:
These are the ranks of resumes for each job:
Evaluation:
We used precision @k metric to evaluate our models' results comparing to manually annotated dataset. Precision at k is the proportion of recommended items in the top-k set that are relevant. In our case k is 15. These are the results for each job and model:
The total precision for each model on 5 jobs:
Visit this notebook for the full code of the semantic similarity approach.
We have chosen 'bert-base-nli-mean-tokens' as skills matching model for its high precision.- 3rd step: We Calculated the final similarity score (the mean of skills matching score, degrees matching score and majors matching score) and returned the resumes with the highest similarity score.
- We retrieved information from the job description using Spacy rule-based PhraseMatcher
- We implemented matching rules for the degrees' levels and the acceptable majors
- We compared between 6 powerful word embedding models to generate the skills matching scores and we chose 'bert-base-nli-mean-tokens' for its high precision
Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks paper