CLARITAS: Digitizing Educational Feedback Collection Through AI-Driven Sentiment Analysis and Agentic Orchestration

Abstract

Traditional feedback collection in educational institutions is often paper-based, resulting in delayed analysis and limited utilization of student insights. This paper presents CLARITAS, an AI-powered educational feedback analytics platform designed to digitize and automate feedback collection and analysis for classrooms and training sessions. Featuring a role-based architecture for administrators, teachers, and students, the system analyses feedback using locally deployed DistilBERT for sentiment classification, Term Frequency-Inverse Document Frequency (TF-IDF) for keyword extraction, and word-frequency analysis. A self-built, restricted Agentic AI layer coordinates these components to generate structured summaries without relying on external LLMs or third-party APIs. A fully local analytical workflow supports data privacy, institutional governance, and traceability. The proposed approach provides a privacy-conscious and scalable alternative to conventional manual feedback analysis.

Introduction

Student feedback is essential for evaluating teaching effectiveness, understanding student experiences, and identifying areas for improvement in educational institutions. However, traditional feedback systems rely on paper forms, spreadsheets, and basic online forms, making manual analysis time-consuming, inconsistent, and inefficient. These methods often fail to identify recurring themes in textual responses and tend to focus more on numerical ratings than written comments.

To address these limitations, this paper proposes CLARITAS, an AI-powered educational feedback analytics platform that automates feedback collection, sentiment analysis, keyword extraction, and summary generation while prioritizing student data privacy. Unlike systems that depend on external AI services, CLARITAS performs its entire analytical workflow locally, ensuring that student feedback is not transmitted to third-party AI APIs.

The literature review highlights the effectiveness of transformer-based models, particularly DistilBERT, for sentiment classification with reduced computational requirements. It also demonstrates the usefulness of TF-IDF for identifying important keywords. However, existing approaches often focus on individual analytical tasks rather than integrating feedback collection, sentiment analysis, keyword extraction, visualization, and automated summarization into a unified educational platform. CLARITAS addresses this gap through an integrated analytics system.

The proposed system consists of the following major components:

1. Feedback collection and management: Teachers create feedback forms containing rating-based and text-based questions. Students log in to submit their responses, which are stored in a database. A duplicate-submission mechanism prevents repeated responses to the same form.

2. Sentiment analysis using DistilBERT: A locally deployed DistilBERT model classifies textual feedback into positive, negative, and neutral categories, helping teachers understand students' overall opinions.

3. TF-IDF-based keyword extraction: TF-IDF identifies significant terms within feedback by considering their frequency in individual responses and across the complete collection. This helps reveal important topics and concerns.

4. Word-frequency analysis: Frequently occurring words are identified to highlight common themes, recurring complaints, and positive aspects of the learning experience.

5. Restricted Agentic AI orchestration: A self-developed orchestration layer coordinates the outputs from DistilBERT, TF-IDF, and word-frequency analysis. It validates and combines the results through predefined, rule-based workflows to produce structured feedback summaries without using external large language models or third-party AI APIs.

6. Analytics dashboard: Teachers can access sentiment distributions, frequently occurring terms, important keywords, rating analysis, and structured summaries to interpret feedback efficiently and identify areas requiring attention.

The methodology follows a systematic workflow involving feedback collection, text preprocessing, sentiment classification, keyword extraction, word-frequency analysis, analytical integration, and summary generation. The platform uses a layered architecture comprising a frontend, FastAPI backend, MySQL database, local NLP processing, orchestration, and analytics dashboard. Role-based access provides separate functionalities for administrators, teachers, and students.

A major contribution of CLARITAS is its privacy-preserving, locally executed analytical pipeline. By keeping feedback processing within the local deployment environment, the system reduces dependence on external services and provides greater control over sensitive educational data. Its modular architecture also supports maintainability and the integration of different analytical components.

Conclusion

CLARITAS presents an AI-powered educational feedback analytics system that automates feedback collection, analysis, and visualization through a unified platform. The system integrates locally deployed DistilBERT for sentiment classification, TF-IDF for keyword extraction, and word-frequency analysis for identifying meaningful patterns in textual feedback. A restricted Agentic AI orchestration layer coordinates these analytical components through a controlled workflow to generate structured feedback summaries without relying on external LLMs or third-party AI APIs. By performing the analysis locally, CLARITAS helps reduce the need to transmit student feedback to external AI services, supporting a privacy-conscious approach to educational data processing. Overall, the system provides an integrated approach for transforming student feedback into structured insights that can assist educators in identifying trends and areas requiring attention.

References

[1] A. Joshy and S. Sundar, “Analyzing the Performance of Sentiment Analysis using BERT, DistilBERT, and RoBERTa,” in Proc. IEEE Int. Power Renew. Energy Conf. (IPRECON), Kollam, India, 2022, pp. 1–6, doi: 10.1109/IPRECON55716.2022.10059542. [2] H. Xiao and L. Luo, “An Automatic Sentiment Analysis Method for Short Texts Based on Transformer-BERT Hybrid Model,” IEEE Access, vol. 12, pp. 93305–93317, 2024, doi: 10.1109/ACCESS.2024.3422268. [3] C.-S. Lin, C.-N. Tsai, J.-S. Jwo, C.-H. Lee, and X. Wang, “Heterogeneous Student Knowledge Distillation From BERT Using a Lightweight Ensemble Framework,” IEEE Access, vol. 12, pp. 33079–33088, 2024, doi: 10.1109/ACCESS.2024.3372568. [4] P. Riedel, M. Reichert, R. von Schwerin, A. Hafner, D. Schaudt, and G. Singh, “Performance Analysis of Federated Learning Algorithms for Multilingual Protest News Detection Using Pre-Trained DistilBERT and BERT,” IEEE Access, vol. 11, pp. 134009–134022, 2023. [5] Ü. Özmen and E. O. Y?ld?r?m, “DistilBERT-Based Hybrid Architecture for Phishing URL Detection,” IEEE Access, vol. 14, pp. 71720–71737, 2026, doi: 10.1109/ACCESS.2026.3684855. [6] Amanullah, S. Ullah, R. T. Chohan, and M. N. Khan, “Understanding Perceived Quality of Mobile Apps Using DistilBERT Model,” IEEE Access, vol. 14, pp. 3560–3576, 2026, doi: 10.1109/ACCESS.2025.3648279.

Copyright

Copyright © 2026 Vishva Sruthi N K , Shahana R, Venkata Lakshmi S. 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.