Multilabel Skill Classifier & Job-Resume Skill Matcher

Multilabel skill classifier from tech job descriptions + Resume-Job matching system

This project features a multilabel text classification model to extract hard and soft skills from tech job descriptions. The model was trained using data scraped from Indeed. The data was collected in two steps:

1. Job URL collection: Due to limitations I could only scrape 26k job links scraped from the first page of various job locations. The URL scraper can be found in scraper\link_scraper.py .
2. Job Description Scraping: Using the job URLs, the job descriptions were scraped with scraper\job_description_scraper.py. All the scraped data can be found in the data folder.

Key Features

• Detects 83 different skills (hard + soft skills)
  • Hard skills: Python, SQL, JavaScript, React, Node.js, AWS, Docker, Git, Java, C++, etc.
  • Soft skills: Teamwork, Collaboration, Communication, Critical Thinking, Problem Solving, Leadership, Adaptability, etc.
• State-of-the-art transformer-based multilabel classification
• Best performing model: ModernBERT — near-perfect results
• Fast inference using ONNX runtime
• Live interactive demo on Hugging Face Spaces
• Full-featured Flask web application including:
  • Job Description → Skills Extraction
  • Resume ↔ Job Description → Skill Matching (shows matching & missing skills + confidence scores)

Automated Labeling Pipeline

It seemed like a big task to manually label all 22k job descriptions. Hence, I used a rule-based multi-labeling system using regex.

Here's how the labeling works:

1. Comprehensive Skills Dictionary
   A total of 83 target skills (hard + soft) were defined covering the most frequently mentioned competencies in tech job postings.

2. Regex-Based Pattern Matching
   Each skill is associated with a carefully crafted regular expression that captures:

   • Common abbreviations used to define the skills
   • Different spellings & variations
   • Full names and short forms
   • Case-insensitive matching

3. One-Hot Encoding
   For every job description, it was checked if the text contains any of the defined patterns for each skill.

Model Training and ONNX Inference

I initally started out with distilroberta-base. Then I explored 3 more transformer-based models and found modernbert to be the best among them with an accuracy score of 0.99. Finally, I converted the trained model into ONNX.

Model Performance Comparison

Model	Accuracy	F1-Samples	F1-Macro	F1-Micro
distilroberta-base	0.9700	0.9446	0.9243	0.9498
modernbert (best)	0.9996	0.9928	0.9969	0.9990
all-MiniLM-L6-v2	0.9700	0.9028	0.8782	0.9081
bert-base-uncased	0.9800	0.9304	0.9110	0.9348

Model Deployment

The model was depployed in HugginFace Spaces Gradio App. You'll get the implementation in the deployment folder or in the gradio app

Fig: HuggingFace Spaces Gradio App Demo

Web Deployment

The Flask-based web app lets user extract skills from job descriptions and also match their resume with the required skills. Try the Skill Extractor and Resume-Matcher on Render.

Fig: Skill Analyzer and Resume-Matcher App Homepage

Fig: Skill Analyzer Demo

Run Locally

# Clone the repo
git clone https://github.com/Naawshin/Multilabel-Skill-Classifier.git

# Switch to flask
git switch flask

# Create virtual environment
python -m venv venv
venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

# Start the app
python app.py

Feedback, feature requests, bug reports, and pull requests are very welcome!

Feel free to reach out:

• ✉️ Email: nawshintabassum88@gmail.com
• 🔗 LinkedIn: https://www.linkedin.com/in/nowshin-tabasum/

---

If this project helped you or you just found it interesting, please consider giving it a star ⭐ It really helps the project grow!