BCSE497J Project-I

Automated Interview Question Generator Application from

Resume using NLP and Generative AI

22BCE3576 ARPIT PAL

22BCE2062 AAKRITI GOENKA
22BCE2089 PRIYA RAJ

Under the Supervision of -

Dr VETRISELVI T
School of Computer Science and Engineering (SCOPE)

[Link].
in
Computer Science and Engineering

School of Computer Science and Engineering

16 September 2025

1
 ABSTRACT

This project lies in the broader domain of Natural Language Processing (NLP) and
Generative Artificial Intelligence (AI), with a focus on modern recruitment and
interview preparation. Today, recruiters spend a great deal of time manually reading
resumes, identifying important skills, and then preparing suitable interview questions.
Existing tools in the hiring ecosystem mostly help in resume screening, keyword
extraction, or candidate ranking. Some advanced platforms offer predefined question
banks, but these are often repetitive, generic, and not truly personalized to the
individual’s profile. As a result, candidates may not be fairly assessed, and recruiters
lose valuable time designing tailored questions for every applicant.

The technical gap identified here is that current systems fail to generate context-
aware, skill-specific, and adaptive interview questions directly from resumes. They
either rely on static sets of questions or do not integrate advanced Generative AI
models capable of creating natural, human-like questions. Moreover, many tools
struggle with processing unstructured resume data effectively, making it difficult to
extract meaningful insights such as project experience, domain expertise, or behavioral
competencies.

To address this limitation, we propose an Automated Interview Question Generator

Application that combines resume analysis using NLP with dynamic question
generation using Generative AI. The system will accept resumes in digital formats
(PDF/Word), apply NLP techniques such as Named Entity Recognition (NER),
dependency parsing, and skill extraction to identify education, skills, projects, and
work history, and then feed this structured data into pre-trained large language
models (LLMs) such as GPT-based models or T5 for generating interview-style
questions. The generated questions will be categorized into technical questions
(domain-specific, coding, or problem-solving based on skills) and behavioral
questions (situation-based, teamwork, leadership, adaptability).

Technically, the application is built using Python, open-source NLP libraries (spaCy,
NLTK, Transformers), and freely available generative AI models, making it cost-
effective and easy to deploy. A simple and intuitive interface will allow recruiters to
upload resumes and instantly receive personalized question sets, while candidates can
use the system to practice for interviews with their own resumes.

The expected result of the model is a scalable, accurate, and human-centered system
that significantly reduces the time and effort required for interview preparation. Unlike
existing solutions, this approach ensures that every question is personalized, clear, and
relevant, improving both recruitment efficiency and candidate experience. For
recruiters, it means less time spent on manual preparation and more time spent on
meaningful evaluation. For candidates, it provides a realistic practice tool that reflects
the kind of questions they are likely to face in actual interviews. Overall, this project

2
demonstrates how combining NLP and Generative AI can transform the hiring
process into a smarter, fairer, and more engaging experience.

3
 TABLE OF CONTENTS
[Link] Contents Page No.
ABSTRACT 2
1. INTRODUCTION
1.1 BACKGROUND 4
1.2 MOTIVATIONS 5
1.3 SCOPE OF THE PROJECT 6
2. PROJECT DESCRIPTION AND GOALS
2.1 LITERATURE REVIEW 7
2.2 GAPS IDENTIFIED 7
2.3 OBJECTIVES 8
2.4 PROBLEM STATEMENT 8
2.5 PROJECT PLAN 9
3. REQUIREMENT ANALYSIS *
3.1 HARDWARE REQUIREMENTS 10
3.2 SOFTWARE REQUIREMENTS 10
3.3 DATASET REQUIREMENTS 11
3.4 FUNCTIONAL REQUIREMENTS 11
3.5 NON-FUNCTIONAL REQUIREMENTS 11
4. SYSTEM DESIGN*
4.1 SYSTEM ARCHITECTURE 12
4.2 ARCHITECTURE DIAGRAM 13
5. REFERENCES 14

[Link]
4
1. Background

Interviews are an important part of the hiring process. They help employers understand the
skills, knowledge, and personality of a candidate. Preparing interview questions is a key
step for recruiters, but it can be a slow process when there are many applicants. Recruiters
often have to read through each resume, identify important skills, and then think of suitable
questions. This process becomes even more difficult when interviews are for technical jobs
that require different types of questions for different skill sets.

Natural Language Processing (NLP) and Generative AI are two technologies that can make
this process faster and easier. NLP allows computers to read and understand human
language, making it possible to extract useful information from resumes. Generative AI can
then use this information to create new, relevant questions in natural language.

By combining these two technologies, it is possible to build a system that takes a resume as
input and produces a list of interview questions that are customized for that person. This
means recruiters can focus on evaluating candidates instead of spending time preparing
questions.

2. Problem Statement

The traditional process of preparing interview questions is manual and time-consuming.

Recruiters need to carefully read each resume, identify key skills, and then decide what
questions to ask. This approach is not efficient when there are many candidates. It also
increases the chances of missing important skills or asking generic questions that may not
reveal a candidate’s true abilities.

There is a need for a system that can automatically analyze resumes and generate
personalized interview questions. Such a system should be accurate, easy to use, and
capable of producing questions that match both the technical and behavioral aspects of the
candidate’s profile.

3. Objectives

The main objectives of this project are:

 To develop a system that can extract important information from resumes using
NLP.

 To identify the candidate’s skills and areas of expertise from the extracted data.
5
  To generate relevant technical and behavioral interview questions using Generative
AI.

 To create an easy-to-use application for recruiters and candidates.

4. Scope

This project focuses on resumes written in English and in digital formats such as PDF or
Word documents. It covers both technical and non-technical job roles. The system will use
free and open-source tools so that it can be deployed without high costs. It is mainly
designed for recruitment purposes but can also be used by individuals for self-practice.

5. Expected Outcomes

By the end of this project, the system will:

 Accept a resume as input.

 Extract skills, education, projects, and work experience.

 Use these details to create a set of personalized interview questions.

Provide questions in simple, clear language suitable for an interview setting

[Link] REVIEW

Over the past two decades, the recruitment process has been steadily evolving with
the help of technology. Traditionally, interview preparation and resume screening
have been manual tasks handled entirely by recruiters. They would carefully go
through resumes, underline important skills, and then prepare questions accordingly.
While this approach gave recruiters control, it was time-consuming, inconsistent, and
often led to repetitive or generic interview questions. These challenges motivated
researchers and practitioners to explore automated systems that could reduce manual
effort and make the process more objective.

[Link] Approaches:
The first wave of automation in recruitment focused largely on resume parsing and
keyword matching. Systems relied on manually crafted rules and keyword
dictionaries to identify skills and experiences from resumes. While these methods
were straightforward, they struggled with variations in resume formats and natural
language. For example, the same programming skill might be written as “Python,”
“Python 3,” or “experience with scripting in Python,” and simple keyword searches
often failed to capture such nuances. These limitations opened the door for more
advanced Natural Language Processing (NLP) techniques.

6
[Link] Learning Models:
As machine learning matured, researchers began to apply statistical models like
Support Vector Machines (SVM), Random Forests, and logistic regression to
classify resumes and extract structured information. At the same time, NLP
techniques such as TF-IDF, bag-of-words, and word embeddings were used to
capture more meaningful representations of text. These models were then applied to
tasks like ranking candidates, identifying relevant skills, and even matching
candidates to job descriptions. However, when it came to generating interview
questions, these methods were still limited because they could only retrieve pre-
written templates instead of creating new, context-aware questions.

[Link] Learning and Transformer Models:

The rise of deep learning marked a turning point in both NLP and recruitment
technologies. Models such as CNNs and RNNs allowed systems to extract richer
features from resumes and job descriptions. More recently, transformer-based
architectures such as BERT, RoBERTa, and GPT have become state-of-the-art in
natural language tasks. These models not only improved the accuracy of skill
extraction but also made it possible to generate natural-sounding text. For interview
preparation, transformer-based models offered a way to generate customized,
coherent, and context-sensitive questions rather than relying on rigid templates.

[Link] AI for Question Creation:

Building on these advances, researchers started experimenting with question
generation (QG) models, originally used in education and training. Models like T5
(Text-to-Text Transfer Transformer) and GPT-series models have shown
remarkable results in generating natural and diverse questions. In the recruitment
context, these models can be adapted to generate technical questions (based on skills
or projects mentioned in resumes) and behavioral questions (based on work
experience or soft skills). Unlike traditional methods, generative AI offers flexibility,
creativity, and personalization, making it highly suitable for interview preparation
tools.

[Link] Gaps:
Despite these advances, there are still several gaps in the literature. Many existing
tools focus on resume screening rather than question generation, and the few that
attempt question generation often rely on rigid templates. Another limitation is the
lack of personalization, as most systems fail to adapt questions to the unique profile
of a candidate. Furthermore, while large language models are powerful, they require
careful fine-tuning to avoid irrelevant, biased, or repetitive questions. These gaps
highlight the need for a system that combines robust resume analysis with generative
AI to create personalized, relevant, and practical interview questions.

[Link] AND TECHNIQUES

7
Tools Used in the Project

[Link] Language

Python - We chose Python as the core programming language because it is simple, flexible, and has a
strong ecosystem of libraries. Its readability makes development faster, and it is widely used in both
Natural Language Processing (NLP) and AI projects, which made it the perfect choice for building this
application.

[Link] Framework

Streamlit - Instead of building a heavy web application from scratch, Streamlit gave us a quick and
elegant way to design an interactive interface. Recruiters and candidates can upload resumes, view
extracted content, and generate questions with just a click. Its clean, user-friendly design ensures that
even non-technical users can use the system easily.

[Link] for Resume Handling & NLP

 PyPDF2 -This library helps in reading resumes in PDF format and extracting text from them.
Without this, recruiters would need to copy-paste text manually.

 spaCy / NLTK (optional) - If deeper parsing is required, these libraries help in breaking down
sentences, tagging parts of speech, and identifying named entities such as skills, degrees, and
company names.

 Regular Expressions (Regex) - Regex is used for detecting structured data like phone numbers
and email addresses directly from text.

[Link] AI Model

 Groq API (LLaMA-3.1-8b-instant model) -This is the brain of the project. After resume data is
extracted, this model takes the information and creates personalized technical and behavioral
questions. By adjusting parameters like “temperature,” we balance between creativity and
relevance.

 Open-source alternatives - In case Groq is not available, GPT-like models from Hugging Face
can be used, ensuring flexibility in deployment.

[Link] Tools

8
  Text Preprocessing Tools - Removing stopwords, stemming, and lemmatization ensure that
extracted text is clean and focused on what matters.

 Vectorization Techniques - TF-IDF, Word2Vec, or Sentence Transformers can convert skills into
vectors, allowing the system to understand skill similarity and group related expertise.

 Download Utility -Using Streamlit’s st.download_button, recruiters or students can save

generated interview questions as a .txt file for later use.

Techniques Applied

[Link] Language Processing (NLP)

 Information Extraction (IE) - Used to pull out critical details from resumes like skills, education,
and work experience.

 Named Entity Recognition (NER) - Helps the system identify important entities such as
candidate names, universities, or company names automatically.

 Text Summarization (lightweight) - Large resumes can be shortened so that only the most
relevant parts are used for question generation.

[Link] AI (LLMs)

 Prompt Engineering - Carefully structured prompts guide the AI to first extract structured data
and then generate meaningful questions.

 Controlled Text Generation - With a temperature setting of 0.7, the system balances between
creative and precise question generation, ensuring that the questions are unique yet relevant.

[Link] & Statistical Techniques (Behind the Scenes)

 Vector Representations - Words and skills are converted into mathematical vectors so that the
system can “understand” their meaning.

 Cosine Similarity - This helps the system measure how close a candidate’s skills are to
predefined categories or keywords, ensuring alignment in question generation.

 Probability Distributions - The generative model works by predicting the next word using
probability scores, which ensures that generated questions flow naturally.

9
[Link] Engineering Practices

 Modular Coding - We separated tasks like PDF extraction into helper functions, making the code
clean and reusable.

 User Interface Design - The design is kept simple and recruiter-friendly, focusing on usability
rather than complexity.

 Error Handling - Checks are added for empty resumes, wrong file formats, or missing text,
making the system more robust in real-world scenarios.

IV. PROPOSED MODEL

Module 1 — Text extraction & preprocessing

 Input: PDF/DOCX → extracted raw text R

10
  Preprocessing tasks: normalize whitespace, remove headers/footers, preserve
section headings, sentence split.
 No learning math here; just deterministic transforms R↦text-cleaned.

Module 2 — Section segmentation & NER (token

classification)
Goal: identify sections and label tokens/entities (NAME, EMAIL, PHONE, SKILL,
DEGREE, ORG, ROLE, PROJECT, DATE).

Model: fine-tuned transformer encoder (BERT/RoBERTa).

For token xt transformer embedding ht∈R^d

11
Module 3 — Skill extraction & ontology mapping
(normalization)

Skill extraction & ontology mapping (normalization)

Two steps: (A) extract skill spans from NER and phrase-chunking, (B) map each
extracted phrase to canonical skill IDs in an ontology (e.g., ESCO/O*NET-like)

12
Module 4- Profile vectorization (candidate embedding)

Create a compact vector vvv representing the candidate’s profile for conditioning generation and scoring.

Module 5- Question generation (Seq2Seq or LLM conditional

generation)

Input: linearized profile X (fields + resume snippets) or vector v plus prompts.

Output: a sequence Q=(q1,…,qL), the question (or multiple Q's).

Model choices:

 Seq2Seq transformer (T5/BART): fine-tune on (profile → question) pairs.

 Decoder-only LLM (GPT-family): prompt with structured context; may be few-shot or fine-
tuned.

13
14
Module 6 — Candidate scoring & ranking

Module 7 — Diversity selection (MMR)

Module 8 — Bias & safety filter

15
Working model-

16
17
Data requirements & collection suggestions
 NER: labeled resumes (token-level) ~ few thousand documents across formats.

 Skill mapping: pairs (span → canonical ID) few-thousand examples or build by semi-automatic
mapping + human curation.

 Generation pairs: (profile → question set) — ideally 10k+ Q pairs, but for bootstrapping you can
create synthetic pairs from job descriptions and interview transcripts, then refine via human
validation.

 Human ratings: for RLHF collect recruiter ratings (relevance, clarity, usefulness) on generated
Qs (hundreds to thousands).

How this connects to your Streamlit + Groq app

 Keep your PyPDF2 extraction and preprocessing as input to the pipeline.

 Use the NER & skill-mapper modules to produce the short structured prompt (instead of sending
full resume raw text).

18
  For generation you can either:

o Use Groq/OpenAI/GPT via prompt (few-shot) with the structured prompt and post-
process outputs with the ranking & safety modules, or

o Fine-tune a seq2seq model offline and host it (lower hallucination, more control).

 Return questions along with the Mapping: line (which skill or resume line motivated the Q) to
satisfy the metadata requirement in your project.

V. DATASETS AND DESCRIPTION

5. REFERENCES
[Link], C.Y., “A Survey of Approaches to Automatic Question Generation,” ACL Anthology /
RoCLing, 2021. ACL Anthology
[Link], S., “A Survey on Neural Question Generation: Methods, Benchmarks, and
Directions,” IJCAI 2024 proceedings, 2024. IJCAI
[Link], N., “Automatic question generation: a review of methodologies,” Frontiers in AI /
PMC, 2023. PubMed Central
[Link]-Torrealba, R., et al., “End-to-End generation of Multiple-Choice questions
using neural language models,” Neural Networks / Elsevier, 2022. ScienceDirect
5.“Automatic Skill-Oriented Question Generation and Recommendation for Intelligent Job
Interviews,” ACM / RecSysHR, (paper/proceedings). ACM Digital Library+1

19
[Link], Y., “Enhancing Resume Content Extraction in Question …” (RecSys/HR workshop),
2023 — shows T5 variants for CV extraction. CEUR Workshop Proceedings
7.“DeepResume: Deep Learning-Based Resume Parsing” (IJ journal / 2024) —
transformer-based resume entity extraction. RJPN
8.“Repurposing General-Purpose LLMs for Skill Extraction,” arXiv, 2024 — skill-LLM
fine-tuning for job/skill extraction. arXiv
9.“Automatic question generation and answer assessment: a survey,” ResearchGate/2019–
2021 survey — overview of AQG techniques & evaluation. ResearchGate
10. “A Survey of Approaches to Automatic Question Generation from 2019 to early 2021”
(Academia / extended survey). Academia
11.“Turing-NLG and large generative models” (background on large LMs useful for AQG
contexts). WIRED
12.“Scalable Concept-Driven Question Generation” (arXiv preprint — conceptual AQG
advances). arXiv
13.“Automatic question generation using T5 — educational experiments” (research outputs
& preprints). ResearchGate
14.“Design and development of machine learning based automated screening,”
ScienceDirect (2022) — automated resume screening background. ScienceDirect
[Link] / other journals detailing advanced resume parsing pipelines and NER usage.
Ijeast
[Link] practical demos and agent descriptions for interview question generators
(ZBrain, product pages) — useful to cite real-world systems. ZBrain
17.“Scoping studies and benchmarks on AQG” (various proceedings & datasets references
in surveys). ACL Anthology
[Link] on combining CV extraction with question generation (RecSysHR paper +
workshop materials).

20