Enhancing Sentiment Analysis: A Comparative Study of Rule-Based Discourse Models and Graph-Based Neural Networks

Abstract

Sentiment analysis plays a critical role in understanding opinions in text. Traditional rule
based approaches typically leverage linguistic cues, sentiment lexicons, and discourse markers to detect 
sentiment polarity in text. However, these models face several limitations when dealing with contextual 
dependencies, sentiment shifts, and long-range interactions. In this study, we contrast a rule-based 
discourse model (Benamara et al., 2017) with a more advanced sentiment analysis method based on a 
combination of Graph Neural Networks (GCN) and BERT embeddings. We construct a discourse-aware 
sentiment graph in which nodes represent individual sentences, words, or aspects, and edges capture 
syntactic dependencies, discourse relations, and contextual links. Our experiments on IMDB, Amazon 
Reviews, and Twitter Sentiment datasets demonstrate that the proposed GCN-based model significantly 
outperforms traditional rule-based systems, achieving a 12% accuracy improvement and better handling 
of sentiment propagation, context shifts, and complex sentence structures.