ResuNER : Intelligent Resume-Parsing NLP-NER Model Web Pipeline

May 2022 - Sep 2022

Associated with The Reflect OrganizationAssociated with The Reflect Organization

From the AI perspective, our main goal was to create a machine learning model which was able to generate cover letters undistinguished from a human-written one. After reviewing different NLG approaches we have decided for the training already existing transformer model, as the most efficient approach.

The first training of the small 124-M parameters GPT-2 model with the dataset of collected cover letters gave us very impressive results (loss=0.19, avg=0.64, 1500 steps): the newly trained model was able to create the text on a very high qualitative level, but the content was not relevant to the candidate skills and job title.

To get more control over the model output we have decided also to add to the training dataset skills extracted from cover letters and job titles. To find this approach helped us the article from Ivan Lai.
It was a challenge to find the most suitable model, which was able to extract correctly the skills from cover letters, but after some experiments with different NLP frameworks (BERT, TFIDF), we decided to use SpaCy Named Entity Recognition model (en_core_web_sm) combined with EntityRuler for custom skills labeling.

The skills dataset we found in Microsoft repository on Github. The created NER (named entity recognition) framework helped us to add into the training dataset skills and job titles. The retrained GPT-2 based model was able to create a cover letter by prompting a set of skills, job title, or combination of both.

In the results we were able to see that the generated text was relevant to the passed into the model information: We need to mention that the description of experiences, the name of the candidate as well as the recipient was created here by the model randomly.