AI Based questions generation system

The problem is to design a system that generates useful, source-grounded questions from educational content while reducing repetitive manual work and preserving teacher control over accuracy, difficulty, and suitability. 1.2 Objectives • Accept educational text or uploaded notes and prepare them for processing. • Generate multiple question formats with configurable quantity and intended difficulty. • Provide suggested answers and references to source material. • Allow users to edit, approve, save, and export question sets. • Evaluate quality using transparent criteria and human review. 1.3 Scope The system is intended for practice and draft assessment creation. It does not replace educators, guarantee syllabus alignment, or independently determine a learner’s ability. Questions used in high-stakes assessments should be reviewed and approved by qualified teachers. AI-Based Questions Generation System | Page 2

Automatic question generation has been studied using rule-based templates, syntactic transformations, semantic representations, and neural language models. Template-based approaches are predictable and easy to inspect, but may produce repetitive questions and depend on the structure of the input. NLP pipelines can identify named entities, noun phrases, definitions, and important statements that may become question targets. These methods provide a transparent baseline for a student project. Neural language models can generate fluent questions from passages and support different question formats. However, fluency does not guarantee correctness. A question may introduce facts absent from the source, use an unclear reference, or allow multiple answers. Multiple-choice distractors are particularly challenging because they must be plausible enough to test understanding but clearly incorrect according to the source and learning objective. Evaluation of generated language often combines automatic metrics with human assessment. Text-overlap metrics can miss valid alternative questions, because several different questions may test the same concept. Human reviewers should therefore assess clarity, answerability, relevance, and educational value. Source traceability and the ability to edit or reject outputs are important for trustworthy educational use.

A practical educational tool should measure more than how many questions it generates. It should retain source links, suggest answers, detect duplicates, support difficulty controls, and include an approval workflow. This proposal integrates these features into a modular design and defines evaluation criteria that include factual support and usability.

The proposed system contains an input module, text-processing module, concept extractor, question-generation engine, answer and distractor builder, quality-check module, and user interface. A user provides notes or a topic, selects question types and quantity, and receives draft questions with suggested answers and source references. The user can edit, approve, regenerate, or export selected items.

Figure 1. Proposed workflow: Educational Text → Text Extraction and Cleaning → Key Concept Identification → Question Generation → Suggested Answers and Distractors → Quality Checks → Educator/Student Review → Export

A prototype can use Python for NLP, FastAPI for backend APIs, a database for saved question sets, and a web frontend for configuration and review. Document processing should use file-size limits and access controls. The generation engine can start with templates and later integrate a language-model API or locally hosted model. The interface should clearly distinguish generated drafts from educator-approved questions. AI-Based Questions Generation System | Page 3 4. Methodology

The system validates uploaded files and extracts text from supported formats. It detects empty, unreadable, or very short documents and reports issues to the user. Text is normalised while preserving meaningful punctuation and technical terms. Long documents can be divided into coherent chunks, and each chunk retains its source location so generated questions can be traced to their origin. Scanned documents may require OCR, and extraction quality should be checked before generation.

Key concepts can be identified using noun-phrase extraction, named-entity recognition, keyword ranking, term frequency, or course-specific vocabularies. Definitions, comparisons, cause-and-effect statements, process steps, and factual claims can be prioritised because they support different question formats. Keyword extraction alone cannot fully interpret meaning, so candidate concepts should be evaluated with surrounding context and the learning objectives.

Templates can transform a definition into a question such as “What is …?”

• Sentence transformation can replace a key phrase with an appropriate question word.
• A language model can draft natural questions from a selected passage and requested format.
• Descriptive questions can target explanation, comparison, application, or process steps.
• Each item should retain its source passage and suggested answer for verification.

An MCQ contains a clear stem, one correct answer, and plausible distractors. Distractors can be selected from related concepts or common misconceptions, but must not be correct under a reasonable interpretation. The system should check for repeated options, grammatical clues, and answer ambiguity. If reliable distractors cannot be produced, the item should be marked for manual editing rather than filled with invented alternatives.

Difficulty may be approximated through cognitive demand, such as recall, understanding, application, analysis, or evaluation. This is only a suggested label: actual difficulty depends on learner background, course level, language, and context. Educators should be able to edit the label and associate a question with a learning outcome.

1. Receive source text and configuration. 2. Validate and extract text. 3. Split text into coherent sections. 4. Identify key concepts and candidate answers. 5. Generate questions in selected formats. 6. Create suggested answers and distractors. 7. Check duplicates, source support, and answerability. 8. Present drafts for human review. 9. Save and export approved items.

Quality checks may flag questions without a source-supported answer, duplicate questions, options with repeated wording, and questions that require knowledge absent from the input. Automated checks cannot reliably detect every ambiguity, so educator review remains necessary. The system should preserve the source excerpt and allow the reviewer to edit both the question and answer. AI-Based Questions Generation System | Page 4

Accept pasted content and supported educational documents.

• Let users choose question formats, quantity, and intended difficulty.
• Display suggested answers and source references where available.
• Allow editing, regeneration, approval, deletion, and export.
• Save content with appropriate access controls and user consent.
• Allow a reviewer to flag incorrect or ambiguous generated items.

Reliability: handle malformed inputs with clear error messages.

• Usability: provide readable question formatting and simple controls.
• Security: validate uploads and restrict access to saved content.
• Performance: process typical passages within a reasonable time.
• Maintainability: separate parsing, generation, validation, and export modules.

A basic database can include User, SourceDocument, QuestionSet, Question, AnswerOption, ReviewRecord, and ExportHistory entities. A Question record can store its type, text, suggested answer, difficulty label, source passage reference, generation method, review status, and creation time. For an MCQ, answer options can be stored separately with a flag identifying the proposed correct option. Only approved users should access stored documents and question sets.

Field Example / Description question_id Unique identifier for each generated item. source_id Reference to the uploaded chapter or passage. question_type MCQ, short answer, or descriptive. question_text Draft question shown to the learner. suggested_answer Answer generated from the source content. review_status Draft, needs revision, or approved. difficulty Suggested recall, understanding, application, or analysis level. This data structure supports traceability and makes it possible to update or remove a question without modifying the original source document. A production system should also implement backup, retention, and deletion policies. AI-Based Questions Generation System | Page 5

A controlled evaluation can select passages from multiple topics and generate a fixed number of questions per passage. At least two reviewers can independently rate a sample using a documented rubric. Disagreements should be resolved through discussion or a third reviewer. The evaluation should compare the template-based baseline with any optional model-based approach using the same input passages and comparable output requirements.

Important error categories include unsupported facts, vague wording, multiple possible answers, incorrect suggested answers, implausible distractors, repetition, and failure to cover important concepts. Errors should be counted by type and linked to examples. This can show whether the main weakness lies in text extraction, concept selection, generation, or quality checking.

No numerical results are presented because the proposed system has not been evaluated in a controlled experiment. A completed study should report the number of passages and questions reviewed, evaluator guidelines, agreement between reviewers where possible, generation settings, model version, and limitations. Automatic text-overlap scores may be supplementary, but should not replace human judgment about educational value.

7.1 Advantages • Can reduce repetitive drafting work and speed up practice-question creation. • Supports multiple formats from the same source passage. • Source links and suggested answers help users verify output. • Modular components can be tested and improved independently. • Can support self-study and classroom revision when questions are reviewed. 7.2 Limitations • Questions may be inaccurate, ambiguous, repetitive, or outside the syllabus. • Distractors may be implausible or accidentally correct. • Scanned or poorly formatted documents may be extracted incorrectly. • Automatic difficulty labels may not match actual learner experience. • Language-model use may involve latency, cost, privacy, and data-retention risks. AI-Based Questions Generation System | Page 6

Educational material may be copyrighted or restricted to a course or institution. Users should have permission to upload and process content, and the system should minimise document retention. Generated questions must be reviewed before formal assessment, especially when marks or progression depend on them. The system should support accessibility and disclose that content is AI-generated so educators and students apply appropriate scrutiny. The system should not create discriminatory or humiliating questions about individual learners. If student performance data is later used to personalise practice, access should be limited and the purpose explained. Data should not be reused for unrelated model training without appropriate permission. A teacher should be able to inspect, correct, or delete question sets and source materials.

Use only source materials that the user is authorised to process.

• Keep a clear distinction between AI-generated drafts and approved questions.
• Review answers and distractors before using them for grading.
• Avoid retaining unnecessary personal or student information.
• Provide a correction mechanism for inaccurate or confusing output.

Future work may include multilingual question generation, syllabus mapping, diagram-based questions, adaptive practice, learning-management-system integration, and question-bank analytics. Retrieval-augmented generation could ground outputs in approved sources, while educator feedback could guide improvements. A learning platform might recommend revision questions based on a learner’s selected topics without treating generated scores as definitive evidence of ability. Another extension is to support Bloom’s taxonomy or institutional learning outcomes. The system could tag questions as recall, comprehension, application, analysis, or evaluation, and help instructors inspect whether a question bank is balanced. However, such tags should remain editable because automatic classification may be inaccurate. Additional testing is needed for different subjects, languages, and learner groups.

Before deployment, the system should be tested for file-processing safety, access control, backup and recovery, and handling of model-provider outages. API keys should be stored securely, not embedded in frontend code. Logs should avoid full source text unless required for troubleshooting and explicitly authorised. Educators should be informed about the system’s limitations and the review steps required before questions are used. AI-Based Questions Generation System | Page 7 10. Conclusion This paper proposed an AI-Based Questions Generation System that converts educational content into draft questions through text processing, concept extraction, generation, answer suggestion, and quality checks. It supports multiple question formats and keeps links to source passages to make verification easier. The proposed workflow includes review, editing, approval, and export, enabling educators to remain responsible for the final assessment content. The design emphasises that automated generation can reduce repetitive work but cannot guarantee correctness or educational suitability. Reliable use depends on source-grounded answers, clear wording, useful distractors, alignment with learning outcomes, privacy protection, and transparent evaluation. The proposal can serve as a foundation for an academic prototype, with future work focused on multilingual support, improved grounding, and systematic educator evaluation.

formats and a question count. Generation settings. 3. Extract System identifies key concepts and candidate answer spans. Concept list with source locations. 4. Generate System drafts questions and suggested answers. Draft question set. 5. Validate Automated checks flag duplicates and unsupported items. Drafts with quality warnings. 6. Review Educator edits and approves or rejects questions. Approved question bank. 7. Export System formats the approved set for printing or sharing. PDF or document output. Appendix C: Example Question Formats C.1 Multiple-Choice Question Question: Which statement best describes the purpose of text preprocessing in an NLP system? A. To prepare raw text into a consistent form for analysis. B. To guarantee every prediction is correct. C. To replace the need for training data. D. To automatically approve all generated questions. Suggested answer: A. The item is an illustrative example and should be checked against the relevant lesson before classroom use. C.2 Short-Answer Question Question: Why should generated questions retain a reference to the source passage? Suggested answer: Source references help reviewers verify that the question and suggested answer are supported by the supplied material and make corrections easier. C.3 Descriptive Question Question: Explain two limitations of automatic question generation and describe how human review can reduce their impact. Suggested answer points: The system may create inaccurate or ambiguous questions and may generate weak distractors. Human reviewers can verify factual support, improve clarity, check learning-objective alignment, and approve only suitable questions. Appendix D: Proposed Development Milestones Phase Activity Deliverable 1 Requirements and sample content selection Requirements and data plan 2 Text extraction and preprocessing Validated input module 3 Baseline question templates and answer extraction Initial generator 4 Review interface and saving Editable question bank 5 Quality checks and evaluation Test report and error analysis 6 Final review and documentation Demonstration-ready prototype End of paper