Skip to content

scoring_methodology.md

Latest commit

 

History

History
104 lines (77 loc) · 4.23 KB

scoring_methodology.md

File metadata and controls

104 lines (77 loc) · 4.23 KB
Raw

Scoring Methodology

Overview

ResuMatch scores resumes across three modules, inspired by VMock's patent (US10346803B2) and university career center methodologies. The final score is a weighted sum classified into zones.

Modules

Impact (40% weight)

Measures whether bullet points demonstrate concrete achievements.

Metric Target What it measures
Action verb rate 90%+ Bullets starting with strong verbs (Led, Developed, Implemented...)
Quantification rate 60%+ Bullets containing numbers, percentages, dollar amounts
Specificity score 70%+ Concrete vs vague language (technologies, project names, outcomes)

Scoring formula:

  • Action verb component: verb_rate / target (capped at 1.0) × 40
  • Quantification component: quant_rate / target (capped at 1.0) × 35
  • Specificity component: specificity / target (capped at 1.0) × 25
  • Weak verb penalty: -5 per unique weak verb (helped, assisted, worked on...)

Presentation (20% weight)

Measures resume structure and polish.

Metric Target What it measures
Page length 1-2 pages Ideal resume length
Grammar 0 errors Spelling, grammar, punctuation
Section completeness All required Experience, Education, Skills present
Formatting Consistent Dates, bullets, headings uniform
Contact info Complete Name, email, phone, LinkedIn, location

Scoring formula:

  • Page length: 1.0 if 1-2 pages, 0.5 if 3, 0.3 if 4+
  • Grammar: 100 - (errors × penalty), capped at penalty_cap

About the grammar check. Spelling and grammar come from LanguageTool running locally through language-tool-python, which needs a Java runtime. Without Java those checks are skipped and only formatting consistency (double spaces, bullet punctuation) is measured.

LanguageTool's dictionary is general English, so it does not recognise most technology names. Rather than penalise a resume for naming its tools, flagged words that read as technology or proper nouns are ignored: anything containing a digit or . / + # _, anything in all capitals, anything with an internal capital, and any capitalised word that is not opening a sentence. So FastAPI, AWS, CI/CD, Redis and Northwind pass, while recieved and a bullet-opening Builded are still caught. The trade-off is deliberate: an unrecognised lowercase tool name can still be flagged, which is the safer direction to fail in. Grammar and agreement errors are never filtered.

  • Sections: present_required / total_required
  • Contact: filled_fields / total_fields (5)
  • Formatting: heuristic score from consistency checks

Competencies (40% weight)

Measures skills breadth and relevance.

Metric Target What it measures
Skills count 15+ Distinct identifiable skills
Category coverage 4+ categories Technical, leadership, communication, analytical, domain
Relevance Industry-standard Recognized skill keywords vs informal descriptions

Scoring formula:

  • Breadth: min(skills_count / min_skills, 1.0) × 40
  • Category: min(category_count / min_categories, 1.0) × 30
  • Relevance: relevance_score × 30

Overall Score

overall = (impact_score × 0.40) + (presentation_score × 0.20) + (competencies_score × 0.40)

Zones

Zone Range Meaning
Green 85-100 Strong resume, minor tweaks only
Yellow 50-84 Good foundation, needs improvement
Red 0-49 Significant revision needed

JD Matching (Separate)

When a job description is provided:

  1. Keyword extraction: KeyBERT extracts top-N keywords from both resume and JD
  2. Keyword match: Jaccard-like overlap of extracted keywords
  3. Semantic match: Cosine similarity of sentence-transformer embeddings (all-MiniLM-L6-v2)
  4. Overall match: (keyword_weight × keyword_match) + (semantic_weight × semantic_match)

Default weights: 50% keyword, 50% semantic.

References

  • VMock Patent: US10346803B2 (SoftScore, keyword matching, feedback loops)
  • University career center scoring rubrics (Impact/Presentation/Competencies framework)
  • Resume NLP research (action verb detection, quantification extraction)