Análisis de sentimientos basado en aspectos (ABSA) y grafos de conocimiento aplicados al contexto educativo

Abstract

Abstract: Although student feedback is the driving force behind academic improvement, its current use is inefficient; traditional analyses collapse the wealth of opinion into blind global metrics. To address this, this research developed an aspect-based sentiment analysis (ABSA) model based on knowledge graphs. It is not just a matter of classifying the text; rather, we seek to identify specific dimensions and relate their semantic connections. In this context, the architecture we propose integrates several elements: it combines GPT-4o-mini inferences with a domain-adapted SKOS ontology and joins textual embeddings for multimodal fusion. We tested the system in a real-world scenario, using data from the OpenCampus platform of the Universidad Técnica Particular de Loja, where we persisted and processed more than 4,300 records in Neo4j. Based on this, we evaluated its performance with the LLM-as-a-Judge strategy, and the results suggest reasonable robustness: we achieved an accuracy in aspect alignment between 85 and 90%, along with a polarity fidelity of 80 to 85%. These numbers indicate that the approach helps convert unstructured data clutter into useful information for academic administration.