Aspect-Specific Sentiment Classification using a RoBERTa BiLSTM Hybrid Model with Hierarchical Attention

Mulkuri, Manupavan (2025) Aspect-Specific Sentiment Classification using a RoBERTa BiLSTM Hybrid Model with Hierarchical Attention. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (926kB) | Preview
Preview	PDF (Configuration Manual) Download (755kB) | Preview

Abstract

Since sentiment analysis has become increasingly relevant to interpreting the opinion of users, conventional approaches that give only a positive or negative score tend to be inadequate in modeling the fine details of real world reviews. Individuals tend to have varied sentiment about a particular attribute of the product or service as in the case of people saying good about the plot of a movie and bad about its acting. Current transformer models such as RoBERTa are able to capture lots of contextual information and in general can capture the contextual information well but they lack transparency and ignore the sequential nature of language. This thesis proposes a hybrid machine learning model to overcome these limitations based on RoBERTa embeddings and the use of BiLSTM layers and a hierarchical attention mechanism. In addition to increasing accuracy, the model is aimed at increasing interpretability by making it clear how each word is helping make aspect-specific sentiment decisions. We consider two datasets SST-2 and IMDb, with frozen and fine-tuned settings of the model evaluation. The fine-tuned model of the performance that gave the accuracy of 93.43% and 93.81% respectively outperforms the baseline of RoBERTa-BiLSTM. Also, SHAP analysis helps to visualise the aspect-level prediction contributions giving evidence of the transparency of the model. On the whole, this paper indicates that there are benefits to incorporating hierarchical attention with the contextual and sequential models in making sentiment analysis more precise and explainable, particularly with practical implications both on the research side and in practical implementations.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Shahid, Abdul UNSPECIFIED
Subjects:	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 Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Machine learning
Divisions:	School of Computing > Master of Science in Artificial Intelligence
Depositing User:	Ciara O'Brien
Date Deposited:	02 Jun 2026 11:32
Last Modified:	02 Jun 2026 11:32
URI:	https://norma.ncirl.ie/id/eprint/9334

Actions (login required)

View Item