An Explainable Rule-Based Resume-Job Description Skill Matching System with LLM-Assisted Feedback

Abstract

Automated resume screening and Applicant Tracking Systems (ATS) increasingly incorporate machine learning components and neural language models. While deep learning methods can capture complex linguistic patterns, their computational overhead, potential opacity, and sensitivity to unvetted training distributions pose practical and ethical challenges for lightweight, transparent academic deployments. In this paper, we describe an open, lightweight, and explainable resume–job description matching system that couples deterministic rule-based information extraction with optional Large Language Model (LLM) qualitative feedback. The core pipeline processes PDF, DOCX, and plain-text resumes, detects functional sections, extracts technical competencies using an auditable 108-skill taxonomy, and evaluates resume–job description compatibility using explicit set operations and a bounded 100-point heuristic score covering Skills, Experience, Projects, Education, and Certifications. To benchmark the pipeline without making real-world recruitment claims, we developed a deterministic synthetic evaluation harness containing 100 resumes, 20 job descriptions, 320 evaluation pairs, and 20 candidate ranking cohorts. Under the synthetic benchmark, the system achieved F1-scores of 0.791 for skill extraction, 0.827 for education, 0.772 for experience, 0.759 for projects, and 0.865 for missing-skill identification. Robustness testing exposed 0.0% recall when experience titles and dates were separated across lines and complete section collapse under unsupported headings. Synonym normalization increased skill recall from 0.896 to 0.968 under controlled conditions. The results characterize the behavior and limitations of the proposed deterministic baseline and provide a reproducible basis for future evaluation with real, anonymized and human-annotated resume data. The benchmark also reports multi-format consistency and controlled baseline comparisons to expose formatting sensitivity and vocabulary-related error patterns.

Introduction

The text presents a research study on an explainable and lightweight resume–job description matching system designed to improve automated recruitment while avoiding the opacity and computational demands of complex AI-based Applicant Tracking Systems (ATS).

Modern recruitment platforms process large numbers of CVs and job descriptions, increasingly using machine learning, NLP, deep learning, and transformer models for candidate screening. Although these methods can identify complex semantic relationships, they also create concerns about lack of transparency, high computational requirements, privacy issues, and questionable predictive validity. Candidates and recruiters may not understand why a particular resume received a certain score.

Proposed Approach

The study proposes a deterministic, rule-based ATS system that provides transparent and auditable results. The system separates its core screening process from optional Large Language Model (LLM) features.

The core system:

• Accepts PDF, DOCX, and plain-text resumes.

• Cleans and normalizes the extracted text.

• Detects important resume sections such as Skills, Education, Experience, Projects, and Certifications.

• Extracts skills using a predefined 108-skill taxonomy.

• Uses regular expressions to identify education qualifications, experience dates, job titles, institutions, projects, and certifications.

• Extracts skills from multiple parts of a resume rather than only the Skills section.

• Extracts required and preferred skills from job descriptions.

• Compares resume and job-description skills using set intersection.

• Produces a transparent 100-point heuristic ATS score based on identifiable matching criteria.

The system's optional LLM functions, such as resume bullet improvement and personalized cover-letter generation, are kept separate from the core screening process. Therefore, if an external LLM service is unavailable, the main resume-matching system continues to operate normally.

Dataset and Evaluation

To avoid making unsupported claims about real recruitment outcomes, the researchers created a deterministic synthetic evaluation environment using random.seed(42).

The benchmark includes:

• 100 synthetic resumes

• 20 job descriptions

• 8 technical roles

• 4 seniority levels

• 5 resume layouts

• 4 job-description skill-density categories

• 320 resume–job pairs

• 20 evaluation cohorts

The study evaluates information extraction, resume–job matching, formatting robustness, baseline matching methods, and common parsing failures.

Main Contributions

The research identifies four major contributions:

1. Explainable recruitment baseline: A transparent system based on a 108-skill taxonomy and a 100-point scoring mechanism.

2. Decoupled architecture: Optional LLM capabilities are separated from the deterministic screening process, allowing the system to work even when external AI services are unavailable.

3. Reproducible benchmark: A controlled synthetic dataset and automated evaluation framework are provided.

4. Failure-mode analysis: The study identifies weaknesses in rule-based extraction, including a 0% recall rate when experience date/title layouts do not follow expected single-line formats and complete section-detection failure when non-standard headings are used.

Literature Review

The literature review covers three major areas:

• Resume segmentation and parsing: Previous research shows that resumes require structural segmentation before information extraction because they contain different layouts and section formats.

• Skill extraction: Dictionary-based approaches are transparent and computationally efficient but can struggle with spelling variations, abbreviations, and implicit skills.

• Person–job matching: Modern neural models can capture sophisticated semantic relationships between resumes and job descriptions, but their decisions can be difficult to audit and explain.

The study therefore chooses a simpler deterministic approach to prioritize transparency, reproducibility, and auditability.

Research Gap

The main research gap is the lack of lightweight, open, reproducible, and explainable recruitment systems that combine:

• Deterministic information extraction

• Transparent scoring

• Optional LLM assistance

• Robustness testing

• Controlled evaluation of structural failure cases

The proposed system attempts to fill this gap without requiring large proprietary hiring datasets or expensive computational infrastructure.

Research Questions

The research investigates five main questions:

1. Information extraction: How accurately can the system extract skills, education, experience, projects, and certifications?

2. Matching: How effectively can it identify matching skills and missing job requirements?

3. Robustness: How well does it handle PDF, DOCX, plain text, formatting differences, and non-standard headings?

4. Baseline comparison: How do case-sensitive, case-insensitive, and synonym-based skill matching methods compare?

5. Score validity: How closely does the 100-point ATS score correspond to the synthetic evaluation rubric and candidate rankings?

System Methodology

The system follows a sequential pipeline:

Resume → Document Extraction → Text Cleaning → Section Detection → Information Extraction → Skill Matching → ATS Score

For document processing:

• PDFs are processed using pdfplumber.

• DOCX files are processed using python-docx.

• Plain-text resumes are read directly.

• Text is normalized using Unicode normalization and removal of unnecessary characters and spaces.

The section detector identifies headings such as Summary, Skills, Experience, Education, Projects, and Certifications using regular-expression patterns.

Skill Extraction

The system uses a 108-skill dictionary divided into 10 categories, including:

• Programming languages

• Frontend technologies

• Backend technologies

• Databases

• AI and machine learning

• Cloud and DevOps

• Business intelligence and analytics

• Development tools

• Testing and quality

• Soft skills

Case-insensitive regular-expression matching is used to identify skills. Skills found in the Skills, Experience, and Projects sections are combined into one candidate skill set.

Conclusion

In this work, we presented an explainable, rule-based resume–job description matching system that prioritizes algorithmic transparency, computational efficiency, and auditability over black-box predictive claims. By decoupling deterministic extraction from optional generative LLM feedback, the architecture keeps the core deterministic analysis independent of the external LLM service and provides explicit degradation behavior when the LLM service is unavailable. Benchmarking the pipeline against a controlled synthetic corpus of 100 resumes, 20 job descriptions, 320 evaluation pairs, and 20 cohorts achieved F1-scores of 0.791 for skill extraction and 0.865 for missing-skill identification, while exposing critical failure modes: a 0.0% recall rate on separated date-title lines and complete section collapse under non-standard headings. Furthermore, empirical correlation analysis showed weak agreement (?? = 0.272) between the heuristic ATS-style score and the synthetic gold-standard rubric, highlighting the necessity of caution when deploying automated screening scores. By releasing our system architecture, evaluation harness, and benchmark results openly, we provide a transparent foundation for responsible, verifiable recruitment automation.

References

[1] European Commission. (2022). European Skills, Competences, Qualifications and Occupations (ESCO) Taxonomy v1.1. Directorate-General for Employment, Social Affairs and Inclusion. URL: https://esco.ec.europa.eu. [2] Gugnani, A. and Misra, H. (2020). Implicit skills extraction using document embedding and its use in job recommendation. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI \'20 / IAAI \'20), Vol. 34, No. 08, pp. 13286–13293. DOI: 10.1609/aaai.v34i08.7038. [3] Khaouja, I., Kassou, I., and Ghogho, M. (2021). A survey on skill identification from online job ads. IEEE Access, Vol. 9, pp. 118134–118153. DOI: 10.1109/ACCESS.2021.3106120. [4] National Center for O*NET Development. (2023). O*NET OnLine: Skills and Knowledge Taxonomies. U.S. Department of Labor, Employment and Training Administration. URL: https://www.onetonline.org. [5] Qin, C., Zhu, H., Xu, T., Zhu, C., Jiang, L., Chen, E., and Xiong, H. (2018). Enhancing person-job fit for talent recruitment: An ability-aware neural network approach. In The 41st International ACM SIGIR Conference on Research & Development in Information Retrieval (SIGIR \'18), pp. 25–34. DOI: 10.1145/3209978.3210025. [6] Raghavan, M., Barocas, S., Kleinberg, J., and Levy, K. (2020). Mitigating bias in algorithmic hiring: Evaluating claims and practices. In Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency (ACM FAccT \'20), pp. 469–481. DOI: 10.1145/3351095.3372828. [7] Sajid, H., Kanwal, J., Bhatti, S. U. R., Qureshi, S. A., Basharat, A., Hussain, S., and Khan, K. U. (2022). Resume parsing framework for e-recruitment. In 2022 16th International Conference on Ubiquitous Information Management and Communication (IMCOM), pp. 1–8. DOI: 10.1109/IMCOM53663.2022.9721762. [8] Yu, K., Guan, G., and Zhou, M. (2005). Resume information extraction with cascaded hybrid model. In Proceedings of the 43rd Annual Meeting of the Association for Computational Linguistics (ACL \'05), pp. 499–506. DOI: 10.3115/1219840.1219902. [9] Zhang, M., Jensen, K. N., Sonniks, S. D., and Plank, B. (2022). SkillSpan: Hard and soft skill extraction from English job postings. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT 2022), pp. 4962–4984. DOI: 10.18653/v1/2022.naacl-main.366. [10] Zhu, C., Zhu, H., Xiong, H., Zhang, P., and Ding, N. (2018). Person-job fit: Adapting the right talent for the right job with machine learning. ACM Transactions on Management Information Systems (TMIS), Vol. 9, No. 3, Article 12, pp. 1–17. DOI: 10.1145/3234691.

Copyright

Copyright © 2026 S. Jathin Chowdary, K. Ravi Kumar, S. Mallikarjun, N. Srishma, R. Vivek. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.