Dynamic ABSA for E-commerce Customer Feedback Optimisation

Rodrigues, Jenat Jiffna (2025) Dynamic ABSA for E-commerce Customer Feedback Optimisation. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (802kB) | Preview
Preview	PDF (Configuration Manual) Download (446kB) | Preview

Abstract

The blistering development of e-commerce has produced enormous amounts of customer reviews, and overall ratings are insufficient to provide comparative sentiment analysis. Conventional sentiment analysis does not attempt to acknowledge opinions regarding particular elements of products, whereas most aspect-based sentiment analysis (ABSA) frameworks assume training on features of a set of predefined aspects and are therefore restricted in their potential to adapt to novel or emergent product characteristics. The study will suggest a completely dynamic and least supervised ABSA pipeline that can identify details in raw e-commerce reviews in a fully automatic way and categorise sentiments related to the study. The system trains DistilBERT on 34,660 Amazon customer reviews to create contextual sentence embeddings, uses PCA and UMAP to reduce dimensions and clusters the data with MiniBatchKMeans to identify coherent aspect groups and labels the data using a weighted combination of KeyBERT, SBERT semantic similarity, and corpus specificity. A narrow-focused DistilBERT model for sentiment (positive, negative, neutral) prediction is done by aspect. The best setting (K=9 clusters) was found to have silhouette score of 0.78, classification accuracy of 93.76 in sentiment and 1,000 reviews per second of real time inference at consumer grade hardware (Google Colab GPU). These were nine meaningful aspects found (e.g., Books, Bluetooth, Battery Life, Display) and actionable business insights were obtained based on the distributions of aspect-sentiment. The hybrid methodology proposed is much more accurate, interpretable, and deployable than the traditional TF-IDF and BERTopic baselines, and provides a scalable solution to real-time e-commerce feedback optimisation.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Agarwal, Bharat UNSPECIFIED
Subjects:	Q Science > QA Mathematics > Electronic computers. Computer science T Technology > T Technology (General) > Information Technology > Electronic computers. Computer science P Language and Literature > P Philology. Linguistics > Computational linguistics. Natural language processing H Social Sciences > HF Commerce > Marketing > Consumer Behaviour H Social Sciences > HF Commerce > Electronic Commerce
Divisions:	School of Computing > Master of Science in Data Analytics
Depositing User:	Ciara O'Brien
Date Deposited:	09 Sep 2026 08:13
Last Modified:	09 Sep 2026 08:13
URI:	https://norma.ncirl.ie/id/eprint/9901

Actions (login required)

View Item