A Knowledge Graph-Driven Approach to Aspect-Based Sentiment Analysis for Environmental Discourse: Evaluating the Impact of Embedding Techniques

Despite the effectiveness of traditional embedding techniques like Word2Vec and GloVe in Aspect-Based Sentiment Analysis (ABSA), these methods often struggle to capture complex contextual and relational nuances in natural language—particularly within environmentally focused discussions involving long or intricate sentences. This paper introduces a novel, fully data-driven ABSA framework that integrates knowledge graphs with transformer-based models to improve sentiment interpretation in environmental texts. Unlike previous approaches, our system constructs knowledge graphs directly from raw input without relying on external ontologies or resources, enabling adaptability across various domains. By combining BERT's contextual language understanding with Node2Vec's graph-based relational embeddings, the proposed hybrid model captures both semantic depth and entity relationships. We evaluate our model against established techniques such as Word2Vec, GloVe, and BERT alone, using both textual and graph-based embeddings. Experimental results on the SemEval-2015 Restaurant dataset show a classification accuracy of 98%, demonstrating the model’s effectiveness. The framework's modular nature also allows seamless integration of alternative embeddings or graph structures, making it highly applicable for analyzing public sentiment around environmental policies, sustainability initiatives, and ecological issues. This work contributes to the development of more robust ABSA models suited for interpreting complex environmental narratives and stakeholder opinions.