Logo Lanfrica

rebumatadele/ATS-research-release

Domaine:

natural language processing

Type de record:

software
Créateur:
reb
Hôte:
Research code, measured artifacts, and reproducibility materials for a fair bilingual applicant-tracking framework in Ethiopia. # A Hybrid Framework for Fair Bilingual Applicant Tracking System in Ethiopia This repository is the public research release for the bilingual applicant-tracking framework developed by Rebuma Tadele Debebe and Melak Mesfin Ayenew at Addis Ababa Science and Technology University. It contains the source code, experiment configurations, automated tests, derived metric artefacts, and programmatically generated figures used to evaluate the framework. The operational system separates bilingual evidence extraction, ESCO-linked representation, candidate-job ranking, evidence-constrained explanation, and mandatory human review. ## Public-release scope This is a clean research archive created from the private development repository. It intentionally excludes: - raw curriculum-vitae documents and redistributed third-party datasets; - candidate or applicant records; - credentials and local environment files; - MLflow databases, caches, and workstation-specific logs; - duplicate training checkpoints and optimiser state; - journal correspondence and private submission files. Public datasets must be obtained from their original repositories under their respective licences. Dataset provenance and evaluation boundaries are documented in `docs/PUBLIC_HUMAN_ANNOTATED_DATA.md`, `docs/PUBLIC_LABEL_BENCHMARK.md`, and `docs/REPRODUCING.md`. ## Principal measured results | Component | Evaluation | Result | |---|---|---:| | Named-entity recognition | Frozen human-annotated ANEC test | F1 0.7769 | | Named-entity recognition | Synthetic CV development | F1 0.9986 | | Public English CV extraction | Primary sequence-model skill-emission coverage, 150 CVs | 13.3% | | Public English CV extraction | Provenance-preserving hybrid coverage, 150 CVs | 96.7% | | Public English IT-skill extraction | Mehyaar exact-span lexicon evaluation | F1 0.0233 | | Candidate-job ranking | Transparent skill coverage, 200 synthetic queries | nDCG@10 0.8895 | | Candidate-job ranking | TF-IDF, same controlled qu …