RISE: resolution of identity through similarity establishment on unstructured job descriptions
- 1. RISE: Resolution of Identity Through Similarity Establishment on Unstructured Job Description Authors: Rakesh Rameshrao Pimplikar1, Kalapriya Kannan2, Abhik Mondal3, Joydeep Mondal1, Sushant Saxena3, Gyana Parija1 and Chandra Devulapalli4 1 IBM Research Lab, India 2 Hewlett Packard Enterprise, india 3 IBM Research Intern 4 IBM Software Lab, India Presenter: Joydeep
- 2. What exactly it is trying to solve?
- 3. Business Administrator Software Engineer Research Staff Member IBM Business Development Manager Software Developer Researcher Microsoft Identity Resolution JobC JobA JobB Unstructured Job Descriptions are available corresponding to the jobs
- 4. Business Application (Where & Why it is needed?)
- 5. • IBM Talent Management (https://www.ibm.com/talent-management) 1. Job Data Normalization/sanitation 2. Automated Creation of Job Description with context preferences 3. Semantic classification of jobs 4. Candidate routing to the appropriate jobs across the organizations
- 6. How the problem has been solved?
- 7. Attribute Identification Phase (AIP) Classifier Training Phase (CTP) Attribute Association Phase (AAP) Extraction Phase (EP) Similarity Phase Identity Resolution Phase (IRP) RISE
- 8. AIP Depth of knowledge, Process, Tools and Technologies, Skills, Domain Knowledge, Experience, Business Knowledge, Efficiency of communication, Roles, expected to perform, Performance expected, Project Management, Schedule Management, Training undergone, Education level, Education streams, Responsibilities PCA Qualitative Analysis by domain experts Depth of knowledge, Process, Tools and Technologies, Skills, Domain Knowledge, Experience, Business Knowledge, Efficiency of communication, Roles, expected to perform, Performance expected, Project Management, Schedule Management, Training undergone, Education level, Education streams, Responsibilities
- 9. CTP & AAP • Multi-Label Classifier Training for labelling every lines of the Job Descriptions. Apply the classifier on new JD. • Feature Used : Taxonomy of Categories
- 11. Similarity Phase • Jaccard Similarity • Experience Similarity: Skills, Roles, Education level, Responsibilities Overall Similarity :
- 12. System Overview
- 13. Experiments • 27K JDs from 28 different organizations • Total 0.18 Millions of lines • Manually tagged 5K lines • Total ~6K phrases tagged manually • 10 people built the ground truth • 86% agreement on total annotations (every line was annotated by 2 people)
- 14. Results • Classifier Evaluation: F1 Score Based Comparison
- 15. Results • Keyword Extractor Evaluation: • Precision: 0.954 • Recall : 0.842 • F1 Score: 0.896 • Overall Similarity Evaluation: • Precision: 0.88 • Min AU-ROC: 0.642 • Max AU-ROC: 0.9 • Mean AU-ROC: 0.779