Sentiment Analysis of Noisy Student Feedback using SVM and BERT with Explainable AI

-, Sreenivas (2025) Sentiment Analysis of Noisy Student Feedback using SVM and BERT with Explainable AI. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (502kB) | Preview
Preview	PDF (Configuration Manual) Download (537kB) | Preview

Abstract

The online learning landscape has expanded at an unprecedented level which has led to the increased student feedback, most of which is informal and holds recordable information about the quality of courses and experience as a student. Nevertheless, the chaotic and unorganised format in which such feedback takes place is very challenging to a machine when it comes to interpreting sentiment. The present study examines the comparative performance of conventional machine learning and transformer-based solutions with explainable AI (XAI) to the sentiment analysis problem in this context.

Two pipelines, (1)TF-IDF + Linear Support Vector Classifier (LinearSVC) model and (2) a fine-tuned BERT transformer model, were used to process a dataset of GPT-labeled student feedback. Both the models were tested with 1,000 reviews and SHAP-based explainability was used on the 100 top reviews with more than 50 characters, to give local and global feature attributions.

The experimental findings revealed that the SVM attained the test accuracy and F1-score of 99.49% and 99.43% respectively whereas the BERT model has a test accuracy and F1-score of 98.89% and 98.75% respectively. Although SVM had a slightly better performance in classification, SHAP showed a key distinction: BERT with its context aware and multi-word senses was able to value semantically rich tokens like helpful and training, BERT was contextual- and multi-word-sensitive, SVM simply used the high-weighted TF-IDF tokens as a whole, being more sensitive to noise and more effective with explicit sentiment keywords.

These results indicate that transformer modeling based approaches together with XAI can lead to more balanced performance in terms of interpretability and robustness on noisy, less formalized student feedback, but that carefully-pruned traditional models can still deliver competitive baseline performance. Automated educational analytics and adaptive learning systems informed by the sentiment of learners have implications to these findings

Item Type:	Thesis (Masters)
Supervisors:	Name Email Ain, Qurrat Ul UNSPECIFIED
Subjects:	L Education > L Education (General) Q Science > QH Natural history > QH301 Biology > Methods of research. Technique. Experimental biology > Data processing. Bioinformatics > Artificial intelligence Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Artificial intelligence P Language and Literature > P Philology. Linguistics > Computational linguistics. Natural language processing L Education > LC Special aspects / Types of education > E-Learning
Divisions:	School of Computing > Master of Science in Data Analytics
Depositing User:	Ciara O'Brien
Date Deposited:	24 Aug 2026 14:26
Last Modified:	24 Aug 2026 14:26
URI:	https://norma.ncirl.ie/id/eprint/9611

Actions (login required)

View Item