Scholarlink

A common situation today we have large applicant pools, yet many systems depend on keyword matching or fixed rules missing true intent, flexible skills, or evolving position titles. This approach offers a practical way to process resumes and score applicants. It preserves layout exactly as submitted, applies domain-specific named entity recognition (NER), aligns resumes and job descriptions semantically, and performs final ranking through a supervised learning-to-rank (LTR) model (LambdaMART). The system bring together, reliable document conversion of PDF or DOCX, with OCR fallback. NER for skills, tenure, education, and certifications. Then, Sentence-BERT embedding for resume to JD alignment. Besides this, precise cues shape the scoring cosine similarity, normalized job titles, education level, and geographic distance. Trained on 10,000 anonymized resumes and 1,200 postings, our method increases mAP by 17%, while Recall@10 rises 12% over BM25. We report NER F1, ranking metrics, ablations, and latency, provide a top-k explanation view for recruiters, quantify shortlist fairness via disparate impact ratio, and discuss resilience to noisy OCR and behavior in niche domains.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Resume Parsing for Candidate Ranking: A Hybrid Semantic and Learning-to-Rank Approach

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Neemisha Paul,
  • Sakshi Phapale,
  • Ritik Singh,
  • Sudhir Dhage

摘要

A common situation today we have large applicant pools, yet many systems depend on keyword matching or fixed rules missing true intent, flexible skills, or evolving position titles. This approach offers a practical way to process resumes and score applicants. It preserves layout exactly as submitted, applies domain-specific named entity recognition (NER), aligns resumes and job descriptions semantically, and performs final ranking through a supervised learning-to-rank (LTR) model (LambdaMART). The system bring together, reliable document conversion of PDF or DOCX, with OCR fallback. NER for skills, tenure, education, and certifications. Then, Sentence-BERT embedding for resume to JD alignment. Besides this, precise cues shape the scoring cosine similarity, normalized job titles, education level, and geographic distance. Trained on 10,000 anonymized resumes and 1,200 postings, our method increases mAP by 17%, while Recall@10 rises 12% over BM25. We report NER F1, ranking metrics, ablations, and latency, provide a top-k explanation view for recruiters, quantify shortlist fairness via disparate impact ratio, and discuss resilience to noisy OCR and behavior in niche domains.