Automation-of-Ranking-Resumes

Problem Statement

• Many institutions and companies use resumes as a way of short-listing candidates in preliminary rounds.
• Going through every resume and short-listing is an arduous process and in many instances, it is prone to human errors.
• This project aims to automate this short-listing process by retrieval of relevant resumes to a given job description(JD).

Dataset

• Resumes and Job descriptions are collected from indeed.com and linkedIn.
• Also job reviews are scraped from ambitionbox.com. Also data related to the selected candidates are available which is used for further analysis

Tools

• Django framework is used to build the user interface.
• BeautifulSoup in python is used for scrapping documents from websites.
• NLTk and Spacy libraries are use to process documents.
• PDFMiner.six is used to convert pdf to text.

Implementation Details:

The project is developed in 3 phases.

Phase 1

• Data collected from resumes and job sections of Naukri.com and linkedIn.
• Data Cleaning
• Dictionary and Posting list are created.
• Then each resume is vectorized.
• Building search engine.
• Query is Job description given by the company
• Similarity score is computed between job description and resumes.
• Finally, search engine will rank and display the suitable candidates.
• Scope for improvement: incorporating NLP techniques like NER into the above model and improving on ranking methods.

Phase 2

• Vectorizing:

  • With the help of indexing we build document vectors using tf-idf approach(both tf and idf normalized).
  • We get a 570*3500 dimenstional matrix where 570 is the no of documents and 3500 is the dimension of each vector.

• Querying:

  • Here query is of two types. We can either query documents based on keywords(eg: skillset,college,etc) or based on job description document.
  • For the latter approach we have to upload JD do all the pre-processing (as in phase1) and then vectorize the query.
  • We integrated it in django framework.

• Ranking:

  • For ranking initially cosine similarity is used to retrieve top N documents.
  • List of selected candidates is available for cross verification. Also Binomial model was used to rank the documents.
  • Both the ranking methods were evaluated using precision@R metric.

• User interface:

  • UI for querying is built with the help of django framework.
  • Database for resume uploading is also created.

Phase 3

In phase 1 and phase 2 we build a basic search engine to retrieve the resumes which matches the JD. In Phase 3 we improved the model by working on 3 major aspects: Vectorization, Information extraction and an ensemble ranking approach.

• Vectorization:

  • Words are vectorized using GloVe vectorization.
  • GloVe model is built by taking both global statistics and concurent word similarities.
  • So it is a better fit than the tf-idf approach.

• Information Extraction:

  • For the information extraction part we used Named Entity Recognition to extract skillset info from resume.
  • For this we built a custom corpus and trained our own model so as to suit the context of our dataset.
  • This will help us to capture better features.

• Ranking:

  • Finally for the ranking we used a combined strategy of classification and ranking.
  • Classification is sone using Naive Bayes, SVM and RNN models and for ranking BM25 and cosine similarity measures were taken.
  • For a ranking, documents are chosen based on max voting of the classifiers and then the selected documents are ranking based on the average of the 2 ranking scores.

• Evaluation:

  • Models are evaluated using precision @R
  • For the base model we got precision@R value of 32%
  • After adapting research paper and implementing them, we improved precision@R to 71%.

Setting up the Model:

• Create a virtual environment for installing dependencies. creating venv
• Install dependencies from requirements.txt
• Run the djando framework. run django