A Project Report
On
“An Explainable AI-Based Resume Screening
and Semantic Job Matching System Using
Transformer Embeddings”
Submitted in Partial Fulfillment for the Requirements for the Award of the
Degree of
BACHELOR OF TECHNOLOGY
IN
COMPUTER SCIENCE & ENGINEERING
By
SHIVANSHU KUMAR
AANCHAL SINGH
ABHAY RAWAT
(Roll No.:2370 -67,66,03)
Under the Supervision of
Shweta Mallick
(Assistant Professor)
Department of Computer Science & Engineering
SAITM, GURUGRAM
Department of Computer Science & Engineering St.
Andrews Institute of Technology & Management
Gurugram–122506
Affiliated to
MAHARSHI DAYANAND UNIVERSITY, ROHTAK (M.D.U)
ABSTRACT
The rapid digitization of recruitment platforms has significantly increased the
volume of job applications received by organizations. Human Resource (HR)
departments often face the challenge of manually screening thousands of
resumes, leading to inefficiencies, delays, and inconsistent shortlisting decisions.
To address this issue, Applicant Tracking Systems (ATS) have been
introduced; however, most existing ATS solutions rely heavily on keyword-
based filtering and rule-based matching techniques [1]. These approaches fail
to capture contextual and semantic relationships between candidate skills and
job descriptions, resulting in inaccurate candidate ranking and potential rejection
of qualified applicants.
Traditional vector space models such as TF-IDF represent documents
statistically but do not account for semantic meaning [4]. Word embedding
techniques such as Word2Vec improve semantic representation at the word
level [2], yet they lack deep contextual understanding. Transformer-based
architectures, particularly Bidirectional Encoder Representations from Transformers
(BERT), have demonstrated superior contextual modeling capabilities in various
Natural Language Processing (NLP) tasks [3]. Despite these advancements, their
integration into intelligent and explainable recruitment systems remains limited.
This project proposes an Explainable AI-Based Resume Screening and
Semantic Job Matching System that integrates TF-IDF, Word2Vec, and BERT
embeddings for comparative analysis. The system converts resumes and job
descriptions into high-dimensional semantic vector representations and computes
similarity scores using cosine similarity measures. Additionally, Explainable
Artificial Intelligence (XAI) techniques such as SHAP are incorporated to
interpret ranking decisions and enhance transparency [5]. A bias detection
module is further integrated to analyze potential demographic or institutional
discrimination in automated hiring processes [6].
The proposed system aims to improve semantic matching accuracy, enhance
fairness, and provide interpretable candidate ranking. Experimental evaluation will
be conducted using metrics such as Precision, Recall, F1-score, Mean
Reciprocal Rank (MRR), and Normalized Discounted Cumulative Gain (NDCG).
The study contributes to bridging the gap between advanced NLP models
and transparent AI-driven recruitment frameworks.
PROBLEM STATEMENT
The modern recruitment process involves analyzing large volumes of resumes to identify
suitable candidates for specific job roles. Manual screening is labor-intensive, subjective,
and prone to human bias. To automate this process, many organizations deploy Applicant
Tracking Systems (ATS). However, most ATS solutions rely on keyword-based matching
techniques, where resumes are filtered based on the presence of predefined keywords [1].
This approach presents several critical limitations:
1. Lack of Semantic Understanding
Keyword-based systems fail to recognize semantically similar terms such as
“Artificial Intelligence” and “Machine Learning,” resulting in inaccurate filtering.
TF-IDF models also suffer from similar limitations, as they treat documents as
bags of words without contextual interpretation [4].
2. Limited Contextual Representation
While Word2Vec introduces distributed word representations [2], it does not fully
capture sentence-level contextual relationships. Consequently, nuanced skills and
experience descriptions may not be accurately evaluated.
3. Black-Box Decision Making
Advanced machine learning models often function as opaque systems, providing no
explanation for ranking decisions. This lack of transparency reduces trust among
HR professionals and candidates [5].
4. Algorithmic Bias and Fairness Concerns
Automated hiring systems may unintentionally favor certain demographic groups or
educational institutions, reinforcing systemic bias [6].
5. Inefficient Ranking Mechanisms
Many systems compute similarity scores but fail to optimize ranking quality using
structured evaluation metrics such as MRR and NDCG.
Given these limitations, there is a strong need for an intelligent, explainable, and bias-
aware recruitment system that leverages transformer-based embeddings to perform accurate
semantic job matching and transparent candidate ranking.
RESEARCH GAP
A comprehensive review of prior research highlights significant limitations in current AI-
based recruitment systems:
1. Over-Reliance on Keyword-Based Matching
Existing systems primarily utilize keyword frequency models such as TF-IDF [4] or rule-
based filters. These approaches lack contextual awareness and fail to capture semantic
relationships between job descriptions and resumes [1]. There is limited research
exploring transformer-based contextual embeddings specifically tailored for recruitment
applications.
2. Inadequate Comparative Evaluation of Embedding Techniques
Although Word2Vec [2] and BERT [3] have been extensively studied in general NLP
tasks, few studies provide a systematic comparison of TF-IDF, Word2Vec, and BERT in
the domain of resume-job matching. The absence of structured comparative evaluation
creates uncertainty regarding model effectiveness in recruitment systems.
3. Lack of Explainability in Recruitment Models
Many AI-driven hiring tools operate as black-box systems without providing interpretable
reasoning for candidate ranking [5]. Explainable AI methods such as SHAP are widely
used in financial and healthcare domains but remain underutilized in recruitment
frameworks.
4. Limited Research on Bias Detection in Automated Hiring
Algorithmic bias in hiring has become a critical ethical concern [6]. However, most
recruitment systems do not integrate fairness evaluation or bias detection mechanisms.
There is a lack of frameworks that simultaneously combine semantic matching with
fairness analysis.
5. Insufficient Use of Advanced Ranking Metrics
While similarity computation is common, few studies evaluate ranking effectiveness using
metrics such as Mean Reciprocal Rank (MRR) and Normalized Discounted Cumulative
Gain (NDCG). These metrics are essential for assessing the quality of candidate ordering
in recommendation-style systems.
Research Contribution of This Project
This project aims to bridge the identified gaps by:
• Performing comparative evaluation of TF-IDF, Word2Vec, and BERT embeddings
• Integrating transformer-based contextual embeddings into recruitment systems
• Implementing Explainable AI for transparent ranking
• Incorporating bias detection mechanisms
• Evaluating ranking performance using advanced metrics (MRR, NDCG)